Map Jobs-rich education as 21st C economy on every belt road

questions on NY's AI Now Report
Celebrating JIm Yong Kim
Hackhikers guide to girls sdg generation
Tsinghua #1 DigitalCooperation MIT #2
18th annual update of hubsworld
happy 2019- 2'oclock rock on stopping extinction of species
Do you know womens most valuable lesson from 20th century?
13 Belt Roads -TUPU
Exploring 100 Jack Ma Partnerships
10 times more economical health care 1984-2025
what is world's most valuable resource? - your place will grow or decline 5 fold by 2050- which? how?
Food security - insects plus
Sustainability - the most urgent games people will ever play
AI round the world
50 years mapping why sustainability needs East to be in top 2 all AI markets by 2025
what if AI world's robots were programmed by Trump?
GPLETFORC League of Nations
This months youthcooperation meetings
world record job creators before moon landing
#BR0 young leaders : naturally china largest gdp by 2030 whatever trump does
9* Sir Fazle Abed
China Capitalism - How Women & Youth were asked to Change www sustainability
supernation economics valuetrue maths crisis: nature values geo & eco systems more than humans
How can you help students and teachers who discover sustainability's missing solutions?
Love & Sustainability: remembered by son and grandddaughter
2019 Geneva to honor huawei and top 20 of made in the internet
Listing 100 reasons why big organisations have zero value unless they serve small enterprises
Huawei wizards number 1 brand
#BR2 South Asia's Coastal Belt - further references
Year in 2019 from
questions on NY's AI Now Report
where can people chat seriously about youth livelihoods
how to help next 2 years of research of world record jobs
what supercities does a nation need to sustain fifth of world people and worldwide youth
HubWorld Update
Game- choose 5 co-leaders to sustain mother earth
DigitalCooperation- Sustainability Youth World Trade Flight Map
Did positive world trade mapping end with silk road 500 years ago
Can the old world save america's new world
Why women dont see g7 nations as leading sustainability solutions
Where do gov value young technologists most
Case 1 learning from world's largest livelihood educator and cashless bnaker
Sports Arts top 10 non-sustainable market
Will USA lose First 100 Nations of Sustainability Status?
#BR1 East of China &
Do you know the 10 most disastrous mistakes macroeconomists can make
universe of world record jobs and sustainability
#BR6 #TheEconomist do americans want to save humans from extinction?
#BR10 latin america
news blockchain & edu belt road
stories developing for new uni alibabauni and arcticuni ...
#BR5 #TheEconomist West Europe
#br3 #theeconomist russia what if america's politicians and media wasted last third of century
Prince Charles
Can Artificials win-win with human Intelligences
its not social unless its 51% social
old homepage
Paradox Francis
youth's tech glossary to 2023 (2000 times moore e- than 1946)
sustainability2019-2020- linkin 1000 times moore rising suns with japan
50 shades of green
Mapping Sustainabililty WWW 1946-2030
?Mapping 3 millennia of world trade
Nobel Dr Yunus did not invent banking for the poor, what he did do...
MICES- Map-Integrate-Community-Economics-Sustainability
How English Failed Democracy's FInal Exam 2 Centuries in a Row
Guide to 4000 times Moore tech era of WorldRecordJobs Creators tsinghua &
unctad dialogue
Yunus top 10 concepts - what to learn, what not to copy?
Was JFK last president to explore how to SME every possible value chain
education's 3 greatest jobs creators
Do you know how to map world trade routes?
How about Soros & Francis & Kim if your politicians ban ending poverty with Asian Girl Power
Place branding- how singapore rose to top of class as US. EU bottomed
which markets will sustainability 30000 most vital solutions come from
Norman Macrae Economist's view of why 21st C needs to get happier and happier alumni clubs
The United Nations of Girls
Jobs Belt Roads Top 13
What is Sustainability?
17 goals of SDG nations - or 2 or 3 or 4 or 5
BENT - did Americans Lose IT?
news from china *and delas " which we dont quite understand yet
Studying Bangladesh - Economists' and Sustainability's Greatest MIracle
10 most exciting developments students 2018-2019
pro-youth financial genius -Rumee Ali
pro-youth banking 2012) and jobs education (2017)
Which of 7 Belt Road Movements are you Mapping
Asian Research Future of Education
Technology leapfrogging and the end of poverty
Is it possible to sustain planet earth without west's g7 nations?
what were the 2 world wars really about? freedom to trade...
Shall we design markets for 10 billion, 3 bn or zero people?
learn from most conscious network in girls world
MIT Matech US-East
Mapping Economics of Ports & Railroads & Digital
Livelihood Education Open Space
Global Uni of Poverty - why SDgirls value china's BAT not US FAG
brac 49 years of 7 by 7 webs of ngo world partners
Privatization - good, bad, terribly no-sustainable?
Cost of Girls True News
Wang Jianlin why under 30s need to value boats & trains more than cars and planes
My favorite europe university east of glasgow is in mortal danger?
Who's Afraid of Diversity of 195 nations
economy of billion poorest girls - bangla &
economies where mass thrives, do no evil to poorest, east belt road economy since 1950
economics of sports and
amy space
what scottish economists actually said
WRJC Missionaries Impossible
top 8 job creators jinping, pope francis, sir fazle, jack ma, jae-in &
Brac's curriculum - of how girls built 8th most populated nation on trust and love
BELT - mapping future of places if youth are to be sustainable everywhere in 21st C
jinping transcript 14 may 2017 project of the century - belt road
marketing's new ;playbook
lessons from first 3000 US alumni trained by jack ma detroit
New York Sept 2018 - Is teaching sustainability possible?
g20 argentina summit july 2018
100 cities -sustainability's last call
belts 65 industrial zones- exploring links to jobs, tech hubs and learning exchanges
survey - where are the world's favorite coding schools
help draft urgent request to world bank jim kim
Optimistic reasons 2018's sustainbility summits will leap to changing education round 3 bn new jobs
ObamaUni #TheEconomist
Feb 2018 par 1/6 months girlsnchnage education at United Nations
imagine if media was used for youth to debate top 7 job creators why BAT is not FAG
Mastermind Quiz- saving human race from extinction 2030-1946
History's WRJC including Gandhi Marx Smith Keynes &
can education help youth save humanity from extinction?
who do you know - world record job creators or destroyers?
100 years of reporting the market for poverty alleviation
2019-2020 Beijing Belt Road Forum 2 May 2019; Japan G20 and Alibaba Olympics
2018 year that india and china raced to end poverty : together!
Celebrating East West South North- What 3.5 billion elders Can do in next 1461 days
aiib 2018 mumbai june
WISE@UNGA 2018 september new york
4.5A - changing education
trade maps - development of peoples by goodwill (health) and finance (wealth)
do you know joyful stories of sustainability gen's students and teachers
3 new banking summits: aiib mumbai, sco qingdao , brics joburg
3 WISE SUMMITS - Accra May, UN New York Sept, Paris March 2019
Ali Baba - TOP 5 ECONOMY OF 21st C
WRJCbook not facebook
ngo rankings by friends of norman macrae foundation
entrepreneur military
Exciting Development Economists - Norman Macrae
whathappened USA
13 years left for 3.5 bn explorers of sustainability 2.0
is diesel one of the dirtiest of energies?
Belt up 9 regions
top 5 micro word trade platforms for creating 3 billion news jobs for under 30s
lessons from lessons from 80 country club redesigning investment banks round green infrastructure
who's who of celebrating end of tv ad age
how south korea is pullin all stops out to be china's best supporter of sustainability generation
characters at aiib 2017 include
East's 70 years of amazing development economic models
exploring the world's history and future through eyes of young chinese ladies & auld scottish allies
fast changing question - who can finance humanity's life critical goals?
explore young world's top 25 job creating platforms
how can everyone build the most valuable idea the world has ever heard of
year 34 of why not free trade with russian people
un year of ecotourism
4th grade Girl's 20 stories of futures worthy of human race
is cnn destroying its value
rejuventaion of media - why youth alumni of happy east will sustain humanity, dismal west wont
chinese millennials -sustainability half billion lead storytellers
world cultural entrepreneur - who?
100 days -- 1/14.6 #TheEconomist
what is the future of retail?
How to design 3 billion new jobs around youth
#youthtech - worldwide brand charter search by Baltimore
world's most valuable question- which markets are good for all the peoples?
2020,2019,2018 = now!
west baltiimore - the most collaborative urban city space for sustainability (goal 11)
Sustaining fully employed youth and 100$ graduate degrees
Conscious Post Its
why parents need to mentor kids in email curriculum k-12
Can world's biggest broken system be fixed? - yes only with mass collaboration
The Games of World Record Job Creation
how to value partners
10* jinping and other world record place leaders
1* Jack Ma
where is #learninggeneration going to and coming from?
future history sustainability #1461
8* Muhammad Yunus
which corporations/sectors are making education their main corporate responsibility
America's number 1 crisis - antisocial media -god bless us
What would world miss if JYK hadnt existed 2012-2022?
HOwen OPEN space
Alumni of Gandhi/Mandela
Emperor Hirohito ( Showa ) One of Greatest Leaders ever by Norman Macrae
Newest branches of POP - West Africa 014, Vatican 013, World Bank 012
the most valuable satellite guided learning tour - sustaining 7 billion being's community nursing
Grassroots rural networks save the world
Diary of , and myUNuslab
When are Bono's Pop Stars On Song in Claiming investing 10% in Agriculture best way to end poverty
Exponential Goals
44th annual newsletter of elearning millennials and job creation started at The Economist
Comparing world bank #2030now structure wit other millennials world class end poverty networks
business models
bottom up pieces of solar, and linksin to future of green finance
bangla economic miracle lesson 1 first trillion dollar audit of a sustainable nation is healthcare
valuing millennials
Help curate Soros ineteconomic invitation to millennials to rethink economics and open society
inbox; could the next 5 years of elearning make or Break our human RACE
Game 1 Top 10 Open Universities that value youth most?
valuetrue search for most human value of internet
Saving Youth - Top 100 Videos to Viralise people i wish i had introduced the poor world's greatest jobs creator to
searching bookworm
Blank page
20 classes -cataloguing frames of partners in publishing world record job creation
Why nature will not sustain human species unless act now on biggest mistakes economists made Q3 C21
Which brand most collaboratively values millennials goals to 2030?
will partnerships of catholic universities be first to free business curriculum 21st women and youth
Course World Record Job Creators -by friends of The Economist's Macrae's Net Generation's Heroes
Leapfrogging curriculum- humanity's greatest value multiplying revolution
MillennialHealth Curriculum - next half billion jobs
Dont you just love economics and media? aka jobs and the curriculum of youth economics
1758 birth of moral viewpoint of economics as social action
Will enough Under 35 year olds know how to map goodwill value chains for 21st to sustain human race?
Mindset's Great Escape: elders economics war on youth
Curriculum of safe community banking
Curriculum of The Web as I envisaged it, we have not seen it yet. The future is still so much bigger
Diary of when/where youth can linkin to sustain world
2030 curriculum of Gandhi - year 14 of gossiping good news of OMAGINE
Nobel Peace Summit Curriculum Competition
curriculum of washington dc - worst of best pro-youth capital
Conscious Capitalism $64 Trillion Dollar Curriculum - Purpose of Hi-Trust banking 99% of humans need
Who was missing from cast of first social good summit-mooc?
9 minute mooc - youth futures depend on whether ceos sustain or destroy value chains
Which trillion dollar markets have even one ceo leading best for youth futures
debates with big funders of end poverty schemes
will media barons ever learn to value connections between likes and dislikes?
The Future History of Social Business Since 1976
how does conscious capitalism relate to valuetrue exchange
VT & collaboration entrepreneur revolution of micrifranchies and bottom-up value chain mapping
VT and the compound risks of unseen wealth
The Economist & Bangladesh - VT & 1976's 2 great calls for wholeplanet redesign of 21st C systems :
3 most important metrics of pro-youth economics: goodwill, sustainability, transparency
Book Introduction to economics of youth
chris macrae linkedin
UNwomens - 10 years of leapfroging - bangladesh paradise lost?
#BR0 China & #TheEconomist
#BR9 Africa #theeconomist
#BR3 Russia #BR4 East Europe #TheEconomist
#BR7 Middle East #TheEconomist
#BR11 Arctic Belt Road
Will Americans lost First 100 Sustainability Nations Status
Dear Parents - do you want children to still be born in 4 generations time?
Can geneva sustain the world

NY team are to be congratulated on their goal to update where AI has got to

question 1 can an update of AI be understood without at least short summary unpdates of other tech-change movements of our era to 2023 acclerating through 1000 to 2000 moore tech than 1946- we are hoping the UB's digitalCooperation will clarify what this combo comprises- there isn't much standard language for world citizens to understand whats possible let alone vote on - and we also hope that when jack ma returns full time to education he will publish a newsletter showing his latest updates - this is synergetic with his leading connections with open source reasrach of AI - the head of unhabitat is due to publish a league table of SD-goal smart cities to start the 2020- submission case deadline march 2019- parallel reserarh has shown that 9 year old girls should be asked to audit cities for safety - as their views are more accurately demanding than most (for obvious reasons given the safety chalenges of a girl growing into adolesence- all of this connects with the future of education - see the range of edutech 

tech Belt Road - there is no better time or space to celebarte positive tech conmnecyivity when natiosn arfe newly bridged by transportation infrastructure- china recommends to the 100+ national leaders who brainstorm around different belt road trade maps- create a dual nation youth hub ahead of opening your bridge- start by celebrating ecah others languages, cultures and sharing tech moocs of a sort that encourage hackathons across nations- many of the nations that meet in annual meetings to value every way nations can cooperate ask that such new connections are launched with celebrations round the most practical masterclasses - see eg jack ma's 3000 person training format 

-his own 15 billion dollar damo academy

the parallel ispiration of wef's founder charles schwab;s IR 4 hubs linking its world inteligence hq in geneva to beijing tokyo and san francisco

the fact that jack ma's main investirs - softbank has liquified a third of its mobile business value to invest in ai; at stanford jerry yang co-founder of yahoo plays the role of guide to every digital coats most open ai researchers - chinese americans are playing the role that basic ai needs to be as open sopurce as basic genome mapping

out of tsinghua university you have a ring of partners in curricula public servants need to be future back -= ie exploring and co-creative not examining over past perfection- mit is one us partner of tsinghua; the former beijing green thinktank leader is brooking at tsinghua partner; unesco embeds itself in tsinghua foriegn langiage department - down the road is beijing normal respomsible for transforations in feminising early childhood learning across china- announced last month is that china is introducing ai as a primary school curricula- where are the wiki tools of this world- how do we keep ahead of ai  

every city with a future needs to clarify which is its most prolifuc industrial revolution 4 and ai hub- axios does quite a good job in dc and amazonuni cant come soone enough to linkin seatlle and new york tech leaders than DC inside the betway policymakers patently lacks; but what are the media modalities for keeping a back from future understanding updated at each education grade if are children are to grow up as mother earth;'s sustainability generation- in some supercities excellent pods seem to be the way passionate but relatively small resourced correspodents share with the world but for the reader this is a very inefficient way compared with transripts and with a well edited on-edmand cooperation mooc -something tricky us understanding what ways networks of infomred people use to chat- in china because wechat is nore poular than email instant meesgage theme groups can scale- and AI analysis of these expoerts fgroups conversations may produce huge leaps forward in innovation - in this mode ai is in effect the way machine intelligence can do a 1000 times more ingtegrated job than any professor or journalist or judge - 

what the world needed by now was a subset of united nations who each voluntered to assign best brains to a collaboration sustainability race arpound mother earth - now we are over 1000 times more connected with tech than racing to the moon - you can assess what population of tech and community conscious youth that needed 

some combo tools

ai -pure and as indicated above taking on integration roles that professors once believed they were best at

big data small- whiuch are the big data platforms eg nilekani's billion person digital id for rebooting india- how small are their enterprising connections


new levels of supercomputing and clouds which may be orders of magnitude faster and capable of taking on big 

g5 internet of things where any object you pass may have the same mobile connectivity as you hjave buit also be fed with its own ai and in some cases its own actions (inclduing sending artioficaislk to arrest you?)

3d printing ie augmented engineering- in the past to make something you had to chissel out, now scanners can make things by adding in- potentially making copying of any scale economic-  Fab City Global Initiative  -augmented reality googles now ;permit schools across the world to share what its like to live in each others communities or extended classrooms where most livelihoods are now prepped

Jack Ma, Founder of Alibaba Group and an Education Commission Commissioner, shares his thoughts on why L-Q is most ...

Founder and Executive Chairman Alibaba Group and Commissioner Jack Ma sits down with The International Commission on ... 

at least a decade ago mit replicated fab labs across several smart cities including shenzen- this linked technologies at a more personal level: 2d , etextiles ie sensors a person can wear either to communicate with the outside or to augment senses- a blind person can now be offered spectacles which enable the person to see in the same way that a robot "sees"; this is part of an area with various names like open source medicine

cyber and other risks of many thousand times more connecetd as everyone's responsibility 


Meredith Whittaker , AI Now Institute, New York University, Google Open Research
Kate Crawford , AI Now Institute, New York University, Microsoft Research
Roel Dobbe , AI Now Institute, New York University
Genevieve Fried , AI Now Institute, New York University
Elizabeth Kaziunas , AI Now Institute, New York University
Varoon Mathur , AI Now Institute, New York University
Sarah Myers West , AI Now Institute, New York University
Rashida Richardson , AI Now Institute, New York University
Jason Schultz , AI Now Institute, New York University School of Law
Oscar Schwartz , AI Now Institute, New York University
With research assistance from Alex Campolo and Gretchen Krueger (AI Now Institute, New York
1.1 AI is Amplifying Widespread Surveillance 12
The faulty science and dangerous history of affect recognition 13
Facial recognition amplifies civil rights concerns 15
1.2 The Risks of Automated Decision Systems in Government 18
1.3 Experimenting on Society: Who Bears the Burden? 22
2.1 Bias Busting and Formulas for Fairness: the Limits of Technological “Fixes” 24
Broader approaches 27
2.2 Industry Applications: Toolkits and System Tweaks 28
2.3 Why Ethics is Not Enough 29
3.1 From Fairness to Justice 32
3.2 Infrastructural Thinking 33
3.3 Accounting for Hidden Labor in AI Systems 34
3.4 Deeper Interdisciplinarity 36
3.5 Race, Gender and Power in AI 37
3.6 Strategic Litigation and Policy Interventions 39
3.7 Research and Organizing: An Emergent Coalition 40
This work is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License
The AI Now Institute at New York University is an interdisciplinary research institute dedicated to
understanding the social implications of AI technologies. It is the first university research center
focused specifically on AI’s social significance. Founded and led by Kate Crawford and Meredith
Whittaker, AI Now is one of the few women-led AI institutes in the world.
AI Now works with a broad coalition of stakeholders, including academic researchers, industry,
civil society, policy makers, and affected communities, to identify and address issues raised by
the rapid introduction of AI across core social domains. AI Now produces interdisciplinary
research to help ensure that AI systems are accountable to the communities and contexts they
are meant to serve, and that they are applied in ways that promote justice and equity. The
Institute’s current research agenda focuses on four core areas: bias and inclusion, rights and
liberties, labor and automation, and safety and critical infrastructure.
Our most recent publications include:
● Litigating Algorithms , a major report assessing recent court cases focused on
government use of algorithms
● Anatomy of an AI System , a large-scale map and longform essay produced in partnership
with SHARE Lab , which investigates the human labor, data, and planetary resources
required to operate an Amazon Echo
● Algorithmic Impact Assessment (AIA) Report , which helps affected communities and
stakeholders assess the use of AI and algorithmic decision-making in public agencies
● Algorithmic Accountability Policy Toolkit , which is geared toward advocates interested
in understanding government use of algorithmic systems
We also host expert workshops and public events on a wide range of topics. Our workshop on
Immigration, Data, and Automation in the Trump Era , co-hosted with the Brennan Center for
Justice and the Center for Privacy and Technology at Georgetown Law, focused on the Trump
Administration’s use of data harvesting, predictive analytics, and machine learning to target
immigrant communities. The Data Genesis Working Group convenes experts from across
industry and academia to examine the mechanics of dataset provenance and maintenance. Our
roundtable on Machine Learning, Inequality and Bias , co-hosted in Berlin with the Robert Bosch
Academy, gathered researchers and policymakers from across Europe to address issues of bias,
discrimination, and fairness in machine learning and related technologies.
Our annual public symposium convenes leaders from academia, industry, government, and civil
society to examine the biggest challenges we face as AI moves into our everyday lives. The AI
Now 2018 Symposium addressed the intersection of AI ethics, organizing, and accountability,
examining the landmark events of the past year. Over 1,000 people registered for the event, which
was free and open to the public. Recordings of the program are available on our website .
More information is available at .
1. Governments need to regulate AI by expanding the powers of sector-specific agencies to
oversee, audit, and monitor these technologies by domain. The implementation of AI
systems is expanding rapidly, without adequate governance, oversight, or accountability
regimes. Domains like health, education, criminal justice, and welfare all have their own
histories, regulatory frameworks, and hazards. However, a national AI safety body or general
AI standards and certification model will struggle to meet the sectoral expertise requirements
needed for nuanced regulation. We need a sector-specific approach that does not prioritize
the technology, but focuses on its application within a given domain. Useful examples of
sector-specific approaches include the United States Federal Aviation Administration and the
National Highway Traffic Safety Administration.
2. Facial recognition and affect recognition need stringent regulation to protect the public
interest. Such regulation should include national laws that require strong oversight, clear
limitations, and public transparency. Communities should have the right to reject the
application of these technologies in both public and private contexts. Mere public notice of
their use is not sufficient, and there should be a high threshold for any consent, given the
dangers of oppressive and continual mass surveillance. Affect recognition deserves particular
attention. Affect recognition is a subclass of facial recognition that claims to detect things
such as personality, inner feelings, mental health, and “worker engagement” based on images
or video of faces. These claims are not backed by robust scientific evidence, and are being
applied in unethical and irresponsible ways that often recall the pseudosciences of phrenology
and physiognomy. Linking affect recognition to hiring, access to insurance, education, and
policing creates deeply concerning risks, at both an individual and societal level.
3. The AI industry urgently needs new approaches to governance. As this report
demonstrates, internal governance structures at most technology companies are failing to
ensure accountability for AI systems. Government regulation is an important component,
but leading companies in the AI industry also need internal accountability structures that go
beyond ethics guidelines. This should include rank-and-file employee representation on the
board of directors, external ethics advisory boards, and the implementation of independent
monitoring and transparency efforts. Third party experts should be able to audit and publish
about key systems, and companies need to ensure that their AI infrastructures can be
understood from “nose to tail,” including their ultimate application and use.
4. AI companies should waive trade secrecy and other legal claims that stand in the way of
accountability in the public sector. Vendors and developers who create AI and automated
decision systems for use in government should agree to waive any trade secrecy or other
legal claim that inhibits full auditing and understanding of their software. Corporate secrecy
laws are a barrier to due process: they contribute to the “black box effect” rendering systems
opaque and unaccountable, making it hard to assess bias, contest decisions, or remedy
errors. Anyone procuring these technologies for use in the public sector should demand that
vendors waive these claims before entering into any agreements.
5. Technology companies should provide protections for conscientious objectors, employee
organizing, and ethical whistleblowers. Organizing and resistance by technology workers
has emerged as a force for accountability and ethical decision making. Technology
companies need to protect workers’ ability to organize, whistleblow, and make ethical choices
about what projects they work on. This should include clear policies accommodating and
protecting conscientious objectors, ensuring workers the right to know what they are working
on, and the ability to abstain from such work without retaliation or retribution. Workers raising
ethical concerns must also be protected, as should whistleblowing in the public interest.
6. Consumer protection agencies should apply “truth-in-advertising” laws to AI products and
services. The hype around AI is only growing, leading to widening gaps between marketing
promises and actual product performance. With these gaps come increasing risks to both
individuals and commercial customers, often with grave consequences. Much like other
products and services that have the potential to seriously impact or exploit populations, AI
vendors should be held to high standards for what they can promise, especially when the
scientific evidence to back these promises is inadequate and the longer-term consequences
are unknown.
7. Technology companies must go beyond the “pipeline model” and commit to addressing the
practices of exclusion and discrimination in their workplaces. Technology companies and
the AI field as a whole have focused on the “pipeline model,” looking to train and hire more
diverse employees. While this is important, it overlooks what happens once people are hired
into workplaces that exclude, harass, or systemically undervalue people on the basis of
gender, race, sexuality, or disability. Companies need to examine the deeper issues in their
workplaces, and the relationship between exclusionary cultures and the products they build,
which can produce tools that perpetuate bias and discrimination. This change in focus needs
to be accompanied by practical action, including a commitment to end pay and opportunity
inequity, along with transparency measures about hiring and retention.
8. Fairness, accountability, and transparency in AI require a detailed account of the “full stack
supply chain.” For meaningful accountability, we need to better understand and track the
component parts of an AI system and the full supply chain on which it relies: that means
accounting for the origins and use of training data, test data, models, application program
interfaces (APIs), and other infrastructural components over a product life cycle. We call this
accounting for the “full stack supply chain” of AI systems, and it is a necessary condition for a
more responsible form of auditing. The full stack supply chain also includes understanding
the true environmental and labor costs of AI systems. This incorporates energy use, the use of
labor in the developing world for content moderation and training data creation, and the
reliance on clickworkers to develop and maintain AI systems.
9. More funding and support are needed for litigation, labor organizing, and community
participation on AI accountability issues. The people most at risk of harm from AI systems
are often those least able to contest the outcomes. We need increased support for robust
mechanisms of legal redress and civic participation. This includes supporting public
advocates who represent those cut off from social services due to algorithmic decision
making, civil society organizations and labor organizers that support groups that are at risk of
job loss and exploitation, and community-based infrastructures that enable public
10. University AI programs should expand beyond computer science and engineering
disciplines. AI began as an interdisciplinary field, but over the decades has narrowed to
become a technical discipline. With the increasing application of AI systems to social
domains, it needs to expand its disciplinary orientation. That means centering forms of
expertise from the social and humanistic disciplines. AI efforts that genuinely wish to address
social implications cannot stay solely within computer science and engineering departments,
where faculty and students are not trained to research the social world. Expanding the
disciplinary orientation of AI research will ensure deeper attention to social contexts, and
more focus on potential hazards when these systems are applied to human populations.
At the core of the cascading scandals around AI in 2018 are questions of accountability: who is
responsible when AI systems harm us? How do we understand these harms, and how do we
remedy them? Where are the points of intervention, and what additional research and regulation is
needed to ensure those interventions are effective? Currently there are few answers to these
questions, and the frameworks presently governing AI are not capable of ensuring accountability.
As the pervasiveness, complexity, and scale of these systems grow, the lack of meaningful
accountability and oversight – including basic safeguards of responsibility, liability, and due
process – is an increasingly urgent concern.
Building on our 2016 and 2017 reports, the AI Now 2018 Report contends with this central
problem and addresses the following key issues:
1. The growing accountability gap in AI, which favors those who create and deploy these
technologies at the expense of those most affected
2. The use of AI to maximize and amplify surveillance, especially in conjunction with facial
and affect recognition, increasing the potential for centralized control and oppression
3. Increasing government use of automated decision systems that directly impact
individuals and communities without established accountability structures
4. Unregulated and unmonitored forms of AI experimentation on human populations
5. The limits of technological solutions to problems of fairness, bias, and discrimination
Within each topic, we identify emerging challenges and new research, and provide
recommendations regarding AI development, deployment, and regulation. We offer practical
pathways informed by research so that policymakers, the public, and technologists can better
understand and mitigate risks. Given that the AI Now Institute’s location and regional expertise is
concentrated in the U.S., this report will focus primarily on the U.S. context, which is also where
several of the world’s largest AI companies are based.
The AI accountability gap is growing: The technology scandals of 2018 have shown that the gap
between those who develop and profit from AI—and those most likely to suffer the consequences
of its negative effects—is growing larger, not smaller. There are several reasons for this, including
a lack of government regulation, a highly concentrated AI sector, insufficient governance
structures within technology companies, power asymmetries between companies and the people
they serve, and a stark cultural divide between the engineering cohort responsible for technical
research, and the vastly diverse populations where AI systems are deployed. These gaps are
producing growing concern about bias, discrimination, due process, liability, and overall
responsibility for harm. This report emphasizes the urgent need for stronger, sector-specific
research and regulation.
AI is amplifying widespread surveillance: The role of AI in widespread surveillance has expanded
immensely in the U.S., China, and many other countries worldwide. This is seen in the growing use
of sensor networks, social media tracking, facial recognition, and affect recognition. These
expansions not only threaten individual privacy, but accelerate the automation of surveillance, and
thus its reach and pervasiveness. This presents new dangers, and magnifies many longstanding
concerns. The use of affect recognition, based on debunked pseudoscience, is also on the rise.
Affect recognition attempts to read inner emotions by a close analysis of the face and is
connected to spurious claims about people’s mood, mental health, level of engagement, and guilt
or innocence. This technology is already being used for discriminatory and unethical purposes,
often without people’s knowledge. Facial recognition technology poses its own dangers,
reinforcing skewed and potentially discriminatory practices, from criminal justice to education to
employment, and presents risks to human rights and civil liberties in multiple countries.
Governments are rapidly expanding the use of automated decision systems without adequate
protections for civil rights: Around the world, government agencies are procuring and deploying
automated decision systems (ADS) under the banners of efficiency and cost-savings. Yet many of
these systems are untested and poorly designed for their tasks, resulting in illegal and often
unconstitutional violations of individual rights. Worse, when they make errors and bad decisions,
the ability to question, contest, and remedy these is often difficult or impossible. Some agencies
are attempting to provide mechanisms for transparency, due process, and other basic rights, but
trade secrecy and similar laws threaten to prevent auditing and adequate testing of these
systems. Drawing from proactive agency efforts, and from recent strategic litigation, we outline
pathways for ADS accountability.
Rampant testing of AI systems “in the wild” on human populations: Silicon Valley is known for
its “move fast and break things” mentality, whereby companies are pushed to experiment with
new technologies quickly and without much regard for the impact of failures, including who bears
the risk. In the past year, we have seen a growing number of experiments deploying AI systems “in
the wild” without proper protocols for notice, consent, or accountability. Such experiments
continue, due in part to a lack of consequences for failure. When harms occur, it is often unclear
where or with whom the responsibility lies. Researching and assigning appropriate responsibility
and liability remains an urgent priority.
The limits of technological fixes to problems of fairness, bias, and discrimination: Much new
work has been done designing mathematical models for what should be considered “fair” when
machines calculate outcomes, aimed at avoiding discrimination. Yet, without a framework that
accounts for social and political contexts and histories, these mathematical formulas for fairness
will almost inevitably miss key factors, and can serve to paper over deeper problems in ways that
ultimately increase harm or ignore justice. Broadening perspectives and expanding research into
AI fairness and bias beyond the merely mathematical is critical to ensuring we are capable of
addressing the core issues and moving the focus from parity to justice.
The move to ethical principles: This year saw the emergence of numerous ethical principles and
guidelines for the creation and deployment of AI technologies, many in response to growing
concerns about AI’s social implications. But as studies show, these types of ethical commitments
have little measurable effect on software development practices if they are not directly tied to
structures of accountability and workplace practices. Further, these codes and guidelines are
rarely backed by enforcement, oversight, or consequences for deviation. Ethical codes can only
help close the AI accountability gap if they are truly built into the processes of AI development and
are backed by enforceable mechanisms of responsibility that are accountable to the public
The following report develops these themes in detail, reflecting on the latest academic research,
and outlines seven strategies for moving forward:
1. Expanding AI fairness research beyond a focus on mathematical parity and statistical
fairness toward issues of justice
2. Studying and tracking the full stack of infrastructure needed to create AI, including
accounting for material supply chains
3. Accounting for the many forms of labor required to create and maintain AI systems
4. Committing to deeper interdisciplinarity in AI
5. Analyzing race, gender, and power in AI
6. Developing new policy interventions and strategic litigation
7. Building coalitions between researchers, civil society, and organizers within the technology
These approaches are designed to positively recast the AI field and address the growing power
imbalance that currently favors those who develop and profit from AI systems at the expense of
the populations most likely to be harmed.
The Social Challenges of AI in 2018
The past year has seen accelerated integration of powerful artificial intelligence systems into core
social institutions, against a backdrop of rising inequality, political populism, and industry
scandals. 1 There have been major movements from both inside and outside technology
companies pushing for greater accountability and justice. The AI Now 2018 Report focuses on
these themes and examines the gaps between AI ethics and meaningful accountability, and the
role of organizing and regulation.
In short, it has been a dramatic year in AI. In any normal year, Cambridge Analytica seeking to
manipulate national elections in the US and UK using social media data and algorithmic ad
targeting would have been the biggest story. 2 But in 2018, it was just one of many scandals.
Facebook had a series of disasters, including a massive data breach in September, 3 multiple class
action lawsuits for discrimination, 4 accusations of inciting ethnic cleansing in Myanmar, 5 potential
violations of the Fair Housing Act, 6 and hosting masses of fake Russian accounts. 7 Throughout
the year, the company’s executives were frequently summoned to testify, with Mark Zuckerberg
facing the US Senate in April and the European Parliament in May. 8 Zuckerberg mentioned AI
technologies over 30 times in his Congressional testimony as the cure-all to the company’s
problems, particularly in the complex areas of censorship, fairness, and content moderation. 9
But Facebook wasn’t the only one in crisis. News broke in March that Google was building AI
systems for the Department of Defense’s drone surveillance program, Project Maven. 10 The news
kicked off an unprecedented wave of technology worker organizing and dissent across the
industry. 11 In June, when the Trump administration introduced the family separation policy that
forcibly removed immigrant children from their parents, employees from Amazon, Salesforce, and
Microsoft all asked their companies to end contracts with U.S. Immigration and Customs
Enforcement (ICE). 12 Less than a month later, it was revealed that ICE modified its own risk
assessment algorithm so that it could only produce one result: the system recommended “detain”
for 100% of immigrants in custody. 13
Throughout the year, AI systems continued to be tested on live populations in high-stakes
domains, with some serious consequences. In March, autonomous cars killed drivers and
pedestrians. 14 Then in May, a voice recognition system in the UK designed to detect immigration
fraud ended up cancelling thousands of visas and deporting people in error. 15 Documents leaked
in July showed that IBM Watson was producing “unsafe and incorrect” cancer treatment
recommendations. 16 And an investigation in September revealed that IBM was also working with
the New York City Police Department (NYPD) to build an “ethnicity detection” feature to search
faces based on race, using police camera footage of thousands of people in the streets of New
York taken without their knowledge or permission. 17
This is just a sampling of an extraordinary series of incidents from 2018. 18 The response has
included a growing wave of criticism, with demands for greater accountability from the
technology industry and the systems they build. 19 In turn, some companies have made public
calls for the U.S. to regulate technologies like facial recognition. 20 Others have published AI ethics
principles and increased efforts to produce technical fixes for issues of bias and discrimination in
AI systems. But many of these ethical and technical approaches define the problem space very
narrowly, neither contending with the historical or social context nor providing mechanisms for
public accountability, oversight, and due process. This makes it nearly impossible for the public to
validate that any of the current problems have, in fact, been addressed.
As numerous scholars have noted, one significant barrier to accountability is the culture of
industrial and legal secrecy that dominates AI development. 21 Just as many AI technologies are
“black boxes”, so are the industrial cultures that create them. 22 Many of the fundamental building
blocks required to understand AI systems and to ensure certain forms of accountability – from
training data, to data models, to the code dictating algorithmic functions, to implementation
guidelines and software, to the business decisions that directed design and development – are
rarely accessible to review, hidden by corporate secrecy laws.
The current accountability gap is also caused by the incentives driving the rapid pace of technical
AI research. The push to “innovate,” publish first, and present a novel addition to the technical
domain has created an accelerated cadence in the field of AI, and in technical disciplines more
broadly. This comes at the cost of considering empirical questions of context and use, or
substantively engaging with ethical concerns. 23 Similarly, technology companies are driven by
pressures to “launch and iterate,” which assume complex social and political questions will be
handled by policy and legal departments, leaving developers and sales departments free from the
responsibility of considering the potential downsides. The “move fast and break things” culture
provides little incentive for ensuring meaningful public accountability or engaging the
communities most likely to experience harm. 24 This is particularly problematic as the accelerated
application of AI systems in sensitive social and political domains presents risks to marginalized
The challenge to create better governance and greater accountability for AI poses particular
problems when such systems are woven into the fabric of government and public institutions.
The lack of transparency, notice, meaningful engagement, accountability, and oversight creates
serious structural barriers for due process and redress for unjust and discriminatory decisions.
In this year’s report, we assess many pressing issues facing us as AI tools are deployed further
into the institutions that govern everyday life. We focus on the biggest industry players, because
the number of companies able to create AI at scale is very small, while their power and reach is
global. We evaluate the current range of responses from industry, governments, researchers,
activists, and civil society at large. We suggest a series of substantive approaches and make ten
specific recommendations. Finally, we share the latest research and policy strategies that can
contribute to greater accountability, as well as a richer understanding of AI systems in a wider
social context.
In identifying the most pressing social implications of AI this year, we look closely at the role of AI
in widespread surveillance in multiple countries around the world, and at the implications for
rights and liberties. In particular, we consider the increasing use of facial recognition, and a
subclass of facial recognition known as affect recognition, and assess the growing calls for
regulation. Next, we share our findings on the government use of automated decision systems,
and what questions this raises for fairness, transparency, and due process when such systems
are protected by trade secrecy and other laws that prevent auditing and close examination. 25
Finally, we look at the practices of deploying experimental systems “in the wild,” testing them on
human populations. We analyze who has the most to gain, and who is at greatest risk of
experiencing harm.
1.1 AI is Amplifying Widespread Surveillance
This year, we have seen AI amplify large-scale surveillance through techniques that analyze video,
audio, images, and social media content across entire populations and identify and target
individuals and groups. While researchers and advocates have long warned about the dangers of
mass data collection and surveillance, 26 AI raises the stakes in three areas: automation, scale of
analysis, and predictive capacity. Specifically, AI systems allow automation of surveillance
capabilities far beyond the limits of human review and hand-coded analytics. Thus, they can serve
to further centralize these capabilities in the hands of a small number of actors. These systems
also exponentially scale analysis and tracking across large quantities of data, attempting to make
connections and inferences that would have been difficult or impossible before their introduction.
Finally, they provide new predictive capabilities to make determinations about individual character
and risk profiles, raising the possibility of granular population controls.
China has offered several examples of alarming AI-enabled surveillance this year, which we know
about largely because the government openly acknowledges them. However, it’s important to
note that many of the same infrastructures already exist in the U.S. and elsewhere, often
produced and promoted by private companies whose marketing emphasizes beneficial use
cases. In the U.S. the use of these tools by law enforcement and government is rarely open to
public scrutiny, as we will review, and there is much we do not know. Such infrastructures and
capabilities could easily be turned to more surveillant ends in the U.S., without public disclosure
and oversight, depending on market incentives and political will.
In China, military and state-sanctioned automated surveillance technology is being deployed to
monitor large portions of the population, often targeting marginalized groups. Reports include
installation of facial recognition tools at the Hong Kong-Shenzhen border, 27 using flocks of robotic
dove-like drones in five provinces across the country, 28 and the widely reported social credit
monitoring system, 29 each of which illustrates how AI-enhanced surveillance systems can be
mobilized as a means of far-reaching social control. 30
The most oppressive use of these systems is reportedly occuring in the Xinjiang Autonomous
Region, described by The Economist as a “police state like no other.” 31 Surveillance in this Uighur
ethnic minority area is pervasive, ranging from physical checkpoints and programs where Uighur
households are required to “adopt” Han Chinese officials into their family, to the widespread use of
surveillance cameras, spyware, Wi-Fi sniffers, and biometric data collection, sometimes by
stealth. Machine learning tools integrate these streams of data to generate extensive lists of
suspects for detention in re-education camps, built by the government to discipline the group.
Estimates of the number of people detained in these camps range from hundreds of thousands to
nearly one million. 32
These infrastructures are not unique to China. Venezuela announced the adoption of a new smart
card ID known as the “carnet de patria,” which, by integrating government databases linked to
social programs, could enable the government to monitor citizens’ personal finances, medical
history, and voting activity. 33 In the United States, we have seen similar efforts. The Pentagon has
funded research on AI-enabled social media surveillance to help predict large-scale population
behaviors, 34 and the U.S. Immigration and Customs Enforcement (ICE) agency is using an
Investigative Case Management System developed by Palantir and powered by Amazon Web
Services in its deportation operations. 35 The system integrates public data with information
purchased from private data brokers to create profiles of immigrants in order to aid the agency in
profiling, tracking, and deporting individuals. 36 These examples show how AI systems increase
integration of surveillance technologies into data-driven models of social control and amplify the
power of such data, magnifying the stakes of misuse and raising urgent and important questions
as to how basic rights and liberties will be protected.
The faulty science and dangerous history of affect recognition
We are also seeing new risks emerging from unregulated facial recognition systems. These
systems facilitate the detection and recognition of individual faces in images or video, and can be
used in combination with other tools to conduct more sophisticated forms of surveillance, such
as automated lip-reading, offering the ability to observe and interpret speech from a distance. 37
Among a host of AI-enabled surveillance and tracking techniques, facial recognition raises
particular civil liberties concerns. Because facial features are a very personal form of biometric
identification that is extremely difficult to change, it is hard to subvert or “opt out” of its operations.
And unlike other tracking tools, facial recognition seeks to use AI for much more than simply
recognizing faces. Once identified, a face can be linked with other forms of personal records and
identifiable data, such as credit score, social graph, or criminal record.
Affect recognition, a subset of facial recognition, aims to interpret faces to automatically detect
inner emotional states or even hidden intentions. This approach promises a type of emotional
weather forecasting: analyzing hundreds of thousands of images of faces, detecting
“micro-expressions,” and mapping these expressions to “true feelings.” 38 This reactivates a long
tradition of physiognomy – a pseudoscience that claims facial features can reveal innate aspects
of our character or personality. Dating from ancient times, scientific interest in physiognomy grew
enormously in the nineteenth century, when it became a central method for scientific forms of
racism and discrimination. 39 Although physiognomy fell out of favor following its association with
Nazi race science, researchers are worried about a reemergence of physiognomic ideas in affect
recognition applications. 40 The idea that AI systems might be able to tell us what a student, a
customer, or a criminal suspect is really feeling or what type of person they intrinsically are is
proving attractive to both corporations and governments, even though the scientific justifications
for such claims are highly questionable, and the history of their discriminatory purposes
The case of affect detection reveals how machine learning systems can easily be used to
intensify forms of classification and discrimination, even when the basic foundations of these
theories remain controversial among psychologists. The scientist most closely associated with
AI-enabled affect detection is the psychologist Paul Ekman, who asserted that emotions can be
grouped into a small set of basic categories like anger, disgust, fear, happiness, sadness, and
surprise. 41 Studying faces, according to Ekman, produces an objective reading of authentic
interior states—a direct window to the soul. Underlying his belief was the idea that emotions are
fixed and universal, identical across individuals, and clearly visible in observable biological
mechanisms regardless of cultural context. But Ekman’s work has been deeply criticized by
psychologists, anthropologists, and other researchers who have found his theories do not hold up
under sustained scrutiny. 42 The psychologist Lisa Feldman Barrett and her colleagues have
argued that an understanding of emotions in terms of these rigid categories and simplistic
physiological causes is no longer tenable. 43 Nonetheless, AI researchers have taken his work as
fact, and used it as a basis for automating emotion detection. 44
Contextual, social, and cultural factors — how, where, and by whom such emotional signifiers are
expressed — play a larger role in emotional expression than was believed by Ekman and his peers.
In light of this new scientific understanding of emotion, any simplistic mapping of a facial
expression onto basic emotional categories through AI is likely to reproduce the errors of an
outdated scientific paradigm. It also raises troubling ethical questions about locating the arbiter of
someone’s “real” character and emotions outside of the individual, and the potential abuse of
power that can be justified based on these faulty claims. Psychiatrist Jamie Metzl documents a
recent cautionary example: a pattern in the 1960s of diagnosing Black people with schizophrenia
if they supported the civil rights movement. 45 Affect detection combined with large-scale facial
recognition has the potential to magnify such political abuses of psychological profiling.
In the realm of education, some U.S. universities have announced plans to use affect analysis
software on students. 46 The University of St. Thomas, in Minnesota, is using a system based on
Microsoft’s facial recognition and affect detection tools to observe students in the classroom
using a webcam. The system predicts the students’ emotional state. An overview of student
sentiment is viewable by the teacher, who can then shift their teaching in a way that “ensures
student engagement,” as judged by the system. This raises serious questions on multiple levels:
what if the system, with a simplistic emotional model, simply cannot grasp more complex states?
How would a student contest a determination made by the system? What if different students are
seen as “happy” while others are “angry”—how should the teacher redirect the lesson? What are
the privacy implications of such a system, particularly given that, in the case of the pilot program,
there is no evidence that students were informed of its use on them?
Outside of the classroom, we are also seeing personal assistants, like Alexa and Siri, seeking to
pick up on the emotional undertones of human speech, with companies even going so far as to
patent methods of marketing based on detecting emotions, as well as mental and physical
health. 47 The AI-enabled emotion measurement company Affectiva now promises it can promote
safer driving by monitoring “driver and occupant emotions, cognitive states, and reactions to the
driving experience...from face and voice.” 48 Yet there is little evidence that any of these systems
actually work across different individuals, contexts, and cultures, or have any safeguards put in
place to mitigate concerns about privacy, bias, or discrimination in their operation. Furthermore,
as we have seen in the large literature on bias and fairness, classifications of this nature not only
have direct impacts on human lives, but also serve as data to train and influence other AI
systems. This raises the stakes for any use of affect recognition, further emphasizing why it
should be critically examined and its use severely restricted.
Facial recognition amplifies civil rights concerns
Concerns are intensifying that facial recognition increases racial discrimination and other biases
in the criminal justice system. Earlier this year, the American Civil Liberties Union (ACLU)
disclosed that both the Orlando Police Department and the Washington County Sheriff’s
department were using Amazon’s Rekognition system, which boasts that it can perform “real-time
face recognition across tens of millions of faces” and detect “up to 100 faces in challenging
crowded photos.” 49 In Washington County, Amazon specifically worked with the Sheriff’s
department to create a mobile app that could scan faces and compare them against a database
of at least 300,000 mugshots. 50 An Amazon representative recently revealed during a talk that
they have been considering applications where Orlando’s network of surveillance cameras could
be used in conjunction with facial recognition technology to find a “person of interest” wherever
they might be in the city. 51
In addition to the privacy and mass surveillance concerns commonly raised, the use of facial
recognition in law enforcement has also intersected with concerns of racial and other biases.
Researchers at the ACLU and the University of California (U.C.) Berkeley tested Amazon’s
Rekognition tool by comparing the photos of sitting members in the United States Congress with
a database containing 25,000 photos of people who had been arrested. The results showed
significant levels of inaccuracy: Amazon’s Rekognition incorrectly identified 28 members of
Congress as people from the arrest database. Moreover, the false positives disproportionately
occurred among non-white members of Congress, with an error rate of nearly 40% compared to
only 5% for white members. 52 Such results echo a string of findings that have demonstrated that
facial recognition technology is, on average, better at detecting light-skinned people than
dark-skinned people, and better at detecting men than women. 53
In its response to the ACLU, Amazon acknowledged that “the Rekognition results can be
significantly skewed by using a facial database that is not appropriately representative.” 54 Given
the deep and historical racial biases in the criminal justice system, most law enforcement
databases are unlikely to be “appropriately representative.” 55 Despite these serious flaws, ongoing
pressure from civil rights groups, and protests from Amazon employees over the potential for
misuse of these technologies, Amazon Web Services CEO Andrew Jassy recently told employees
that “we feel really great and really strongly about the value that Amazon Rekognition is providing
our customers of all sizes and all types of industries in law enforcement and out of law
enforcement.” 56
Nor is Amazon alone in implementing facial recognition technologies in unaccountable ways.
Investigative journalists recently disclosed that IBM and the New York City Police Department
(NYPD) partnered to develop such a system that included “ethnicity search” as a custom feature,
trained on thousands of hours of NYPD surveillance footage. 57 Use of facial recognition software
in the private sector has expanded as well. 58 Major retailers and venues have already begun using
these technologies to detect shoplifters, monitor crowds, and even “scan for unhappy customers,”
using facial recognition systems instrumented with “affect detection” capabilities. 59
These concerns are amplified by a lack of laws and regulations. There is currently no federal
legislation that seeks to provide standards, restrictions, requirements, or guidance regarding the
development or use of facial recognition technology. In fact, most existing federal legislation
looks to promote the use of facial recognition for surveillance, immigration enforcement,
employment verification, and domestic entry-exit systems. 60 The laws that we do have are
piecemeal, and none specifically address facial recognition. Among these is the Biometric
Information Privacy Act, a 2008 Illinois law that sets forth stringent rules regarding the collection
of biometrics. While the law does not mention facial recognition, given that the technology was
not widely available in 2008, many of its requirements, such as obtaining consent, are reasonably
interpreted to apply. 61 More recently, several municipalities and a local transit system have
adopted ordinances that seek to create greater transparency and oversight of data collection and
use requirements regarding the acquisition of surveillance technologies, which would include
facial recognition based on the expansive definition in these ordinances. 62
Opposition to the use of facial recognition tools by government agencies is growing. Earlier this
year, AI Now joined the ACLU and over 30 other research and advocacy organizations calling on
Amazon to stop selling facial recognition software to government agencies after the ACLU
uncovered documents showing law enforcement use of Amazon’s Rekognition API. 63 Members of
Congress are also pushing Amazon to provide more information. 64
Some have gone further, calling for an outright ban. Scholars Woodrow Hartzog and Evan Selinger
argue that facial recognition technology is a “tool for oppression that’s perfectly suited for
governments to display unprecedented authoritarian control and an all-out privacy-eviscerating
machine,” necessitating extreme caution and diligence before being applied in our contemporary
digital ecosystem. 65 Critiquing the Stanford “gaydar” study that claimed its deep neural network
was more accurate than humans at predicting sexuality from facial images, 66 Frank Pasquale
wrote that “there are some scientific research programs best not pursued - and this might be one
of them.” 67
Kade Crockford, Director of the Technology for Liberty Program at ACLU of Massachusetts, also
wrote in favor of a ban, stating that “artificial intelligence technologies like face recognition
systems fundamentally change the balance of power between the people and the
government...some technologies are so dangerous to that balance of power that they must be
rejected.” 68 Microsoft President Brad Smith has called for government regulation of facial
recognition, while Rick Smith, CEO of law enforcement technology company Axon, recently stated
that the “accuracy thresholds” of facial recognition tools aren’t “where they need to be to be
making operational decisions.” 69
The events of this year have strongly underscored the urgent need for stricter regulation of both
facial and affect recognition technologies. Such regulations should severely restrict use by both
the public and the private sector, and ensure that communities affected by these technologies are
the final arbiters of whether they are used at all. This is especially important in situations where
basic rights and liberties are at risk, requiring stringent oversight, audits, and transparency.
Linkages should not be permitted between private and government databases. At this point, given
the evidence in hand, policymakers should not be funding or furthering the deployment of these
systems in public spaces.
1.2 The Risks of Automated Decision Systems in
Over the past year, we have seen a substantial increase in the adoption of Automated Decision
Systems (ADS) across government domains, including criminal justice, child welfare, education,
and immigration. Often adopted under the theory that they will improve government efficiency or
cost-savings, ADS seek to aid or replace various decision-making processes and policy
determinations. However, because the underlying models are often proprietary and the systems
frequently untested before deployment, many community advocates have raised significant
concerns about lack of due process, accountability, community engagement, and auditing. 70
Such was the case for Tammy Dobbs, who moved to Arkansas in 2008 and signed up for a state
disability program to help her with her cerebral palsy. 71 Under the program, the state sent a
qualified nurse to assess Tammy to determine the number of caregiver hours she would need.
Because Tammy spent most of her waking hours in a wheelchair and had stiffness in her hands,
her initial assessment allocated 56 hours of home care per week. Fast forward to 2016, when the
state assessor arrived with a new ADS on her laptop. Using a proprietary algorithm, this system
calculated the number of hours Tammy would be allotted. Without any explanation or opportunity
for comment, discussion, or reassessment, the program allotted Tammy 32 hours per week, a
massive and sudden drop that Tammy had no chance to prepare for and that severely reduced
her quality of life.
Nor was Tammy’s situation exceptional. According to Legal Aid of Arkansas attorney Kevin De
Liban, hundreds of other individuals with disabilities also received dramatic reductions in hours, all
without any meaningful opportunity to understand or contest their allocations. Legal Aid
subsequently sued the State of Arkansas, eventually winning a ruling that the new algorithmic
allocation program was erroneous and unconstitutional. Yet by then, much of the damage to the
lives of those affected had been done. 72
The Arkansas disability cases provide a concrete example of the substantial risks that occur
when governments use ADS in decisions that have immediate impacts on vulnerable populations.
While individual assessors may also suffer from bias or flawed logic, the impact of their
case-by-case decisions has nowhere near the magnitude or scale that a single flawed ADS can
have across an entire population.
The increased introduction of such systems comes at a time when, according to the World
Income Inequality Database, the United States has the highest income inequality rate of all
western countries. 73 Moreover, Federal Reserve data shows wealth inequalities continue to grow,
and racial wealth disparities have more than tripled in the last 50 years, with current policies set to
exacerbate such problems. 74 In 2018 alone, we have seen a U.S. executive order cutting funding
for social programs that serve the country’s poorest citizens, 75 alongside a proposed federal
budget that will significantly reduce low-income and affordable housing, 76 the implementation of
onerous work requirements for Medicaid, 77 and a proposal to cut food assistance benefits for
low-income seniors and people with disabilities. 78
In the context of such policies, agencies are under immense pressure to cut costs, and many are
looking to ADS as a means of automating hard decisions that have very real effects on those
most in need. 79 As such, many ADS systems are often implemented with the goal of doing more
with less in the context of austerity policies and cost-cutting. They are frequently designed and
configured primarily to achieve these goals, with their ultimate effectiveness being evaluated
based on their ability to trim costs, often at the expense of the populations such tools are
ostensibly intended to serve. 80 As researcher Virginia Eubanks argues, “What seems like an effort
to lower program barriers and remove human bias often has the opposite effect, blocking
hundreds of thousands of people from receiving the services they deserve.” 81
When these problems arise, they are frequently difficult to remedy. Few ADS are designed or
implemented in ways that easily allow affected individuals to contest, mitigate, or fix adverse or
incorrect decisions. Additionally, human discretion and the ability to intervene or override a
system’s determination is often substantially limited or removed from case managers, social
workers, and others trained to understand the context and nuance of a particular person and
situation. 82 These front-line workers become mere intermediaries, communicating inflexible
decisions made by automated systems, without the ability to alter them.
Unlike the civil servants who have historically been responsible for such decisions, many ADS
come from private vendors and are frequently implemented without thorough testing, review, or
auditing to ensure their fitness for a given domain. 83 Nor are these systems typically built with any
explicit form of oversight or accountability. This makes discovery of problematic automated
outcomes difficult, especially since such errors and evidence of discrimination frequently
manifest as collective harms, only recognizable as a pattern across many individual cases.
Detecting such problems requires oversight and monitoring. It also requires access to data that is
often neither available to advocates and the public nor monitored by government agencies.
For example, the Houston Federation of Teachers sued the Houston Independent School District
for procuring a third-party ADS to use student test data to make teacher employment decisions,
including which teachers were promoted and which were terminated. It was revealed that no one
in the district – not a single employee – could explain or even replicate the determinations made
by the system, even though the district had access to all the underlying data. 84 Teachers who
sought to contest the determinations were told that the “black box” system was simply to be
believed and could not be questioned. Even when the teachers brought a lawsuit, claiming
constitutional, civil rights, and labor law violations, the ADS vendor fought against providing any
access to how its system worked. As a result, the judge ruled that the use of this ADS in public
employee cases could run afoul of constitutional due process protections, especially when trade
secrecy blocked employees’ ability to understand how decisions were made. The case has
subsequently been settled, with the District agreeing to abandon the third-party ADS.
Similarly, in 2013, Los Angeles County adopted an ADS to assess imminent danger or harm to
children, and to predict the likelihood of a family being re-referred to the child welfare system
within 12 to 18 months. The County did not perform a review of the system or assess the efficacy
of using predictive analytics for child safety and welfare. It was only after the death of a child
whom the system failed to identify as at-risk that County leadership directed a review, which
raised serious questions regarding the system’s validity. The review specifically noted that the
system failed to provide a comprehensive picture of a given family, “but instead focus[ed] on a few
broad strokes without giving weight to important nuance.” 85 Virginia Eubanks found similar
problems in her investigation of an ADS developed by the same private vendor for use in
Allegheny County, PA. This system produced biased outcomes because it significantly
oversampled poor children from working class communities, especially communities of color, in
effect subjecting poor parents and children to more frequent investigation. 86
Even in the face of acknowledged issues of bias and the potential for error in high-stakes
domains, these systems are being rapidly adopted. The Ministry of Social Development in New
Zealand supported the use of a predictive ADS system to identify children at risk of maltreatment,
despite their recognizing that the system raised “significant ethical concerns.” They defended this
on the grounds that the benefits “plausibly outweighed” the potential harms, which included
reconfiguring child welfare as a statistical issue. 87
These cases not only highlight the need for greater transparency, oversight, and accountability in
the adoption, development, and implementation of ADS, but also the need for examination of the
limitations of these systems overall, and of the economic and policy factors that accompany the
push to apply such systems. Virginia Eubanks, who investigated Allegheny County’s use of an
ADS in child welfare, looked at this and a number of case studies to show how ADS are often
adopted to avoid or obfuscate broader structural and systemic problems in society – problems
that are often beyond the capacity of cash-strapped agencies to address meaningfully. 88
Other automated systems have also been proposed as a strategy to combat pre-existing
problems within government systems. For years, criminal justice advocates and researchers have
pushed for the elimination of cash bail, which has been shown to disproportionately harm
individuals based on race and socioeconomic status while at the same time failing to enhance
public safety. 89 In response, New Jersey and California recently passed legislation aimed at
addressing this concern. However, instead of simply ending cash bail, they replaced it with a
pretrial assessment system designed to algorithmically generate “risk” scores that claim to
predict whether a person should go free or be detained in jail while awaiting trial. 90
The shift from policies such as cash bail to automated systems and risk assessment scoring is
still relatively new, and is proceeding even without substantial research examining the potential to
amplify discrimination within the criminal justice system. Yet there are some early indicators that
raise concern. New Jersey’s law went into effect in 2017, and while the state has experienced a
decline in its pretrial population, advocates have expressed worry that racial disparities in the risk
assessment system persist. 91 Similarly, when California’s legislation passed earlier this year, many
of the criminal justice advocates who pushed for the end of cash bail, and supported an earlier
version of the bill, opposed its final version due to the risk assessment requirement. 92
Education policy is also feeling the impact of automated decision systems. A University College
London professor is among those who argued for AI to replace standardized testing, suggesting
that UCL Knowledge Lab’s AIAssess can be “trusted...with the assessment of our children’s
knowledge and understanding,” and can serve to replace or augment more traditional testing. 93
However, much like other forms of AI, there is a growing body of research that shows automated
essay scoring systems may encode bias against certain linguistic and ethnic groups in ways that
replicate patterns of marginalization. 94 Unfair decisions based on automated scores assigned to
students from historically and systemically disadvantaged groups are likely to have profound
consequences on children’s lives, and to exacerbate existing disparities in access to employment
opportunities and resources. 95
The implications of educational ADS go beyond testing to other areas, such as school
assignments and even transportation. The City of Boston was in the spotlight this year after two
failed efforts to address school equity via automated systems. First, the school district adopted a
geographically-driven school assignment algorithm, intended to provide students access to higher
quality schools closer to home. The city’s goal was to increase the racial and geographic
integration in the school district, but a report assessing the impact of the system determined that
it did the opposite: while it shortened student commutes, it ultimately reduced school
integration. 96 Researchers noted that this was, in part, because it was impossible for the system
to meet its intended goal given the history and context within which it was being used. The
geographic distribution of quality schools in Boston was already inequitable, and the pre-existing
racial disparities that played a role in placement at these schools created complications that
could not be overcome by an algorithm. 97
Following this, the Boston school district tried again to use an algorithmic system to improve
inequity, this time designing it to reconfigure school start times – aiming to begin high school
later, and middle school earlier. This was done in an effort to improve student health and
performance based on a recognition of students’ circadian rhythms at different ages, and to
optimize use of school buses to produce cost savings. It also aimed to increase racial equity,
since students of color primarily attended schools with inconvenient start times compounded by
long bus rides. The city developed an ADS that optimized for these goals. However, it was never
implemented because of significant public backlash, which ultimately resulted in the resignation
of the superintendent. 98
In this case, the design process failed to adequately recognize the needs of families, or include
them in defining and reviewing system goals. Under the proposed system, parents with children in
both high school and middle school would need to reconfigure their schedules for vastly different
start and end times, putting strain on those without this flexibility. The National Association for
the Advancement of Colored People (NAACP) and the Lawyers’ Committee for Civil Rights and
Economic Justice opposed the plan because of the school district’s failure to appreciate that
parents of color and lower-income parents often rely on jobs that lack work schedule flexibility
and may not be able to afford additional child care. 99
These failed efforts demonstrate two important issues that policymakers must consider when
evaluating the use of these systems. First, unaddressed structural and systemic problems will
persist and will likely undermine the potential benefits of these systems if they are not addressed
prior to a system’s design and implementation. Second, robust and meaningful community
engagement is essential before a system is put in place and should be included in the process of
establishing a system’s goals and purpose.
In AI Now’s Algorithmic Impact Assessment (AIA) framework, community engagement is an
integral part of any ADS accountability process, both as part of the design stage as well as before,
during, and after implementation. 100 When affected communities have the opportunity to assess
and potentially reject the use of systems that are not acceptable, and to call out fundamental
flaws in the system before it is put in place, the validity and legitimacy of the system is vastly
improved. Such engagement serves communities and government agencies: if parents of color
and lower-income parents in Boston were meaningfully engaged in assessing the goals of the
school start time algorithmic intervention, their concerns might have been accounted for in the
design of the system, saving the city time and resources, and providing a much-needed model of
Above all, accountability in the government use of algorithmic systems is impossible when the
systems making recommendations are “black boxes.” When third-party vendors insist on trade
secrecy to keep their systems opaque, it makes any path to redress or appeal extremely
difficult. 101 This is why vendors should waive trade secrecy and other legal claims that would
inhibit the ability to understand, audit, or test their systems for bias, error, or other issues. It is
important for both people in government and those who study the effects of these systems to
understand why automated recommendations are made, and to be able to trust their validity. It is
even more critical that those whose lives are negatively impacted by these systems be able to
contest and appeal adverse decisions. 102
Governments should be cautious: while automated decision systems may promise short-term
cost savings and efficiencies, it is governments, not third party vendors, who will ultimately be
held responsible for their failings. Without adequate transparency, accountability, and oversight,
these systems risk introducing and reinforcing unfair and arbitrary practices in critical
government determinations and policies. 103
1.3 Experimenting on Society: Who Bears the Burden?
Over the last ten years, the funding and focus on technical AI research and development has
accelerated. But efforts at ensuring that these systems are safe and non-discriminatory have not
received the same resources or attention. Currently, there are few established methods for
measuring, validating, and monitoring the effects of AI systems “in the wild”. AI systems tasked
with significant decision making are effectively tested on live populations, often with little
oversight or a clear regulatory framework.
For example, in March 2018, a self-driving Uber was navigating the Phoenix suburbs and failed to
“see” a woman, hitting and killing her. 104 Last March, Tesla confirmed that a second driver had
been killed in an accident in which the car’s autopilot technology was engaged. 105 Neither
company suffered serious consequences, and in the case of Uber, the person minding the
autonomous vehicle was ultimately blamed, even though Uber had explicitly disabled the vehicle’s
system for automatically applying brakes in dangerous situations. 106 Despite these fatal errors,
Alphabet Inc.’s Waymo recently announced plans for an “early rider program” in Phoenix. 107
Residents can sign up to be Waymo test subjects, and be driven automatically in the process.
Many claim that the occasional autonomous vehicle fatality needs to be put in the context of the
existing ecosystem, in which many driving-related deaths happen without AI. 108 However, because
regulations and liability regimes govern humans and machines differently, risks generated from
machine-human interactions do not cleanly fall into a discrete regulatory or accountability
category. Strong incentives for regulatory and jurisdictional arbitrage exist in this and many other
AI domains. For example, the fact that Phoenix serves as the site of Waymo and Uber testing is
not an accident. Early this year, Arizona, perhaps swayed by a promise of technology jobs and
capital, made official what the state allowed in practice since 2015: fully autonomous vehicles
without anyone behind the wheel are permitted on public roads. This policy was put in place
without any of the regulatory scaffolding that would be required to contend with the complex
issues that are raised in terms of liability and accountability. In the words of the Phoenix New
Times : “Arizona has agreed to step aside and see how this technology develops. If something
goes wrong, well, there's no plan for that yet.” 109 This regulatory accountability gap is clearly visible
in the Uber death case, apparently caused by a combination of corporate expedience (disabling
the automatic braking system) and backup driver distraction. 110
While autonomous vehicles arguably present AI’s most straightforward non-military dangers to
human safety, other AI domains also raise serious concerns. For example, IBM’s Watson for
Oncology is already being tested in hospitals across the globe, assisting in patient diagnostics
and clinical care. Increasingly, its effectiveness, and the promises of IBM’s marketing, are being
questioned. Investigative reporters gained access to internal documents that paint a troubling
picture of IBM’s system, including its recommending “unsafe and incorrect cancer treatments.”
While this system was still in its trial phase, it raised serious concerns about the incentives driving
the rush to integrate such technology, and the lack of clinical validation and peer-reviewed
research attesting to IBM’s marketing claims of effectiveness. 111
Such events have not slowed AI deployment in healthcare. Recently, the U.S. Food and Drug
Administration (FDA) issued a controversial decision to clear the new Apple Watch, which
features a built-in electrocardiogram (EKG) and the ability to notify a user of irregular heart
rhythm, as safe for consumers. 112 Here, concerns that the FDA may be moving too quickly in an
attempt to keep up with the pace of innovation have joined with concerns around data privacy and
security. 113 Similarly, DeepMind Health’s decision to move its Streams Application, a tool designed
to support decision-making by nurses and health practitioners, under the umbrella of Google,
caused some to worry that DeepMind’s promise to not share the data of patients would be
broken. 114
Children and young adults are frequently subjects of such experiments. Earlier this year, it was
revealed that Pearson, a major AI-education vendor, inserted “social-psychological interventions”
into one of its commercial learning software programs to test how 9,000 students would respond.
They did this without the consent or knowledge of students, parents, or teachers. 115 The company
then tracked whether students who received “growth-mindset” messages through the learning
software attempted and completed more problems than students who did not. This psychological
testing on unknowing populations, especially young people in the education system, raises
significant ethical and privacy concerns. It also highlights the growing influence of private
companies in purportedly public domains, and the lack of transparency and due process that
accompany the current practices of AI deployment and integration.
Here we see not only examples of the real harms that can come from biased and inaccurate AI
systems, but evidence of the AI industry’s willingness to conduct early releases of experimental
tools on human populations. As Amazon recently responded when criticized for monetizing
people’s wedding and baby registries with deceptive advertising tactics, “we’re constantly
experimenting.” 116 This is a repeated pattern when market dominance and profits are valued over
safety, transparency, and assurance. Without meaningful accountability frameworks, as well as
strong regulatory structures, this kind of unchecked experimentation will only expand in size and
scale, and the potential hazards will grow.
2.1 Bias Busting and Formulas for Fairness: the Limits of
Technological “Fixes”
Over the past year, we have seen growing consensus that AI systems perpetuate and amplify
bias, and that computational methods are not inherently neutral and objective. This recognition
comes in the wake of a string of examples, including evidence of bias in algorithmic pretrial risk
assessments and hiring algorithms, and has been aided by the work of the Fairness,
Accountability, and Transparency in Machine Learning community. 117 The community has been at
the center of an emerging body of academic research on AI-related bias and fairness, producing
insights into the nature of these issues, along with methods aimed at remediating bias. These
approaches are now being operationalized in industrial settings.
In the search for “algorithmic fairness”, many definitions of fairness, along with strategies to
achieve it, have been proposed over the past few years, primarily by the technical community. 118
This work has informed the development of new algorithms and statistical techniques that aim to
diagnose and mitigate bias. The success of such techniques is generally measured against one or
another computational definition of fairness, based on a mathematical set of results. However,
the problems these techniques ultimately aim to remedy have deep social and historical roots,
some of which are more cleanly captured by discrete mathematical representations than others.
Below is a brief survey of some of the more prominent approaches to understanding and defining
issues involving algorithmic bias and fairness.
● Allocative harms describe the effects of AI systems that unfairly withhold services,
resources, or opportunities from some. Such harms have captured much of the attention
of those dedicated to building technical interventions that ensure fair AI systems, in part
because it is (theoretically) possible to quantify such harms and their remediation. 119
However, we have seen less attention paid to fixing systems that amplify and reproduce
representational harms : the harm caused by systems that reproduce and amplify harmful
stereotypes, often doing so in ways that mirror assumptions used to justify discrimination
and inequality.
In a keynote of the 2017 Conference on Neural Information Processing (NeurIPS), AI Now
cofounder Kate Crawford described the way in which historical patterns of discrimination
and classification, which often construct harmful representations of people based on
perceived differences, are reflected in the assumptions and data that inform AI systems,
often resulting in allocative harms. 120 This perspective requires one to move beyond
locating biases in an algorithm or dataset, and to consider “the role of AI in harmful
representations of human identity,” and the way in which such harmful representations are
both shaped, and shape, our social and cultural understandings of ourselves and each
other. 121
● Observational fairness strategies attempt to diagnose and mitigate bias by considering a
dataset (either data used for training an AI model, or the input data processed by such a
model), and applying methods to the data aimed at detecting whether it encodes bias
against individuals or groups based on characteristics such as race, gender, or
socioeconomic standing. These characteristics are typically referred to as protected or
sensitive attributes. The majority of observational fairness approaches can be categorized
as being a form of either anti-classification, classification parity, or calibration, as
proposed by Sam Corbett-Davies and Sharad Goel. 122 Observational fairness strategies
have increasingly emerged through efforts from the community to contend with the
limitations of technical fairness work and to provide entry points for other disciplines. 123
● Anti-classification strategies declare a machine learning model to be fair if it does not
depend on protected attributes in the data set. For instance, this strategy considers a
pretrial risk assessment of two defendants who differ based on race or gender but are
identical in terms of their other personal information to be “fair” if they are assigned the
same risk. This strategy often requires omitting all protected attributes and their “proxies”
from the data set that is used to train a model (proxies being any attributes that are
correlated to protected attributes, such as ZIP code being correlated with race). 124
● Classification parity declares a model fair when its predictive performance is equal across
groupings that are defined by protected attributes. For example, classification parity would
ensure that the percentage of people an algorithm turns down for a loan when they are
actually creditworthy (its “false negative” rate) is the same for both Black and white
populations. In practice, this strategy often results in decreasing the “accuracy” for certain
populations in order to match that of others.
● Calibration strategies look less at the data and more at the outcome once an AI system
has produced a decision or prediction. These approaches work to ensure that outcomes
do not depend on protected attributes. For example, in the case of pretrial risk
assessment, applying a calibration strategy would aim to make sure that among a pool of
defendants with a similar risk score, the proportion who actually do reoffend on release is
the same across different protected attributes, such as race.
Several scholars have identified limitations with these approaches to observational fairness. With
respect to anti-classification, some argue that there are important cases where protected
attributes—such as race or gender— should be included in data used to train and inform an AI
system in order to ensure equitable decisions. 125 For example, Corbett-Davies and Goel discuss
the importance of including gender in pretrial risk assessment. As women reoffend less often
than men in many jurisdictions, gender-neutral risk assessments tend to overstate the recidivism
risk of women, “which can lead to unnecessarily harsh judicial decisions.” As a result, some
jurisdictions use gender-specific risk assessment tools. These cases counter a widespread view
that deleting sufficient information from data sets will eventually “debias” an AI system. Since
correlations between variables in a dataset almost always exist, removing such variables can
result in very little information, and thus poor predictive performance without the ability to
measure potential harms post hoc.
Secondly, some have argued that different mathematical fairness criteria are mutually exclusive.
Hence, it is generally not possible, except in highly constrained cases, to simultaneously satisfy
both calibration and any form of classification parity. 126 These “impossibility results” show how
each fairness strategy makes implicit assumptions about what is and is not fair. They also
highlight the inherent mathematical trade-offs facing those aiming to mitigate various forms of
bias based on one or another fairness definition. Ultimately, these findings serve to complicate the
broader policy debate focused on solving bias issues with mathematical fairness tools. What they
make clear is that solving complex policy issues related to bias and discrimination by
indiscriminately applying one or more fairness metrics is unlikely to be successful. This does not
mean that such metrics are not useful: observational criteria may help understanding around
whether datasets and AI systems meet various notions of fairness and bias and subsequently
help inform a richer discussion about the goals one hopes to achieve when deploying AI systems
in complex social contexts.
The proliferation of observational fairness methods also raises concerns over the potential to
provide a false sense of assurance. While researchers often have a nuanced sense of the
limitations of their tools, others who might implement them may ignore such limits when looking
for quick fixes. The idea that, once “treated” with such methods, AI systems are free of bias and
safe to use in sensitive domains can provide a dangerous sense of false security—one that relies
heavily on mathematical definitions of fairness without looking at the deeper social and historical
context. As legal scholar Frank Pasquale observes, “algorithms alone can’t meaningfully hold
other algorithms accountable.” 127
While increased attention to the problems of fairness and bias in AI is a positive development,
some have expressed concern over a “mathematization of ethics.” 128 As Shira Mitchell has argued:
“As statistical thinkers in the political sphere we should be aware of the hazards of
supplanting politics by an expert discourse. In general, every statistical intervention to
a conversation tends to raise the technical bar of entry, until it is reduced to a
conversation between technical experts…are we speaking statistics to power? Or are
we merely providing that power with new tools for the marginalization of unquantified
political concerns?” 129
Such concerns are not new. Upcoming work by Hutchinson and Mitchell surveys over fifty years
of attempts to construct quantitative fairness definitions across multiple disciplines. Their work
recalls a period between 1964 and 1973 when researchers focused on defining fairness for
educational assessments in ways that echo the current AI fairness debate. Their efforts stalled
after they were unable to agree on “broad technical solutions to the issues involved in fairness.”
These precedents emphasize what the Fairness, Accountability and Transparency in Machine
Learning community has been discovering: without a “tight connection to real world impact,” the
added value of new fairness metrics and algorithms in the machine learning community could be
minimal. 130 In order to arrive at more meaningful research on fairness and algorithmic bias, we
must continue to pair the expertise and perspectives of communities outside of technical
disciplines to those within.
Broader approaches
Dobbe et al. have drawn on the definition of bias proposed in the early value-sensitive design
(VSD) literature to propose a broader view of fairness. 131 VSD, as theorized in the nineties by Batya
Friedman and Helen Nissenbaum, asserts that bias in computer systems pre-exists the system
itself. 132 Such bias is reflected in the data that informs the systems and embedded in the
assumptions made during the construction of a computer system. This bias manifests during the
operation of the systems due to feedback loops and dissonance between the system and our
dynamic social and cultural contexts. 133 The VSD approach is one way to bring a broader lens to
these issues, emphasizing the interests and perspectives of direct and indirect stakeholders
throughout the design process.
Another approach is a “social systems analysis” first described by Kate Crawford and Ryan Calo in
Nature . 134 This is a method that combines quantitative and qualitative research methods by
forensically analyzing a technical system while also studying the technology once it is deployed in
social settings. It proposes that we engage with social impacts at every stage—conception,
design, deployment, and regulation of a technology, across the life cycle.
We have also seen increased focus on examining the provenance and construction of the data
used to train and inform AI systems. This data shapes AI systems’ “view of the world,” and an
understanding of how it is created and what it is meant to represent is essential to understanding
the limits of the systems that it informs. 135 As an initial remedy to this problem, a group of
researchers led by Timnit Gebru proposed “Datasheets for Datasets,” a standardized form of
documentation meant to accompany datasets used to train and inform AI systems. 136 A follow-up
paper looks at standardizing provenance for AI models. 137 These approaches allow AI
practitioners and those overseeing and assessing the applicability of AI within a given context to
better understand whether the data that shapes a given model is appropriate, representative, or
potentially possessing legal or ethical issues.
Advances in bias-busting and fairness formulas are strong signs that the field of AI has accepted
that these concerns are real. However, the limits of narrow mathematical models will continue to
undermine these approaches until broader perspectives are included. Approaches to fairness and
bias must take into account both allocative and representational harms, and those that debate
the definitions of fairness and bias must recognize and give voice to the individuals and
communities most affected. 138 Any formulation of fairness that excludes impacted populations
and the institutional context in which a system is deployed is too limited.
2.2 Industry Applications: Toolkits and System Tweaks
This year, we have also seen several technology companies operationalize fairness definitions,
metrics, and tools. In the last year, four of the biggest AI companies released bias mitigation tools.
IBM released the “AI Fairness 360” open-source tool kit, which includes nine different algorithms
and many other fairness metrics developed by researchers in the Fairness, Accountability and
Transparency in Machine Learning community. The toolkit is intended to be integrated into the
software development pipeline from early stages of data pre-processing, to the training process
itself, through the use of specific mathematical models that deploy bias mitigation strategies. 139
Google’s People + AI Research group (PAIR) released the open-source “What-If” tool, a dashboard
allowing researchers to visualize the effects of different bias mitigation strategies and metrics, as
well as a tool called “Facets” that supports decision-making around which fairness metric to
use. 140 Microsoft released, a Python package meant to help implement a binary
classifier subject to a developer’s intended fairness constraint. 141 Facebook announced the
creation and testing of a tool called “Fairness Flow”, an internal tool for Facebook engineers that
incorporates many of the same algorithms to help identify bias in machine learning models. 142
Even Accenture, a consulting firm, has developed internal software tools to help clients
understand and “essentially eliminate the bias in algorithms.” 143
Industry standards bodies have also taken on fairness efforts in response to industry and public
sector requests for accountability assurances. The Institute of Electrical and Electronics
Engineers (IEEE) recently announced an Ethics Certification Program for Autonomous and
Intelligent Systems in the hopes of creating “marks” that can attest to the broader public that an
AI system is transparent, accountable, and fair. 144 While this effort is new, and while IEEE has not
published the certification’s underlying methods, it is hard to see, given the complexity of these
issues, how settling on one certification standard across all contexts and all AI systems would be
possible—or ultimately reliable—in ensuring that systems are used in safe and ethical ways.
Similar concerns have arisen in other contexts, such as privacy certification programs. 145
In both the rapid industrial adoption of academic fairness methods, and the rush to certification,
we see an eagerness to “solve” and “eliminate” problems of bias and fairness using familiar
approaches and skills that avoid the need for significant structural change, and which fail to
interrogate the complex social and historical factors at play. Combining “academically credible”
technical fairness fixes and certification check boxes runs the risk of instrumenting fairness in
ways that lets industry say it has fixed these problems and may divert attention from examining
ongoing harms. It also relieves companies of the responsibility to explore more complex and
costly forms of review and remediation. Rather than relying on quick fixes, tools, and
certifications, issues of bias and fairness require deeper consideration and more robust
accountability frameworks, including strong disclaimers about how “automated fairness” cannot
be relied on to truly eliminate bias from AI systems.
2.3 Why Ethics is Not Enough
A top-level recommendation in the AI Now 2017 Report advised that “ethical codes meant to steer
the AI field should be accompanied by strong oversight and accountability mechanisms.” 146 While
we have seen a rush to adopt such codes, in many instances offered as a means to address the
growing controversy surrounding the design and implementation of AI systems, we have not seen
strong oversight and accountability to backstop these ethical commitments.
After it was revealed that Google was working with the Pentagon on Project Maven—developing AI
systems for drone surveillance—the debate about the role of AI in weapons systems grew in
intensity. Project Maven generated significant protest among Google’s employees, who
successfully petitioned the company’s leadership to end their involvement with the program when
the current contract expired. 147 By way of response, Google’s CEO Sundar Pichai released a public
set of seven “guiding principles” designed to ensure that the company’s work on AI will be socially
responsible. 148 These ethical principles include the commitment to ”be socially beneficial,” and to
“avoid creating or reinforcing unfair bias.” They also include a section titled, “AI applications we will
not pursue,” which includes “weapons and other technologies whose principal purpose or
implementation is to cause or directly facilitate injury to people”—a direct response to the
company’s decision not to renew its contract with the Department of Defense. But it is not clear to
the public who would oversee the implementation of the principles, and no ethics board has been
Google was not alone. Other companies, including Microsoft, Facebook, and police body camera
maker Axon, also assembled ethics boards, advisors, and teams. 149 In addition, technical
membership organizations moved to update several of their ethical codes. The IEEE reworked its
code of ethics to reflect the challenges of AI and autonomous systems, and researchers in the
Association for Computing Machinery (ACM) called for a restructuring of peer review processes,
requiring the authors of technical papers to consider the potential adverse uses of their work,
which is not a common practice. 150 Universities including Harvard, NYU, Stanford, and MIT offered
new courses on the ethics and ethical AI development practices aimed at identifying issues and
considering the ramifications of technological innovation before it is implemented at scale. 151 The
University of Montreal launched a wide-ranging process to formulate a declaration for the
responsible development of AI that includes both expert summits and open public deliberations
for input from citizens. 152
Such developments are encouraging, and it is noteworthy that those at the heart of AI
development have declared they are taking ethics seriously. Ethical initiatives help develop a
shared language with which to discuss and debate social and political concerns. They provide
developers, company employees, and other stakeholders a set of high-level value statements or
objectives against which actions can be later judged. They are also educational, often doing the
work of raising awareness of particular risks of AI both within a given institution, and externally,
amongst the broader concerned public. 153
However, developing socially just and equitable AI systems will require more than ethical
language, however well-intentioned it may be. We see two classes of problems with this current
approach to ethics. The first has to do with enforcement and accountability. Ethical approaches in
industry implicitly ask that the public simply take corporations at their word when they say they
will guide their conduct in ethical ways. While the public may be able to compare a post hoc
decision made by a company to its guiding principles, this does not allow insight into decision
making, or the power to reverse or guide such a decision. In her analysis of Google’s AI Principles,
Lucy Suchman, a pioneering scholar of human computer interaction, argues that without “the
requisite bodies for deliberation, appeal, and redress” vague ethical principles like “don’t be evil” or
“do the right thing” are “vacuous.” 154
This “trust us” form of corporate self-governance also has the potential to displace or forestall
more comprehensive and binding forms of governmental regulation. Ben Wagner of the Vienna
University of Economics and Business argues, “Unable or unwilling to properly provide regulatory
solutions, ethics is seen as the “easy” or “soft” option which can help structure and give meaning
to existing self-regulatory initiatives.” 155 In other words, ethical codes may deflect criticism by
acknowledging that problems exist, without ceding any power to regulate or transform the way
technology is developed and applied. The fact that a former Facebook operations manager
claims, “We can’t trust Facebook to regulate itself,” should be taken into account when evaluating
ethical codes in industry. 156
A second problem relates to the deeper assumptions and worldviews of the designers of ethical
codes in the technology industry. In response to the proliferation of corporate ethics initiatives,
Greene et al. undertook a systematic critical review of high-profile “vision statements for ethical
AI.” 157 One of their findings was that these statements tend to adopt a technologically
deterministic worldview, one where ethical agency and decision making was delegated to experts,
“a narrow circle of who can or should adjudicate ethical concerns around AI/ML” on behalf of the
rest of us. These statements often assert that AI promises both great benefits and risks to a
universal humanity, without acknowledgement of more specific risks to marginalized populations.
Rather than asking fundamental ethical and political questions about whether AI systems should
be built, these documents implicitly frame technological progress as inevitable, calling for better
building. 158
Empirical study of the use of these codes is only beginning, but preliminary results are not
promising. One recent study found that “explicitly instructing [engineers] to consider the ACM
code of ethics in their decision making had no observed effect when compared with a control
group.” 159 However, these researchers did find that media or historical accounts of ethical
controversies in engineering, like Volkswagen’s Dieselgate, may prompt more reflective practice.
Perhaps the most revealing evidence of the limitations of these emerging ethical codes is how
corporations act after they formulate them. Among the list of applications Google promises not to
pursue as a part of its AI Principles are “technologies whose purpose contravenes widely
accepted principles of international law and human rights.” 160 That was tested earlier this year
after investigative journalists revealed that Google was quietly developing a censored version of
its search engine (which relies extensively on AI capabilities) for the Chinese market, code-named
Dragonfly. 161 Organizations condemned the project as a violation of human rights law, and as
such, a violation of Google’s AI principles. Google employees also organized against the effort. 162
As of writing, the project has not been cancelled, nor has its continued development been
explained in light of the clear commitment in the company’s AI Principles, although Google’s CEO
has defended it as “exploratory.” 163
There is an obvious need for accountability and oversight in the industry, and so far the move
toward ethics is not meeting this need. This is likely in part due to the market-driven incentives
working against industry-driven implementations: a drastic (if momentary) drop in Facebook and
Twitter’s share price occurred after they announced efforts to combat misinformation and
increase spending on security and privacy efforts. 164
This is no excuse not to pursue a more ethically driven agenda, but it does suggest that we should
be wary of relying on companies to implement ethical practices voluntarily, since many of the
incentives governing these large, publicly traded technology corporations penalize ethical action.
For these mechanisms to serve as meaningful forms of accountability requires that external
oversight and transparency be put into place to ensure that there exists an external system of
checks and balances in addition to the cultivation of ethical norms and values within the
engineering profession and technology companies.
When we released our AI Now 2016 Report, fairness formulas, debiasing toolkits, and ethical
guidelines for AI were rare. The fact that they are commonplace today shows how far the field has
come. Yet much more needs to be done. Below, we outline seven strategies for future progress on
these issues.
3.1 From Fairness to Justice
Any debate about bias and fairness should approach issues of power and hierarchy, looking at
who is in a position to produce and profit from these systems, whose values are embedded in
these systems, who sets their “objective functions,” and which contexts they are intended to work
within. 165 Echoing the Association for Computing Machinery (ACM) researcher’s call for an
acknowledgement of “negative implications” as a requirement for peer review, much more
attention must be paid to the ways that AI can be used as a tool for exploitation and control. 166 We
must also be cautious not to reframe political questions as technical concerns. 167
When framed as technical “fixes,” debiasing solutions rarely allow for questions about the
appropriateness or efficacy of an AI system altogether, or for an interrogation of the institutional
context into which the “fixed” AI system will ultimately be applied. For example, a “debiased”
predictive algorithm that accurately forecasts where crime will occur, but that is being used by law
enforcement to harass and oppress communities of color, is still an essentially unfair system. 168
To this end, our definitions of “fairness” must expand to encompass the structural, historical, and
political contexts in which an algorithmic systems is deployed.
Furthermore, fairness is a term that can be easily co-opted: important questions such as “Fair to
whom? And in what context?” should always be asked. For example, making a facial recognition
system perform equally on people with light and dark skin may be a type of technical progress in
terms of parity, but if that technology is disproportionately used on people of color and
low-income communities, is it really “fair?” This is why definitions of fairness face a hard limit if
they remain purely contained within the technical domain: in short, “parity is not justice.” 169
3.2 Infrastructural Thinking
In order to better understand and track the complexities of AI systems, we need to look beyond
the technology and the hype to account for the broader context of how AI is shaping and shaped
by social and material forces. As Edwards et al. argue: “When dealing with infrastructures, we
need to look to the whole array of organizational forms, practices, and institutions which
accompany, make possible, and inflect the development of new technology.” 170 Doing so requires
both experimental methodological approaches and theory building, expanding beyond narrow
analyses of individual systems in isolation to consider them on a local and global scale. It also
requires considering ways in which technologies are entangled in social relations, material
dependencies, and political purposes. 171
In “Anatomy of an AI System,” a 2018 essay and large-scale map, AI Now cofounder Kate
Crawford and Professor Vladan Joler took a single Amazon Echo and analyzed all the forms of
environmental and labor resources required to develop, produce, maintain, and finally dispose of
this sleek and seemingly simple object. When you ask Alexa to play your favorite song, you have
drawn on a massive interlinked chain of extractive processes. It involves lithium mining in Bolivia,
clickworkers creating large-scale training datasets in southeast Asia, container ships and
international logistics, and vast data extraction and analysis by Alexa Voice Service (AVS) across
distributed data centers. The process ends in the final resting place of all AI consumer gadgets: in
e-waste rubbish heaps in Ghana, Pakistan, and China.
The “Anatomy of an AI System” project points to approaches we can employ in contending with
the global implications of AI, and the multi-layered nature of value extraction and exploitation from
the developing world to the developed world. This helps to illuminate the darker corners that are
rarely considered in analysis of AI systems. 172
In particular, an infrastructural analysis of AI shows that there are black boxes within black boxes:
not just at the algorithmic level, but also at the levels of trade secrecy laws, labor practices, and
untraceable global supply chains for rare earth minerals used to build consumer AI devices. These
obscure not only the material impacts of AI systems, but the intensive human work of maintaining
and repairing them through practices like content moderation and data training. 173 As Nick Seaver
puts it, “If you cannot see a human in the loop, you just need to look for a bigger loop.” 174
Only by tracing across these sociotechnical layers can we understand what we are calling the “full
stack supply chain” of AI—the human and nonhuman components that make up the global scale
of AI systems. There are many sociotechnical data infrastructures needed to make AI function:
these include training data, test data, APIs, data centers, fiber networks, undersea cables, energy
use, labor involved in content moderation and training set creation, and a constant reliance on
clickwork to develop and maintain AI systems. We cannot see the global environmental and labor
implications of these tools of everyday convenience, nor can we meaningfully advocate for
fairness, accountability, and transparency in AI systems, without an understanding of this full
stack supply chain.
3.3 Accounting for Hidden Labor in AI Systems
Another emerging research area where we expect to see greater impact focuses on the underpaid
and unrecognized workers who help build, maintain, and test AI systems. This hidden human
labor takes many forms—from supply chain work, to digital crowdsourced “clickwork,” to
traditional service industry jobs. Hidden labor exists at all stages of the AI pipeline, from
producing and transporting the raw minerals required to create the core infrastructure of AI
systems, to providing the invisible human work that often backstops claims of AI “magic” once
these systems are deployed in products and services. 175 Communications scholar Lilly Irani refers
to such hidden labor as “human-fueled automation.” 176 Her research draws attention to the
experiences of clickworkers or “microworkers” who perform the repetitive digital tasks that
underlie AI systems, like labeling training data and reviewing flagged content, as “workers hidden
in the technology.” 177
While this labor is essential to making AI systems “work,” it is usually very poorly compensated. A
2018 study from the United Nations’ International Labor Organization (ILO) surveyed 3,500
microworkers from 75 countries who routinely offered their labor on popular microtask platforms
like Mechanical Turk, Crowdflower, Microworker, and Clickworker. The report found that a
substantial number of people earned below their local minimum wage (despite 57% of
respondents having advanced degrees specializing in science and technology). 178 Similarly, those
who do content moderation work, screening problematic content posted on social media
platforms and news feeds, are also paid poorly, in spite of their essential and emotionally difficult
labor. 179
This has not been lost on some in the technical AI research community, who have begun to call
attention to the crucial and marginalized role of this labor, and to consider their own responsibility
to intervene. Silberman and others discuss how researchers conducting AI studies are
increasingly dependent upon cheap crowdsourced labor. 180 They note that, between the years
2008 and 2016, the term “crowdsourcing” went from appearing in less than 1,000 scientific
articles to over 20,000. With online microworkers unregulated by current labor laws, researchers
are being asked to reconsider what counts as “ethical conduct” in the AI research community.
Silberman et al. argue for treating crowdworkers as coworkers, paying them minimum wage
determined by the client’s location, and the need for additional Institutional Review Board (IRB)
The practice of examining hidden human labor draws on a lineage of feminist research. The
concept of “invisible work,” for instance, originated with studies of unpaid women’s care work and
investigations into organizational settings that relied upon “emotional labor,” particularly
traditionally “feminized” fields like nursing and flight attendants. 181 Researchers found that
common activities taken on by female workers, such as soothing anxious patients or managing
unruly customers, were not formally recognized or compensated as work, in spite of their being
essential. The feminist legacy of invisible work is useful for contextualizing these new forms of
labor, and in understanding the characterization of this work, which, while essential, is often
written out of the AI narrative, rarely counted or compensated.
In her article, “The Automation Charade,” Astra Taylor proposes the term “fauxtomation” to call
attention to the gap between the marketing rhetoric of AI as a seamless product or service and
the messy, lived reality of automation, which frequently relies on such unsung human labor.
“Automation,” Taylor cautions, “has an ideological function as well as a technological
dimension.” 182 In making this case, she critiques popular narratives around the future of labor,
which posit a near-horizon where workers will be replaced by machines. She sees such claims as
functioning to disempower workers: what leverage do workers have to demand better wages and
benefits in the face of impending automation? We saw this narrative deployed in 2016 by former
McDonald’s CEO Ed Rensi, who cited the growing “Fight for $15” movement as the impetus for the
company’s introduction of automated kiosks to replace cashiers. 183 Workers who fought for better
pay would ultimately be worse off, he reasoned, as their demand for living wages would force the
company to automate and eliminate them. Examining his claim two years on, we see that this is
not entirely true. Automation or no, workers are still needed: after McDonald’s added kiosks to its
Chicago flagship store, the location reopened with more employees than before the kiosks were
introduced. 184
The integration of automation and AI in the workplace is aimed not only at automating worker
tasks, but at managing, monitoring, and assessing workers themselves. Alex Rosenblatt’s 2018
ethnography of Uber drivers details the precarity and uncertainty produced by depending on the
whims of a centralized, AI-enabled platform for one’s livelihood. The algorithmic logic that governs
ride-sharing applications can arbitrarily bar drivers from work, result in unreliable wages and
unexpected costs, and nudge people into working longer hours, resulting in unsafe driving
conditions. 185 Such platforms isolate workers from each other, making concerted activity and
labor organizing difficult. They also function to create significant information asymmetries
between data-rich companies aiming to extract value from workers, and the workers themselves.
Even so, 2018 has seen increasing dissent from such workers. Some prominent examples of
worker-driven protest include on-demand delivery riders striking alongside UK fast food industry
employees and rideshare drivers calling for job protections. 186
Silicon Valley contractors working in security, food, and janitorial services within major technology
companies have also organized, seeking a living wage and other protections. 187 They are among
thousands of workers who labor alongside their full-time technology worker peers, but are
classified as independent contractors. Under this designation, they are often paid low wages, and
provided few benefits and protections. They are also rarely counted in official employee numbers,
even though they make up a large portion of most technology industry workforces, and perform
essential work. For example, as of this year, contract workers outnumber Google’s direct
employees for the first time in the company’s history. 188 This increasing wave of dissent makes
visible the social tensions at the heart of the practice of hiding and marginalizing important forms
of labor.
The physical, emotional, and financial costs of treating workers like “bits of code” and devaluing
their work and well-being has been highlighted in recent news articles describing the conditions of
Amazon warehouse workers and contracted Prime delivery drivers. 189 Amazon warehouse
workers recently went on strike in Europe, protesting harsh conditions. According to one striking
worker, “You start at the company healthy and leave it as a broken human,” with many workers
requiring surgeries related to workplace conditions. 190
Recognizing all of the labor required to “make AI work” can help us better understand the
implications of its development and use. Research in these areas also helps us reexamine the
focus on technical talent in narratives describing AI’s creation and recognize that technical skills
account for only a portion of a much larger effort. This enables us to question numerous labor
policies, such as the focus on pushing workers to acquire coding or data science skills as a way
to ensure they are counted and compensated. They also help us identify who is likely to benefit,
and who, along the AI production and deployment pipeline, is likely to be harmed.
3.4 Deeper Interdisciplinarity
AI researchers and developers are engaged in building technologies that have significant
implications for diverse populations in broad fields like law, sociology, and medicine. Yet much of
this development happens far removed from the experience and expertise of these groups. This
has led to a call to expand the disciplinary makeup of those engaged in AI design, development,
and critique, beyond purely technical expertise. 191 Since then, we have seen some movement in
this direction. Recently, MIT announced plans to establish a new college of computing that aims
to “advance pioneering work on AI’s ethical use and societal impact” by fostering integrated
cross-disciplinary training, “educating the bilinguals of the future,” as MIT President L. Rafael Reif
described it. 192
Such initiatives are critical: as AI becomes more deeply embedded in areas like healthcare,
criminal justice, hiring, housing, and educational systems, experts from these domains are
essential if we are to ensure AI works as envisioned. In integrating these disciplinary perspectives,
it is important that they are not merely ”languages” to be acquired by computer scientists and
engineers seeking to expand their work into new areas—especially when other disciplines have
been leading that work. Instead, social science and the humanities should be centered as
contributors to the AI field’s foundational knowledge and future direction, enabling us to leverage
new modes of analysis and methodological intervention. 193
3.5 Race, Gender and Power in AI
This year, a groundswell of political action emerged around issues of discrimination, harassment,
and inequity in the technology industry, especially in the AI field. 194 This rising concern weaves
together a number of related issues, from the biases in AI systems, to failed diversity and
inclusion efforts within industry and academia, to the grassroots efforts to confront sexual
harassment and the abuse of power in workplaces and classrooms.
Resonating with the broader #MeToo movement, we saw issues relating to diversity and inclusion
in artificial intelligence rise on the public agenda:
● Following the 2017 Conference on Neural Information Processing Systems, members of
the artificial intelligence and machine learning communities began voicing concerns about
long standing problems of harassment and discrimination in conference settings, leading
to #ProtestNIPS, a movement aimed at highlighting examples of toxicity in the community
and the need to address them. 195 Among other things, this provoked a change to the
conference acronym, a longstanding subject of sensitivity for its gendered and historical
connotations. The conference, which was previously referred to as NIPS, now goes by
NeurIPS. 196
● We also saw renewed focus on initiatives devoted to creating platforms for inclusion in
the field, such as Black in AI, Women in Machine Learning, Latinx in AI, and Queer in AI,
alongside the appointment of Diversity and Inclusion chairs and a series of other changes
to the design of NeurIPS intended to foster equity and inclusion among participants. 197
● Across the industry, we saw a growing technology worker movement that intersected with
these issues. The Google Walkout, in particular, took on a worker-driven agenda that
acknowledged that race, class, and sexuality are intertwined with forms of gender-based
discrimination. The walkout explicitly aimed to center the needs of the company’s
temporary contract workers and vendors who lack the job security and benefits of more
privileged technology workers. 198 These efforts have led to some significant structural
changes—notably, the end to forced arbitration for sexual harassment claims across a
number of the largest companies in the AI industry. 199
● In other arenas, corporate boards have ignored or otherwise refused to address
shareholder proposals targeting discriminatory workspaces. This year, Google dismissed
a proposal that would tie executive compensation to progress made on diversity and
inclusion, while in 2016, Apple refused a mandate that would require it to diversify its
board and senior management. 200
Across these efforts, advocates of diversity in AI are finding intersections between the move to
address gender and race-based harassment and abuse within the technology community, and
other forms of inequity and abuses of power. But this is still an uphill battle: while there is
increased attention to problems of bias in AI systems, we have yet to see much research within
the fairness and bias debate focused on the state of equity and diversity in the AI field itself.
Indeed, reliable figures on representation in AI are difficult to come by, although some limited data
does exist.
A recent estimate produced by WIRED and Element AI found that only 12% of researchers who
contributed to the three leading machine learning conferences in 2017 were women. This gender
gap is replicated at large technology firms like Facebook and Google, whose websites show that
only 15% and 10% of their AI research staff are women. 201 And there is no reliable data on the
state of racial diversity in the field, or retention rates for people of color. 202 Collectively, the limited
evidence suggests that AI, as a field, is even less diverse than computer science as a whole,
which is itself at a historic low point: women make up only 18% of computer science majors in the
United States, a decline from a high point of 37% in 1984. 203
These trends are even more dramatic when compared to other STEM fields in which gender
diversity has shown a marked improvement. 204 Yet these are not new problems: the
WIRED /Element AI survey is not significantly different from a study of the AI field that was
published by IEEE Expert in 1992, which found that only 13% of published authors in the journal
over the prior four years were women. 205 And in the 1980s, female grad students at MIT’s
Computer Science and Artificial Intelligence Labs thoroughly documented their experiences with
toxic working environments in the report “Barriers to Equality in Academia: Women in Computer
Science at MIT.” 206
It is time to address the connection between discrimination and harassment in the AI community,
and bias in the technical products that are produced by the community. Scholars in science and
technology studies have long observed that the values and beliefs of those who create
technologies shape the technologies they create. 207 Expanding the field’s frame of reference to
recognize this connection will ensure it is better equipped to address the problems raised by its
rapid proliferation into sensitive social domains. As one AI researcher put it, “Bias is not just in our
datasets, it’s in our conferences and community.” 208
A recent example illustrates these connections, and how discriminatory practices within the
culture that produces an AI system can be mirrored and amplified in the system itself. Amazon
recently developed an experimental AI system to help it rank job candidates. It trained the system
on data reflecting the company’s historical hiring preferences, hoping to more efficiently identify
qualified candidates. 209 But the system didn’t work as expected: based on the company’s
historical hiring, it showed a distinct bias against women candidates, downgrading resumes from
candidates who attended two all-women’s colleges, and even penalizing resumes that contained
the word “woman.” After uncovering this bias, the company attempted to fix the system, adjusting
the algorithm to treat these terms more fairly. This did not work, and the project was eventually
scrapped. Gender-based discrimination was embedded too deeply within the system – a system
built to reflect Amazon’s past hiring practices – to be uprooted using the “debiasing” approach
commonly adopted within the AI field.
As scholars like Safiya Noble and Mar Hicks have observed, there is a clear through-line
connecting longstanding patterns of discrimination and harassment in AI to the ways artificial
intelligence technologies can amplify and contribute to marginalization and social inequity. 210
Patterns of cultural discrimination are often embedded in AI systems in complex and meaningful
ways, and we need to better understand how these effects are felt by different communities. 211
This is a space that has too long been overlooked and where research is sorely needed. AI Now is
publishing a dedicated report on these issues in December 2018, and we have a multi-year
research project dedicated to examining these challenges.
3.6 Strategic Litigation and Policy Interventions
This year saw an increase in court challenges to the use of automated systems, particularly when
government agencies use them in decisions that affect individual rights. In a recent AI Now
Report called “Litigating Algorithms,” we documented five recent case studies of litigation
involving the use of automated systems: in Medicaid and disability benefits cases, public teacher
employment evaluations, juvenile criminal risk assessment, and criminal DNA analysis. 212
The findings brought to light several emerging trends. First, these cases provided concrete
evidence that governments are routinely adopting automated decision systems (ADS) as
measures to produce “cost savings” or to streamline work. Yet, they are failing to assess how
these systems might disproportionately harm the populations they are meant to serve,
particularly those who are the most vulnerable and who have little recourse or even knowledge
that these systems are deeply affecting their lives. In many cases, there was not a single
government employee who could explain the automated decision, correct errors, or audit the
results of its determination. Through a series of vendor and contractor agreements, almost all
avenues for understanding or contesting the impact of these systems were shielded by legal
protections such as trade secret law.
Second, few government agencies had invested real efforts to ensure that fairness and due
process protections remained in place when switching from human-driven decisions to
algorithmically-driven ones. The typical audit, appeals, and accountability mechanisms were
totally absent from automated system design. Fortunately, successful strategic litigation by
lawyers from the American Civil Liberties Union (ACLU) of Idaho, Legal Aid of Arkansas, the
Houston Federation of Teachers, The Legal Aid Society of New York, and various public defenders
were able to secure victories for their clients and challenge these unlawful uses based, in part, on
constitutional and administrative due process litigation claims.
The playbook for how to litigate algorithms is still being written, but our report uncovered several
useful strategies to support long-term solutions and protections. First, arguments based on
procedural due process presented serious challenges to the trade secrecy claims of private
vendors, with the vast majority of judges ruling that the right to assert constitutional or civil rights
protections outweighs any risk of intellectual property misappropriation. Second, a failure to notify
affected individuals and communities matters: agencies who neglected to engage community
groups concerning the use of these systems were often judged to have failed to appropriately
provide the opportunity for public notice and comment, meaning that their implementation of AI
systems was potentially unconstitutional. Third, interdisciplinary collaboration is important when
trying to determine where these systems fail, especially when submitting evidence to judges. In
cases in which lawyers worked closely with technical and social science experts, judges were able
to learn about the scientific flaws in these systems as well as the social ramifications and harms.
Looking forward, we anticipate future strategic litigation cases will produce many more lessons.
These interventions generate greater understanding and remedial accountability for these
systems, even in situations where government agencies have attempted to disclaim ownership,
understanding, or control. Combined with tools such as AI Now’s Algorithmic Impact Assessment
framework, alongside robust regulatory oversight regimes, we can begin to identify, measure, and,
when necessary, intervene in efforts to use AI and automated systems in ways that produce
harm. 213 However, in order to continue to build on recent progress, lawyers and community
activists who represent individuals in such suits need greater funding and support, as well as
networks of domain experts that they can draw on to help advise strategy and audit systems.
3.7 Research and Organizing: An Emergent Coalition
The rapid deployment of AI and related systems in everyday life is not a concern for the future—it
is already here, with no signs of slowing down. Recognizing this, a set of strategies have emerged,
drawing on long-standing traditions of activism and organizing to demand structural changes for
greater accountability.
Social activism by technologists is nothing new. In the early 1980s, Computer Professionals for
Social Responsibility formed to oppose the use of computers in warfare. 214 More recently, the
2016 “Never Again” technology pledge rallied thousands of workers in various technology sectors
to sign a promise not to build databases or conduct data collection that could be used to target
religious minorities or facilitate mass deportations. 215 While 2018’s organizing and activism draws
from a long tradition, its scale is new to the technology sector. Technology workers are joining
forces with civil society organizations and researchers in opposition to their employers’ technical
and business decisions.
Google employees kicked off publicly visible organizing in 2018, opposing Project Maven, a
Pentagon effort to apply Google’s machine vision AI capabilities to Department of Defense drone
surveillance. 216 Researchers and human rights organizations joined the cause, and in June,
Google announced it would abandon the project. 217 At Amazon, Salesforce, and Microsoft,
employees petitioned their leadership to end contracts with Immigrations and Customs
Enforcement (ICE), supported by immigration and advocacy organizations. 218 Amazon employees
also joined the ACLU in petitioning the company to stop selling facial recognition to law
enforcement, responding to the ACLU’s work exposing existing contracts. 219 Following Maven,
Google employees again rose up against Project Dragonfly, a version of the Google search engine
enabling government-directed censorship and surveillance, planned for the Chinese market. 220 In
response to media reports that disclosed the secretive effort, employees requested ethical
oversight and accountability, and over 700 of them joined Amnesty International in a call to cancel
the project, signing their name publicly to an open letter which coincided with Amnesty
International protests 221
The biggest moment occurred in early November, when 20,000 Google workers walked out
around the globe in an action called Walkout for Real Change. 222 The walkout characterized
Google as a company at which “abuse of power, systemic racism, and unaccountable
decision-making are the norm.” 223 Organizers called on leadership to meet five demands, including
ending pay and opportunity inequity, eliminating forced arbitration in cases of sexual harassment
and discrimination, and adding an employee representative to the board of directors. A week after
the walkout, Google met a small portion of these demands, agreeing to end forced arbitration in
cases of sexual harassment (but notably ignoring discrimination). 224 This move was quickly
replicated throughout the industry, with Facebook, Square, eBay, and Airbnb following suit. 225
By joining forces with researchers and civil society groups, this new wave of labor organizing
mirrors calls for greater diversity and openness within the AI research domain. 226 These
movements are incorporating diverse perspectives across class, sector, and discipline, working to
ensure they are capable of understanding the true costs of company practices, including the
impact of the systems they build. The Google workers who participated in the walkout expanded
their coalition across class and sector, emphasizing contract workers in their demands, and
situating themselves within a growing movement “not just in tech, but across the country,
including teachers, fast-food workers and others who are using their strength in numbers to make
real change.” 227
The recent surge in activism has largely been driven by whistleblowers within technology
companies, who have disclosed information about secretive projects to journalists. 228 These
disclosures have helped educate the public, which is traditionally excluded from such access, and
helped external researchers and advocates provide more informed analysis. By establishing
shared ground truth, whistleblowing has helped build the broad coalitions that characterize these
movements. The critical role of ethical whistleblowing over the last year has also highlighted both
its social importance, and the lack of protections for those who make such disclosures.
The broad coalition of technology worker organizers, researchers, and civil society is playing an
increasing role in the push for accountability in the technology sector. Many engineering
employees have considerable bargaining power and are uniquely positioned to demand change
from their employers. 229 Applying this power to push for greater accountability presents a hopeful
model for labor organizing in the public interest, especially given the current lack of government
regulation, external oversight, and other meaningful levers capable of reviewing and steering
technology company decision making.
This year saw AI systems rapidly introduced into more social domains, leaving increasing
numbers of people at risk. While AI techniques still offer considerable promise, rapid deployment
of systems without appropriate assessment, accountability, and oversight can create serious
hazards. We urgently need to regulate AI systems sector-by-sector, with particular attention paid
to facial and affect recognition, and to inform those policies with rigorous research.
But regulation can only be effective if the legal and technological barriers that prevent auditing,
understanding, and intervening in these systems are removed. Back in 2016, we recommended in
the first AI Now report that the Computer Fraud and Abuse Act (CFAA) and the Digital Millennium
Copyright Act (DMCA) should not be used to restrict research into AI accountability and
auditing. 230 This year, we go further: AI companies should waive trade secrecy and other legal
claims that would prevent algorithmic accountability in the public sector. Governments and public
institutions must be able to understand and explain how and why decisions are made, particularly
when people’s access to healthcare, housing, and employment is on the line.
The question is no longer whether there are harms and biases in AI systems. That debate has
been settled: the evidence has mounted beyond doubt in the last year. The next task now is
addressing these harms. This is particularly urgent given the scale at which these systems are
deployed, the way they function to centralize power and insight in the hands of the few, and the
increasingly uneven distribution of costs and benefits that accompanies this centralization. We
need deeper analyses of the “full stack supply chain” behind AI systems, to track their
development and deployment across the product life cycle, and to take into account their true
environmental and labor costs. 231
Furthermore, it is long overdue for technology companies to directly address the cultures of
exclusion and discrimination in the workplace. The lack of diversity and ongoing tactics of
harassment, exclusion, and unequal pay are not only deeply harmful to employees in these
companies but also impacts the AI products they release, producing tools that perpetuate bias
and discrimination. 232
The current structure within which AI development and deployment occurs works against
meaningfully addressing these pressing issues. Those in a position to profit are incentivized to
accelerate the development and application of systems without taking the time to build diverse
teams, create safety guardrails, or test for disparate impacts. Those most exposed to harm from
these systems commonly lack the financial means and access to accountability mechanisms
that would allow for redress or legal appeals. 233 This is why we are arguing for greater funding for
public litigation, labor organizing, and community participation as more AI and algorithmic
systems shift the balance of power across many institutions and workplaces.
It is imperative that the balance of power shifts back in the public’s favor. This will require
significant structural change that goes well beyond a focus on technical systems, including a
willingness to alter the standard operational assumptions that govern the modern AI industry
players. The current focus on discrete technical fixes to systems should expand to draw on
socially-engaged disciplines, histories, and strategies capable of providing a deeper
understanding of the various social contexts that shape the development and use of AI systems.
As more universities turn their focus to the study of AI’s social implications, computer science and
engineering can no longer be the unquestioned center, but should collaborate more equally with
social and humanistic disciplines, as well as with civil society organizations and affected
Fortunately, we are beginning to see new coalitions form between researchers, activists, lawyers,
concerned technology workers, and civil society organizations to support the oversight,
accountability, and ongoing monitoring of AI systems. For these important connections to grow,
more protections are needed, including a commitment from technology companies to provide
protections for conscientious objectors who do not want to work on military or policing contracts,
along with protections for employees involved in labor organizing and ethical whistleblowers. 234
The last year revealed many of the hardest challenges for accountability and justice as AI
systems moved deeper into the social world. Yet there have been extraordinary moments of
potential, as well as significant public debates and hopeful forms of protest, that may ultimately
illuminate the pathways for consequential and positive change.
1. As AI pioneers Stuart Russell and Peter Norvig point out, the history of artificial intelligence has not
produced a clear definition of AI, but can be seen as variously emphasizing four possible goals:
“systems that think like humans, systems that act like humans, systems that think rationally, systems
that act rationally.” In this report we use the term AI to refer to a broad assemblage of technologies,
from early rule-based algorithmic systems to deep neural networks, all of which rely on an array of
data and computational infrastructures. These technologies span speech recognition, language
translation, image recognition, predictions, and determinations—tasks that have traditionally relied on
human capacities across the four goals Russell and Norvig identify. While AI is not new, recent
developments in the ability to collect and store large quantities of data, combined with advances in
computational power have led to significant breakthroughs in the field over the last ten years, along
with a strong push to commercialize these technologies and apply them across core social domains.
See: Stuart J. Russell and Peter Norvig, Artificial Intelligence: A Modern Approach , (Englewood Cliffs,
NJ: Prentice Hall, 1995), 2.
2. Carole Cadwalladr and Emma Graham-Harrison, “Revealed: 50 Million Facebook Profiles Harvested
for Cambridge Analytica in Major Data Breach,” The Guardian , March 17, 2018,
on .
3. Guy Rosen, “Security Update,” Facebook Newsroom , September 28, 2018, .
4. Josh Eidelson, “Facebook Tools Are Used to Screen Out Older Job Seekers, Lawsuit Claims,”
Bloomberg, May 29, 2018,
-job-seekers-lawsuit-claims .
5. Bloomberg Editorial Board, “Think the U.S. Has a Facebook Problem? Look to Asia,” Bloomberg ,
October 22, 2017,
hington .
6. Andrew Liptak, “The US Government Alleges Facebook Enabled Housing Ad Discrimination,” The
Verge , August 19, 2018,
-facebook-complaint-race-gender-discrimination .
7. Elizabeth Weise, “Russian Fake Accounts Showed Posts to 126 Million Facebook Users,” USA TODAY ,
October 30, 2017,
n-facebook-users/815342001/ .
8. Hamza Shaban, Craig Timberg, and Elizabeth Dwoskin, “Facebook, Google and Twitter Testified on
Capitol Hill. Here’s What They Said,” Washington Post , October 31, 2017,
e-set-to-testify-on-capitol-hill-heres-what-to-expect/ ; Casey Newton, “Mark Zuckerberg’s Appearance
before European Parliament Yields an Empty Spectacle,” The Verge , May 22, 2018, .
9. Drew Harwell, “AI will solve Facebook’s most vexing problems, Mark Zuckerberg says. Just don’t ask
when or how,” Washington Post , April 11, 2018,
xing-problems-mark-zuckerberg-says-just-dont-ask-when-or-how/ .
10. Kate Conger and Dell Cameron, “Google Is Helping the Pentagon Build AI for Drones,” Gizmodo , March
6, 2018, .
11. Rick Paulas, “A New Kind of Labor Movement in Silicon Valley,” The Atlantic , September 4, 2018, .
12. Hamza Shaban, “Amazon Employees Demand Company Cut Ties with ICE,” Washington Post , June
22, 2018,
ompany-cut-ties-with-ice/ ; Jacob Kastrenakes, “Salesforce Employees Ask CEO to ‘Re-Examine’
Contract with Border Protection Agency,” The Verge , June 25, 2018,
mmigration-cbp ; Colin Lecher, “The Employee Letter Denouncing Microsoft’s ICE Contract Now Has
over 300 Signatures,” The Verge , June 21, 2018, .
13. Nikhil Sonnad, “US Border Agents Hacked Their “Risk Assessment” System to Recommend Detention
100% of the Time,” Quartz , June 26, 2018,
migrant-detention-every-time/ .
14. Daisuke Wakabayashi, “Self-Driving Uber Car Kills Pedestrian in Arizona, Where Robots Roam,” The
New York Times , July 30, 2018, .
15. Nikhil Sonnad, “A Flawed Algorithm Led the UK to Deport Thousands of Students,” Quartz , May 3,
2018, .
16. Casey Ross and Ike Swetlitz, “IBM’s Watson Recommended ‘unsafe and Incorrect’ Cancer
Treatments,” STAT , July 25, 2018, .
17. George Joseph and Kenneth Lipp, “IBM Used NYPD Surveillance Footage to Develop Technology That
Lets Police Search by Skin Color,” The Intercept , September 6, 2018, .
18. To see a large-scale timeline of events in 2018, see: Kate Crawford and Meredith Whittaker, “AI in
2018: A Year in Review,” Medium , October 14, 2018,
19. Jon Evans, “The Techlash,” TechCrunch , June 17, 2018,
20. “Microsoft Calls for Facial Recognition Technology Rules given ‘Potential for Abuse,’” The Guardian ,
July 14, 2018,
-potential-for-abuse .
21. Natalie Ram, “Innovating Criminal Justice,” Northwestern University Law Review 112, no. 4 (February
1, 2018): 659–724, ; Rebecca
Wexler, “Life, Liberty, and Trade Secrets,” Stanford Law Review 70, no. 5 (May 2018): 1343–1429, ; Danielle Keats Citron
and Frank A. Pasquale, “The Scored Society: Due Process for Automated Predictions,” Washington
Law Review 89, no. 1 (2014): 1–33, .
22. See: Frank Pasquale, The Black Box Society: The Secret Algorithms That Control Money and
Information, (Cambridge: Harvard University Press, 2015).
23. D. Sculley, Jasper Snoek, Alex Wiltschko, and Ali Rahimi, “Winner’s Curse? On Pace, Progress and
Empirical Rigor,” 6th International Conference on Learning Representations (ICLR), (Vancouver, 2018), .
24. Kate Crawford, “The Test We Can—and Should—Run on Facebook,” The Atlantic , July 2, 2014,
ok/373819/ ; Molly Jackman and Lauri Kanerva, “Evolving the IRB: Building Robust Review for Industry
Research,” Washington & Lee Law Review Online 72, no. 8 (June 14, 2016): 442–457; Zoltan Boka,
“Facebook’s Research Ethics Board Needs to Stay Far Away from Facebook,” Wired , June 23, 2016, .
25. “Sandvig v. Sessions — Challenge to CFAA Prohibition on Uncovering Racial Discrimination Online,”
September 12, 2017, American Civil Liberties Union ,
nation-online .
26. See: Simone Browne, Dark Matters: On the Surveillance of Blackness (Durham: Duke University Press,
2015); Alvaro M. Bedoya, “What the FBI’s Surveillance of Martin Luther King Tells Us About the
Modern Spy Era,” Slate , January 18, 2016,
odern-spying.html ; James Ball, Julian Borger, and Glenn Greenwald, “Revealed: How US and UK Spy
Agencies Defeat Internet Privacy and Security,” The Guardian , September 6, 2013, ; Shoshana
Zuboff, “Big Other: Surveillance Capitalism and the Prospects of an Information Civilization,” Journal
of Information Technology 30, no. 1 (March 1, 2015): 75–89, .
27. Alice Shen, “Facial Recognition Tech Comes to Hong Kong-Shenzhen Border,” South China Morning
Post , July 24, 2018,
eter-tax-free-traders-hong .
28. Stephen Chen, “China’s Robotic Spy Birds Take Surveillance to New Heights,” South China Morning
Post , June 24, 2018,
ock-robotic-doves-do-they .
29. David Z. Morris, “China Will Block Travel for Those With Bad ‘Social Credit,’” Fortune , March 18, 2018, .
30. Nathan Vanderklippe, “Chinese Blacklist an Early Glimpse of Sweeping New Social-Credit Control,”
The Globe and Mail , January 3, 2018,
credit-control/article37493300/ .
31. “China Has Turned Xinjiang into a Police State like No Other,” T he Economist , May 31, 2018,
o-other .
32. Emily Feng and Louise Lucas, “Inside China’s Surveillance State,” Financial Times , July 20, 2018, .
33. Angus Berwick, “A New Venezuelan ID, Created with China’s ZTE, Tracks Citizen Behavior,” Reuters ,
November 14, 2018, .
34. Nafeez Ahmed, “Pentagon Wants to Predict Anti-Trump Protests Using Social Media Surveillance,”
Motherboard , October 30, 2018,
social-media-surveillance .
35. Karen Hao, “Amazon Is the Invisible Backbone behind ICE’s Immigration Crackdown,” MIT Technology
Review , October 22, 2018,
ation-crackdown/ .
36. “Who’s behind Ice?” (Empower LLC, Mijente, The National Immigration Project, and the Immigrant
Defense Project, October 23, 2018),
ta-Companies-Fueling-Deportations_v3-.pdf .
37. Brendan Shillingford et al., “Large-Scale Visual Speech Recognition,” arXiv preprint [Cs] ,
arXiv:1807.05162, July 13, 2018.
38. “Machine Vision Algorithm Learns to Recognize Hidden Facial Expressions,” MIT Technology Review ,
November 13, 2015,
expressions/ .
39. Richard T. Gray, About Face: German Physiognomic Thought from Lavater to Auschwitz (Detroit:
Wayne State University Press, 2004); Sharrona Pearl, About Faces: Physiognomy in
Nineteenth-Century Britain (Cambridge, MA: Harvard University Press, 2010).
40. Blaise Aguera y Arcas, Margaret Mitchell, and Alexander Todorov, “Physiognomy’s New Clothes,”
Medium , May 7, 2017, .
41. Ruth Leys, “How Did Fear Become a Scientific Object and What Kind of Object Is It?,” Representations
110, no. 1 (2010): 66–104, . Leys has offered a number of
critiques of Ekman’s research program, most recently in Ruth Leys, The Ascent of Affect: Genealogy
and Critique (Chicago: University of Chicago Press, 2017).
42. Alan J. Fridlund, Human Facial Expression: An Evolutionary View (San Diego: Academic Press, 1994).
43. Lisa Feldman Barrett, “Are Emotions Natural Kinds?,” Perspectives on Psychological Science 1, no. 1
(March 2006): 28–58, ; Erika H. Siegel, Molly K.
Sands, Wim Van den Noortgate, Paul Condon, Yale Chang, Jennifer Dy, Karen S. Quigley, and Lisa
Feldman Barrett. 2018. “Emotion Fingerprints or Emotion Populations? A Meta-Analytic Investigation
of Autonomic Features of Emotion Categories.” Psychological Bulletin 144 (4): 343–93, .
44. For example, despite criticism by the U.S. Government Accountability Office, the Transportation
Security Administration invested over one billion dollars in its SPOT program, aimed at identifying
potential terrorists based on these behavioral indicators. See: “Aviation Security: TSA Should Limit
Future Funding for Behavior Detection Activities” (Washington, DC: U.S. Government Accountability
Office, November 13, 2013), .
45. Jonathan Metzl, The Protest Psychosis: How Schizophrenia Became a Black Disease (Boston: Beacon
Press, 2009).
46. Mark Lieberman, “Sentiment Analysis Allows Instructors to Shape Course Content around Students’
Emotions,” Inside Higher Education , February 20, 2018,
uctors-shape-course-content .
47. Will Knight, “Emotional Intelligence Might Be a Virtual Assistant’s Secret Weapon,” MIT Technology
Review , June 13, 2016,
otions ;
48. “Affectiva Automotive AI,” Affectiva , accessed November 18, 2018, .
49. Jeff Weiner, “ACLU: Amazon’s Face-Recognition Software Matched Members of Congress with
Mugshots,” Orlando Sentinel , July 26, 2018,
g-software-congress-20180726-story.html ; “Amazon Rekognition Announces Real-Time Face
Recognition, Text in Image Recognition, and Improved Face Detection,” Amazon Web Services ,
November 21, 2017,
recognition-text-in-image-recognition-and-improved-face-detection/ .
50. Chris Adzima, “Using Amazon Rekognition to Identify Persons of Interest for Law Enforcement,”
Amazon Web Services , June 15, 2017,
nterest-for-law-enforcement/ .
51. Ranju Das, “Image & Video Rekognition Based on AWS” (Amazon Web Services Summit, Seoul, 2018), .
52. Jacob Snow, “Amazon’s Face Recognition Falsely Matched 28 Members of Congress With Mugshots,”
American Civil Liberties Union , July 26, 2018,
alsely-matched-28 .
53. Joy Buolamwini and Timnit Gebru, “Gender Shades: Intersectional Accuracy Disparities in
Commercial Gender Classification,” in Conference on Fairness, Accountability and Transparency (New
York, 2018), 77–91, .
54. Matt Wood, “Thoughts On Machine Learning Accuracy,” AWS News Blog , July 27, 2018, .
55. Sidney Fussell, “Amazon Accidentally Makes Rock-Solid Case for Not Giving Its Face Recognition
Tech to Police,” Gizmodo , July 27, 2018, .
56. Bryan Menegus, “Amazon Breaks Silence on Aiding Law Enforcement Following Employee Backlash,”
Gizmodo , August 11, 2018, .
57. Joseph and Lipp, “IBM Used NYPD Surveillance Footage to Develop Technology That Lets Police
Search by Skin Color.”
58. Jenna Bitar and Jay Stanley, “Are Stores You Shop at Secretly Using Face Recognition on You?,”
American Civil Liberties Union , March 26, 2018,
y-using-face .
59. John Brandon, “Walmart Will Scan for Unhappy Shoppers Using Facial Recognition (Cue the
Apocalypse),” VentureBeat , August 9, 2017,
on-cue-the-apocalypse/ .
60. AG and Legal Workforce Act, H.R. 6417, 115th Cong. (2017-2018), ; Securing America's Future Act
of 2018, H.R. 4760, 115th Cong. (2017-2018), ; Strong Visa Integrity Secures
America Act,” H.R. 2626, 115th Cong. (2017-2018), ; H.R. Security and Immigration
Reform Act of 2018, 6136, 115th Cong. (2017-2018), .
61. Biometric Information Privacy Act, 740 ILCS 14/1 (Statutes current through the end of the 2018
Regular Session of the 100th General Assembly),
:contentItem:5C66-0WY1-6YS3-D06V-00000-00&context=1516831 .
62. Brian Hofer, “BART Board Approves Surveillance Ordinance, Lake Merritt Development,” KTVU ,
September 13, 2018, .
63. “Letter from Nationwide Coalition to Amazon CEO Jeff Bezos Regarding Rekognition,” American Civil
Liberties Union , June 18, 2018, .
64. Abrar Al-Heeti, “Congress Still Wants Answers from Amazon about Its Facial Recognition Tech,”
CNET , November 29, 2018,
-tech/ .
65. Woodrow Hartzog and Evan Selinger, “Facial Recognition Is the Perfect Tool for Oppression,” Medium ,
August 2, 2018, .
66. Yilun Wang and Michael Kosinski, “Deep Neural Networks Are More Accurate than Humans at
Detecting Sexual Orientation from Facial Images,” Journal of Personality and Social Psychology 114,
no. 2 (2018): 246–57, .
67. Frank Pasquale, “When Machine Learning Is Facially Invalid,” Communications of the ACM 61, no. 9
(August 2018): 25–27, .
68. Kade Crockford, “Massachusetts Should Ban Face Recognition Technology,” WBUR , August 1, 2018, . See
also: Hartzog and Selinger, “Facial Recognition Is the Perfect Tool for Oppression.”
69. Brad Smith, “Facial Recognition Technology: The Need for Public Regulation and Corporate
Responsibility,” Microsoft on the Issues , July 13, 2018,
ublic-regulation-and-corporate-responsibility/ ; Sidney Fussell, “Axon CEO Says Face Recognition Isn’t
Accurate Enough for Body Cams Yet,” Gizmodo , August 8, 2018, .
70. “Litigating Algorithms: Challenging Government Use of Algorithmic Decision Systems” (New York: AI
Now Institute, September 2018), ; Dillon Reisman,
Jason Schultz, Kate Crawford, and Meredith Whittaker, “Algorithmic Impact Assessments: A Practical
Framework for Public Agency Accountability” (New York: AI Now Institute, April 2018), ; Micah Altman, Alexandra Wood, and Effy Vayena, “A
Harm-Reduction Framework for Algorithmic Fairness,” IEEE Security Privacy 16, no. 3 (May 2018):
34–45, .
71. Lecher, “A Healthcare Algorithm Started Cutting Care, and No One Knew Why.”
72. Ibid.
73. Carlos Gradín, “World Income Inequality Database (WIID3.4),” United Nations University , October 27,
2015, .
74. Serena Lei, “Nine Charts about Wealth Inequality in America,” The Urban Institute , 2015, ; Lisa J. Dettling, Joanne W. Hsu, Lindsay Jacobs, Kevin B. Moore, and
Jeffrey P. Thompson, "Recent Trends in Wealth-Holding by Race and Ethnicity: Evidence from the
Survey of Consumer Finances," (Washington: Board of Governors of the Federal Reserve System,
September 27, 2017), .
75. Andrew Restuccia, Sarah Ferris, and Helena Bottemiller Evich, “Behind Trump’s Plan to Target the
Federal Safety Net,” POLITICO , December 11, 2017, .
76. Brakkton Booker, “White House Budget Calls For Deep Cuts To HUD,” , February 13, 2018, .
77. Alison Kodjak, “Federal Judge Blocks Medicaid Work Requirements In Kentucky,” , June 29,
requirements-in-kentucky ; Rachel Garfield et al., “Implications of Work Requirements in
Medicaid: What Does the Data Say?,” The Henry J. Kaiser Family Foundation , June 12, 2018,
data-say/ .
78. Stacy Dean, “President’s Budget Would Shift Substantial Costs to States and Cut Food Assistance for
Millions,” Center on Budget and Policy Priorities , May 23, 2017,
and-cut-food .
79. David Pegg and Niamh McIntyre, “Child Abuse Algorithms: From Science Fiction to Cost-Cutting
Reality,” The Guardian , September 16, 2018,
cutting-reality .
80. David Scharfenberg, “Computers Can Solve Your Problem. You May Not like the Answer,” The Boston
Globe , September 21, 2018, ;
Joe Flood, The Fires: How a Computer Formula, Big Ideas, and the Best of Intentions Burned Down
New York City–and Determined the Future of Cities (New York: Riverhead Books, 2011).
81. Virginia Eubanks, “We Created Poverty. Algorithms Won’t Make That Go Away,” The Guardian , May 13,
ake-that-go-away ;
82. Leo Morales, “Federal Court Rules Against Idaho Department of Health and Welfare in Medicaid Class
Action,” ACLU of Idaho , March 30, 2016,
e-medicaid-class-action .
83. “Litigating Algorithms.”
84. Ibid., 10.
85. Michael Nash, “Examination of Using Structured Decision Making and Predictive Analytics in
Assessing Safety and Risk in Child Welfare” (Los Angeles: County of Los Angeles Office of Child
Protection, May 4, 2017),
MandPredictiveAnalytics_.pdf .
86. Virginia Eubanks, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor
(New York: St. Martin’s Press, 2018).
87. “Vulnerable Children Predictive Modelling” (Wellington: New Zealand Ministry of Social Development),
accessed November 18, 2018,
elling/ .
88. Eubanks, Automating Inequality .
89. Margaret Talbot, “The Case Against Cash Bail,” The New Yorker , August 25, 2015, .
90. Sam Levin, “Imprisoned by Algorithms: The Dark Side of California Ending Cash Bail,” The Guardian ,
September 7, 2018,
fornia-ending-cash-bail ; Maddie Hanna, “What Happened When New Jersey Stopped Relying on Cash
Bail,” The Philadelphia Inquirer , February 16, 2018,
ml .
91. Colleen O’Dea, “Civil Rights Coalition Calls for End to Core Element of NJ Bail Reform,” NJ Spotlight ,
July 31, 2018,
il-reform/ .
92. Jeremy B. White, “California Ended Cash Bail. Why Are So Many Reformers Unhappy About It?,”
POLITICO Magazine , August 29, 2018, .
93. 93. Rose Luckin, “Towards Artificial Intelligence-Based Assessment Systems,” Nature Human
Behaviour 1 (March 1, 2017): 1–3, .
94. Brent Bridgeman, Catherine Trapani, and Yigal Attali, “Comparison of Human and Machine Scoring of
Essays: Differences by Gender, Ethnicity, and Country,” Applied Measurement in Education 25, no. 1
(January 2012): 27–40, .
95. Nitin Madnani et al., “Building Better Open-Source Tools to Support Fairness in Automated Scoring,” in
Proceedings of the First ACL Workshop on Ethics in Natural Language Processing (Valencia:
Association for Computational Linguistics, 2017), 41–52, .
96. Daniel T. O’Brien et al., “An Evaluation of Equity in the Boston Public Schools’ Home-Based
Assignment Policy” (Boston: Boston Area Research Initiative, July 2018), .
97. Ibid .
98. Scharfenberg, “Computers Can Solve Your Problem. You May Not like the Answer.”
99. Ibid.
100. Dillon Reisman, Jason Schultz, Kate Crawford, and Meredith Whittaker, “Algorithmic Impact
Assessments: A Practical Framework for Public Agency Accountability” (New York: AI Now Institute,
April 2018), .
101. “Litigating Algorithms.”
102. “Algorithmic Accountability Policy Toolkit” (New York: AI Now Institute, October 2018), .
103. Reisman et al., “Algorithmic Impact Assessments.”
104. Wakabayashi, “Self-Driving Uber Car Kills Pedestrian in Arizona.”
105. Jack Stewart, “Tesla’s Self-Driving Autopilot Involved in Another Deadly Crash,” Wired , March 31, 2018, .
106. “Uber Autonomous-SUV Driver Streamed ‘The Voice’ Just before Deadly Arizona Crash, Report Says,”
Los Angeles Times , June 22, 2018, .
107. “Early Rider Program,” Waymo , accessed November 16, 2018, .
108. Aarian Marshall, “Wanna Save Lots of Lives? Put (Imperfect) Self-Driving Cars on the Road, ASAP,”
Wired , November 7, 2017, .
109. Lindsay Moore, “Autonomous Vehicles Continue to Drive Without Rules,” Phoenix New Times ,
November 9, 2017,
3743 .
110. “Uber Autonomous-SUV Driver Streamed ‘The Voice’ Just before Deadly Arizona Crash, Report Says.”
111. Ross and Swetlitz, “IBM’s Watson Recommended ‘unsafe and Incorrect’ Cancer Treatments.”
112. Angela Chen, “What the Apple Watch’s FDA Clearance Actually Means,” The Verge , September 13,
safe .
113. Mark Vermette, “Apple Watch Approval Marks Shift In Device Development And Approvals,” Med
Device Online , October 24, 2018,
d-approvals-0001 ; Rajiv Leventhal, “4.4M Patient Records Breached in Q3 2018, Protenus Finds,”
Healthcare Informatics Magazine , November 7, 2018,
-2018-protenus-finds .
114. Alex Hern, “Google ‘betrays Patient Trust’ with DeepMind Health Move,” The Guardian , November 14,
hcare-move .
115. Benjamin Herold, “Pearson Tested ‘Social-Psychological’ Messages in Learning Software, With Mixed
Results,” Education Week , April 17, 2018,
?cmp=SOC-SHR-FB .
116. Rolfe Winkler and Laura Stevens, “New Parents Complain Amazon Baby-Registry Ads Are Deceptive,”
Wall Street Journal , November 28, 2018, .
117. Julia Angwin, Jeff Larson, and Laura Kirchner, “Machine Bias: There’s Software Used Across the
Country to Predict Future Criminals. And It’s Biased Against Blacks.,” ProPublica , May 23, 2016, ; Jeffrey
Dastin, “Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women,” Reuters ,
October 10, 2018, .
118. Shira Mitchell, “Mirror Mirror: Reflections on Quantitative Fairness,” 2018, ; Arvind Narayanan, Tutorial: 21 Fairness Definitions and Their
Politics , accessed November 18, 2018, .
119. Solon Barocas, Kate Crawford, Aaron Shapiro, Hanna Wallach, 2017, “The Problem with Bias:
Allocative Versus Representational Harms in Machine Learning” (Measure, Model, Mix: Computer as
Instrument: 9th Annual SIGCIS Conference, Philadelphia, 2017).
120. Kate Crawford, “The Trouble with Bias” (Conference on Neural Information Processing Systems, Long
Beach, CA, 2017), .
121. Ibid.
122. Sam Corbett-Davies and Sharad Goel, “The Measure and Mismeasure of Fairness: A Critical Review of
Fair Machine Learning,” arXiv preprint [CS] arXiv:1808.00023, July 31, 2018.
123. Solon Barocas and Moritz Hardt, “Fairness in Machine Learning” (Conference on Neural Information
Processing Systems, Long Beach, CA, 2017), ; Narayanan, 21 Fairness
Definitions and Their Politics .
124. Tracy Jan, “Redlining Was Banned 50 Years Ago. It’s Still Hurting Minorities Today,” Washington Post ,
March 28, 2018,
ts-still-hurting-minorities-today/ .
125. Cynthia Dwork, Nicole Immorlica, Adam T. Kalai, Mark DM Leiserson, “Decoupled classifiers for
group-fair and efficient machine learning”, Conference on Fairness, Accountability and Transparency
(January 21, 2018), 119-133, .
126. Jon Kleinberg, “Inherent Trade-Offs in Algorithmic Fairness,” in Abstracts of the 2018 ACM
International Conference on Measurement and Modeling of Computer Systems , SIGMETRICS ’18 (New
York: ACM, 2018), 40–40, ; Alexandra Chouldechova, “Fair
Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments,” arXiv preprint
[Cs, Stat] arXiv:1610.07524, October 24, 2016.
127. Frank Pasquale, “Odd Numbers,” Real Life , August 20, 2018, .
128. Sebastian Benthall, “Critical Reflections on FAT* 2018: A Historical Idealist Perspective,” Dataactive ,
April 11, 2018, .
129. Mitchell, “Mirror Mirror: Reflections on Quantitative Fairness.”
130. Ben Hutchinson and Margaret Mitchell, “50 Years of Test (Un)Fairness: Lessons for Machine
Learning,” arXiv preprint [CS] , arXiv:1811.10104, November 25, 2018.
131. Roel Dobbe, Sarah Dean, Thomas Gilbert, and Nitin Kohli, “A Broader View on Bias in Automated
Decision-Making: Reflecting on Epistemology and Dynamics,” arXiv preprint [Cs, Math, Stat] ,
arXiv:1807.00553, July 2, 2018.
132. Batya Friedman and Helen Nissenbaum, “Discerning Bias in Computer Systems,” in INTERACT ’93 and
CHI ’93 Conference Companion on Human Factors in Computing Systems , (New York: ACM, 1993),
141–142, .
133. Batya Friedman and Helen Nissenbaum, “Bias in Computer Systems,” ACM Transactions on
Information Systems 14, no. 3 (July 1996): 330–347, .
134. Kate Crawford and Ryan Calo, “There Is a Blind Spot in AI Research,” Nature 538, no. 7625 (October
20, 2016): 311, .
135. See: Meredith Whittaker, “Data Genesis: AI’s Primordial Soup” (Eyeo Festival, Minneapolis, 2018), .
136. Timnit Gebru et al., “Datasheets for Datasets,” arXiv preprint [Cs] , arXiv:1803.09010, March 23, 2018.
137. Margaret Mitchell et al., “Model Cards for Model Reporting,” arXiv preprint [Cs] arXiv:1810.03993,
October 5, 2018.
138. Jevan A. Hutson, Jessie G. Taft, Solon Barocas and Karen Levy,, “Debiasing Desire: Addressing Bias &
Discrimination on Intimate Platforms,” Proceedings of the ACM On Human-Computer Interaction
(CSCW) 2 (November 2018): 1–18, .
139. Animesh Singh and Michael Hind, “AI Fairness 360: Attacking Bias from All Angles!,” IBM Developer ,
September 19, 2018, .
140. James Wexler, “The What-If Tool: Code-Free Probing of Machine Learning Models,” Google AI Blog ,
September 11, 2018, ;
James Wexler, “Facets: An Open Source Visualization Tool for Machine Learning Training Data,”
Google AI Blog , June 17, 2017, .
141. Alekh Agarwal et al., “A Reductions Approach to Fair Classification,” arXiv preprints [CS] ,
arXiv:1803.02453, March 6, 2018.
142. Dave Gershgorn, “Facebook Says It Has a Tool to Detect Bias in Its Artificial Intelligence,” Quartz , May
3, 2018, .
143. Jeremy Kahn, “Accenture Unveils Tool to Help Companies Insure Their AI Is Fair,” Bloomberg , June 13,
sure-their-ai-is-fair .
144. “The Ethics Certification Program for Autonomous and Intelligent Systems (ECPAIS),” IEEE Standards
Association , accessed November 19, 2018, .
145. “TRUSTe Settles FTC Charges It Deceived Consumers Through Its Privacy Seal Program,” Federal
Trade Commission , November 17, 2014,
sumers-through-its .
146. “AI Now 2017 Report.”
147. Scott Shane, Cade Metz, and Daisuke Wakabayashi, “How a Pentagon Contract Became an Identity
Crisis for Google,” The New York Times , November 2, 2018, .
148. Sundar Pichai, “AI at Google: Our Principles,” The Keyword , June 7, 2018, .
149. Alan Boyle, “Microsoft Is Turning down Some Sales over AI Ethics, Top Researcher Eric Horvitz Says,”
GeekWire , April 10, 2018,
s/ ; Jordan Novet, “Facebook Forms Ethics Team to Prevent Bias in AI Software,” CNBC , May 3, 2018, ; “Axon
AI and Policing Technology Ethics Board,” Axon , accessed November 19, 2018, .
150. “Ethically Aligned Design Version 2: A Vision for Prioritizing Human Well-Being with Autonomous and
Intelligent Systems” (New York: IEEE Global Initiative on Ethics of Autonomous and Intelligent
Systems, 2018), ; Brent Hecht, Lauren Wilcox, Jeffrey P. Bigham,
Johannes Schöning, Ehsan Hoque, Jason Ernst, Yonatan Bisk, Luigi De Russis, Lana Yarosh, Bushra
Anjum, Danish Contractor, and Cathy Wu. “It’s Time to Do Something: Mitigating the Negative Impacts
of Computing Through a Change to the Peer Review Process,” ACM Future of Computing Blog , March
29, 2018, .
151. Alina Tugend, “Colleges Grapple With Teaching the Technology and Ethics of A.I.,” The New York
Times , November 3, 2018, .
152. “La Déclaration de Montréal pour un Développement Responsable de l’Intelligence Artificielle,”
accessed November 19, 2018, .
153. “Microsoft Professional Program for Artificial Intelligence,” Microsoft , accessed November 29, 2018, .
154. Lucy Suchman, “Corporate Accountability,” Robot Futures , June 11, 2018, .
155. Ben Wagner, “Ethics as Escape From Regulation: From Ethics-Washing to Ethics-Shopping?” In Being
Profiling. Cogitas Ergo Sum , ed. Mireille Hildebrandt. (Amsterdam: Amsterdam University Press,
forthcoming 2019),
gulation_2018_BW9.pdf .
156. Sandy Parakilas, “We Can’t Trust Facebook to Regulate Itself,” The New York Times , January 20, 2018. .
157. Dan M. Greene, Anna Laura Hoffman, and Luke Stark, “Better, Nicer, Clearer, Fairer: A Critical
Assessment of the Movement for Ethical Artificial Intelligence and Machine Learning” (Hawaii
International Conference on System Sciences, Maui, forthcoming 2019),
-HICSS-Final-Submission.pdf .
158. Ibid.
159. Andrew McNamara, Justin Smith, and Emerson Murphy-Hill, “Does ACM’s Code of Ethics Change
Ethical Decision Making in Software Development?,” in Proceedings of the 2018 26th ACM Joint
Meeting on European Software Engineering Conference and Symposium on the Foundations of
Software Engineering , ESEC/FSE 2018 (New York: ACM, 2018), 729, .
160. Pichai, “AI at Google: Our Principles.”
161. Ryan Gallagher, “Google Plans to Launch Censored Search Engine in China, Leaked Documents
Reveal,” The Intercept , August 1, 2018, .
162. Ronald Deibert, Rebecca Mackinnon, Xiao Qiang, and Lokman Tsui, “Open Letter to Google on
Reported Plans to Launch a Censored Search Engine in China,” Amnesty International , August 28,
age=en ; Kate Conger and Daisuke Wakabayashi, “Google Employees Protest Secret Work on
Censored Search Engine for China,” The New York Times , September 10, 2018,
a.html .
163. Mark Bergen, “Google CEO Tells Staff China Plans Are ‘Exploratory’ After Backlash,” Bloomberg August
17, 2018,
re-exploratory .
164. Matt Phillips, “Facebook’s Stock Plunge Shatters Faith in Tech Companies’ Invincibility,” The New York
Times , October 17, 2018, ; Rupert Neate,
“Twitter Stock Plunges 20% in Wake of 1m User Decline,” The Guardian , July 27, 2018,
in-three-months .
165. For a more general description of justice as fairness, see: John Rawls, Justice as Fairness: A
Restatement , ed. Erin I. Kelly (Cambridge, MA: Harvard University Press, 2001).
166. Brent Hecht et al., “It’s Time to Do Something,” .
167. Ben Green, “‘Fair’ Risk Assessments: A Precarious Approach for Criminal Justice Reform” (5th
Workshop on Fairness, Accountability, and Transparency in Machine Learning, Stockholm, 2018), .
168. Ben Green, “Putting the J(ustice) in FAT,” Berkman Klein Center , February 26, 2018, .
169. Kate Crawford, “Just An Engineer: The Politics of AI,” (The Royal Society, London, July 2018), .
170. Paul N. Edwards, Geoffrey C. Bowker, Steven J. Jackson, and Robin Williams. “Introduction: An
Agenda for Infrastructure Studies,” Journal of the Association for Information Systems 10, no. 5 (May
28, 2009): 364–74, .
171. Susan Leigh Star, “The Ethnography of Infrastructure,” American Behavioral Scientist 43, no. 3
(November 1, 1999): 377–91, ; Paul N. Edwards,
Thomas J. Misa, and Philip Brey, “Infrastructure and Modernity: Force, Time, and Social Organization
in the History of Sociotechnical Systems,” in Modernity and Technology (Cambridge, MA: MIT Press,
2003), 185–225; Shannon Mattern, “The Big Data of Ice, Rocks, Soils, and Sediments”, Places,
November 2017, ;
Jean-Christophe Plantin, Carl Lagoze, Paul N. Edwards and Christian Sandvig. “Infrastructure studies
meet platform studies in the age of Google and Facebook,” New Media & Society, 20, no. 1 (2018):
293-310, .
172. Kate Crawford and Vladan Joler, “Anatomy of an AI System: The Amazon Echo As An Anatomical
Map of Human Labor, Data and Planetary Resources,” (AI Now Institute and Share Lab, September 7,
2018), .
173. Steven J. Jackson, “Rethinking Repair”, in Media Technologies: Essays on Communication, Materiality,
and Society, Tarleton Gillespie, Pablo J. Boczkowski and Kirsten A. Foot, eds. (Cambridge: MIT Press,
2014), 221-239.
174. Nick Seaver, “What Should An Anthropology of Algorithms Do?” Cultural Anthropology, 33 , No. 3
(2018): 375-385, .
175. Li Yuan, “How Cheap Labor Drives China’s A.I. Ambitions,” The New York Times , November 27, 2018, ; Mary L.
Gray and Siddharth Suri, “The Humans Working Behind the AI Curtain, Harvard Business Review,
January 9, 2017, .
176. Lilly Irani, “The Hidden Faces of Automation,” XRDS 23, no. 2 (December 2016): 34–37, .
177. Ibid.
178. Janine Berg et al., “Digital Labour Platforms and the Future of Work: Towards Decent Work in the
Online World,” (Geneva: International Labour Organization, September 20, 2018), .
179. Sarah T. Roberts, “Commercial Content Moderation: Digital Laborers’ Dirty Work,” in The Intersectional
Internet: Race, Sex, Class and Culture Online , ed. Safiya Umoja Noble and Brendesha M. Tynes (New
York: Peter Lang, 2016), 147–159.
180. M. S. Silberman et al., “Responsible Research with Crowds: Pay Crowdworkers at Least Minimum
Wage,” Communications of the ACM 61, no. 3 (February 2018): 39–41, .
181. Arlene Kaplan Daniels, “Invisible Work,” Social Problems 34, no. 5 (1987): 403–15, ; Arlie Russell Hochschild, The Managed Heart: Commercialization of
Human Feeling (Berkeley: University of California Press, 2012).
182. Astra Taylor, “The Automation Charade,” Logic Magazine , October 2, 2018, .
183. Jana Kasperkevic, “Ex-McDonald’s CEO Suggests Replacing Employees with Robots amid Protests,”
The Guardian , May 25, 2016,
ployees-robots ; Ed Rensi, “Thanks To ‘Fight For $15’ Minimum Wage, McDonald’s Unveils
Job-Replacing Self-Service Kiosks Nationwide,” Forbes , November 29, 2016,
ds-unveils-job-replacing-self-service-kiosks-nationwide/#5defa3eb4fbc .
184. Uliana Pavlova, “McDonald’s Kiosks Mean More Staff at Chicago Flagship, Not Fewer,” Bloomberg ,
August 8, 2018,
ago-flagship-not-fewer .
185. Alex Rosenblat, Uberland: How Algorithms Are Rewriting the Rules of Work (Berkeley: University of
California Press, 2018).
186. Ben Chapman, “Uber Eats and Deliveroo Riders Are Going on Strike This Week,” The Independent ,
October 3, 2018,
poons-mcstrike-industrial-action-a8567286.html ; Carolyn Said, “Uber, Lyft Drivers Fear Getting Booted
from Work,” San Francisco Chronicle , October 14, 2018,
4052.php .
187. Kate Conger, “Google and Facebook’s Security Guards Are Fighting to Earn a Living Wage,” Gizmodo ,
July 27, 2018, .
188. Mark Bergen and Josh Eidelson, “Inside Google’s Shadow Workforce,” Bloomberg , July 25, 2018, .
189. Thuy Ong, “Amazon Patents Wristbands That Track Warehouse Employees’ Hands in Real Time,” The
Verge , February 1, 2018,
mployees ; Hayley Peterson, “Missing Wages, Grueling Shifts, and Bottles of Urine: The Disturbing
Accounts of Amazon Delivery Drivers May Reveal the True Human Cost of ‘Free’ Shipping,” Business
Insider , September 11, 2018,
ns-2018-8 .
190. Catie Keck, “Amazon Workers Across Europe Protest Black Friday, Citing Grueling Work Conditions,”
Gizmodo , November 23, 2018, .
191. “AI Now 2017 Report.”
192. L. Rafael Reif, “Letter to the MIT Community Regarding the MIT Stephen A. Schwarzman College of
Computing,” MIT News , October 15, 2018,
ting ; Steve Lohr, “M.I.T. Plans College for Artificial Intelligence, Backed by $1 Billion,” The New York
Times , October 16, 2018, .
193. Lev Manovich, “Can We Think Without Categories?” Digital Culture & Society, 4, no. 1 (2018): 17-28.
194. AI Now Institute, “Gender, Race and Power: Outlining a New AI Research Agenda,” Medium, November
15, 2018. .
195. Kristian Lum, “Statistics, we have a problem,” Medium, December 13, 2017, .
196. Jennings Brown, “ ‘NIPS’ AI Conference Changes Name Following Protests Over Gross Acronym,”
Gizmodo, November 19, 2018. ; Corinna
Cortes et al., “From the Board: Changing Our Acronym,” NeurIPS November 16, 2018, .
197. “NIPS Name Change,” NeurIPS , October 17, 2018, .
198. Meredith Whittaker, one of the co-founders of AI Now, was one of the organizers of the Google
Walkout. See also: Caroline O’Donovan and Ryan Mac, “Google Engineers Are Organizing A Walkout
To Protest The Company’s Protection Of An Alleged Sexual Harasser”, Buzzfeed, Oct. 30, 2018,
-harassment .
199. Davey Alba and Caroline O’Donovan, “Square, Airbnb, And eBay Just Said They Would End Forced
Arbitration For Sexual Harassment Claims,” Buzzfeed, Nov. 15, 2018,
200. Sharon Florentine, “Alphabet Dismisses Action on Diversity and Inclusion,” CIO , June 15, 2018,
n.html ; Sara Ashley O’Brien, “Apple’s Board Calls Diversity Proposal ‘Unduly Burdensome and Not
Necessary,’” CNNMoney , January 15, 2016, .
201. Tom Simonite, “AI Is the Future - But Where Are the Women?,” Wired , August 17, 2018, .
202. Jackie Snow, “‘We’re in a Diversity Crisis:’ Cofounder of Black in AI on What’s Poisoning Algorithms in
Our Lives,” MIT Technology Review , February 14, 2018,
s-poisoning-the-algorithms-in-our/ .
203. Paula A. Johnson, Sheila E. Widnall, and Frazier F. Benya, eds., Sexual Harassment of Women: Climate,
Culture, and Consequences in Academic Sciences, Engineering, and Medicine (Washington, DC: The
National Academies Press, 2018),
es-in-academic .
204. Caroline Clark Hayes, “Computer Science: The Incredible Shrinking Woman,” in Gender Codes: Why
Women Are Leaving Computing , ed. Thomas J Misa (Hoboken: Wiley, 2010), 25–50.
205. Dale Strok, “Women in AI,” IEEE Expert: Intelligent Systems and Their Applications 7, no. 4 (August
1992): 7–22, .
206. “Barriers to Equality in Academia: Women in Computer Science at M.I.T.” (Cambridge, MA: The
Laboratory of Computer Science and the Artificial Intelligence Laboratory at M.I.T., February 1983), .
207. Langdon Winner, “Do Artifacts Have Politics?,” Daedalus 109, no. 1 (1980): 121–36.
208. Stephen Merity, “Bias is not just in our datasets, it’s in our conferences and community,” ,
December 11, 2017, .
209. Dastin, “Amazon scraps secret AI recruiting tool that showed bias against women.”
210. Safiya U. Noble, Algorithms of Oppression: How Search Engines Reinforce Racism, (New York: NYU
Press, 2018);
211. Clemens Apprich, Wendy Hui Kyong Chun, Florian Cramer, and Hito Steyerl, Pattern Discrimination
(Minneapolis: University of Minnesota Press, forthcoming 2019).
212. “Litigating Algorithms.”
213. Reisman et al., “Algorithmic Impact Assessments.”
214. “CPSR History,” Computer Professionals for Social Responsibility , June 1, 2005, .
215. “The Never Again Pledge,” accessed November 29, 2018,
216. Scott Shane and Daisuke Wakabayashi, “‘The Business of War:’ Google Employees Protest Work for
the Pentagon,” The New York Times , November 2, 2018, .
217. Lucy Suchman, Lilly Irani, and Peter Asaro, “Google’s March to the Business of War Must Be Stopped,”
The Guardian , May 16, 2018, ;
Mary Wareham, “Letter to Sergey Brin and Sundar Pichai,” March 13, 2018, ;
Daisuke Wakabayashi and Scott Shane, “Google Will Not Renew Pentagon Contract That Upset
Employees,” The New York Times , November 2, 2018, .
218. Shaban, “Amazon Employees Demand Company Cut Ties with ICE;” “Who’s Behind Ice?;” Caroline
O’Donovan, “Employees Of Another Major Tech Company Are Petitioning Government Contracts,”
BuzzFeed News , June 26, 2018,
-company-contract ; Sheera Frenkel, “Microsoft Employees Question C.E.O. Over Company’s Contract
With ICE,” The New York Times , July 27, 2018, ; Peter Kotecki,
“Burning Man Protesters Raise Awareness of Palantir, Amazon ICE Ties,” Business Insider , August 31,
2018, .
219. Greg Sandoval, “Over 100 Amazon Employees Sign Letter Asking Jeff Bezos to Stop Selling
Facial-Recognition Software to Police,” Business Insider , June 22, 2018,
ial-recognition-software-police-2018-6 ; Kade Crockford, “Over 150,000 People Tell Amazon: Stop
Selling Facial Recognition Tech to Police,” American Civil Liberties Union , June 18, 2018,
azon-stop-selling-facial ; Matt Cagle and Nicole Ozer, “Amazon Teams Up With Government to Deploy
Dangerous New Facial Recognition Technology,” ACLU Free Future , May 22, 2018,
t-deploy-dangerous-new .
220. Ryan Gallagher “Google China Prototype Links Searches to Phone Numbers” The Intercept ,
September 14, 2018, .
221. Bryan Menegus, “Here’s the Letter 1,400 Google Workers Sent Leadership in Protest of Censored
Search Engine for China,” Gizmodo , August 16, 2018, ; Google
Employees Against Dragonfly, “We Are Google Employees. Google Must Drop Dragonfly.,” Medium ,
November 27, 2018,
nfly-4c8a30c5e5eb ; “Google Must Not Capitulate to China’s Censorship Demands,” Amnesty
International , November 28, 2018,
-demands/ .
222. Co-founder of AI Now, Meredith Whittaker, was one of the eight core organizers of the Google
223. Google Walkout for Real Change, “Google Employees and Contractors Participate in ‘Global Walkout
for Real Change,’” Medium , November 2, 2018,
out-for-real-change-389c65517843 .
224. Richard Waters, “Google Ends Forced Arbitration for Sexual Harassment Claims,” Financial Times ,
November 8, 2018, .
225. Kate Gibson, “Tech Signals End of Forced Arbitration for Sexual Misconduct Claims,” CBS
MoneyWatch , November 16, 2018,
s/ .
226. “AI Now 2017 Report.”
227. Google Walkout for Real Change, “#GoogleWalkout Update: Collective Action Works, but We Need to
Keep Working.,” Medium , November 8, 2018,
o-keep-working-b17f673ad513 ; Noam Scheiber, “Google Workers Reject Silicon Valley Individualism
in Walkout,” The New York Times , November 7, 2018, .
228. For example, Google’s highly secret Dragonfly project to censor search results in China and link
Chinese residents’ phone numbers to search logs. See: “We Are Google Employees. Google Must
Drop Dragonfly.,” Medium , November 27, 2018,
nfly-4c8a30c5e5eb .
229. Cade Metz, “Tech Giants Are Paying Huge Salaries for Scarce A.I. Talent,” The New York Times ,
October 22, 2017, .
230. “Recommendations” in “AI Now 2016 Report,” (New York: AI Now Institute, 2016), .
231. Crawford and Joler, “Anatomy of an AI System,” .
232. See, as just one of many examples: Dastin, “Amazon scraps secret AI recruiting tool that showed bias
against women,”
tool-that-showed-bias-against-women-idUSKCN1MK08G .
233. Lecher, “A Healthcare Algorithm Started Cutting Care, and No One Knew Why.”
234. Sonia Katyal, “Private Accountability in the Age of the Algorithm,” UCLA Law Review 66 (forthcoming
2019), .

Enter content here

Enter content here

Enter content here

Enter supporting content here