Black women, AI, and overcoming historical patterns of abuse

After a 2019 research paper demonstrated that commercially available facial analysis tools fail to work for women with dark skin, AWS executives went on the attack. Instead of offering up more equitable performance results or allowing the federal government to assess their algorithm like other companies with facial recognition tech have done, AWS executives attempted to discredit study coauthors Joy Buolamwini and Deb Raji in multiple blog posts. More than 70 respected AI researchers rebuked this attack, defended the study, and called on Amazon to stop selling the technology to police, a position the company temporarily adopted last year after the death of George Floyd.

But according to the Abuse and Misogynoir Playbook, published earlier this year by a trio of MIT researchers, Amazon’s attempt to smear two Black women AI researchers and discredit their work follows a set of tactics that have been used against Black women for centuries. Moya Bailey coined the term “misogynoir” in 2010 as a portmanteau of “misogyny” and “noir.” Playbook coauthors Katlyn Turner, Danielle Wood, and Catherine D’Ignazio say these tactics were also used to disparage former Ethical AI team co-lead Timnit Gebru after Google fired her in late 2020 and stress that it’s a pattern engineers and data scientists need to recognize.

The Abuse and Misogynoir Playbook is part of the State of AI report from the Montreal AI Ethics Institute and was compiled by MIT professors in response to Google’s treatment of Gebru, a story VentureBeat has covered in depth. The coauthors hope that recognition of the phenomena will prove a first step in ensuring these tactics are no longer used against Black women. Last May, VentureBeat wrote about a fight for the soul of machine learning, highlighting ties between white supremacy and companies like Banjo and Clearview AI, as well as calls for reform from many in the industry, including prominent Black women.

MIT assistant professor Danielle Wood, whose work focuses on justice and space research, told VentureBeat it’s important to recognize that the tactics outlined in the Abuse and Misogynoir Playbook can be used in almost any arena. She noted that while some cling to a belief in the impartiality of data-driven results, the AI field is in no way exempt from this problem.

“This is a process, a series of related things, and the process has to be described step by step or else people won’t get the point,” Wood said. “I can be part of a system that’s actually practicing misogynoir, and I’m a Black woman. Because it’s a habit that is so prolific, it’s something I might participate in without even thinking about it. All of us can.”

Above: The Abuse and Misogynoir Playbook (Design by Melissa Teng)

Image Credit: Design by Melissa Teng

The playbook outlines the intersectional and unique abuse aimed at Black women in five steps:

Step 1: A Black woman scholar makes a contribution that speaks truth to power or upsets the status quo. 

Step 2: Disbelief in her contribution from people who say the results can’t be true and either think a Black woman couldn’t have done the research or find another way to call her contribution into question.

Step 3: Dismissal, discrediting, and gaslighting ensues. AI chief Jeff Dean’s public attempt to discredit Gebru alongside colleagues is a textbook example. Similarly, after current and former Dropbox employees alleged gender discrimination at the company, Dropbox CEO Drew Houston attempted to discredit the report’s findings, according to documents obtained by VentureBeat.

Gaslighting is a term taken from the 1944 movie Gaslight, in which a character goes to extreme lengths to make a woman deny her senses, ignore the truth, and feel like she’s going crazy. It’s not uncommon at this stage for people to consider the targeted Black woman’s contribution an attempt to weaponize pity or sympathy. Another instance that sparked gaslighting allegations involved algorithmic bias, Facebook chief AI scientist Yann LeCun, and Gebru.

Step 4: Erasure. Over time, counter-narratives, deplatforming, and exclusion are used to prevent that person from carrying out their work as part of attempts to erase their contributions.

Step 5: Revisionism seeks to paper over the contributions of Black women and can lead to whitewashed versions of events and slow progress toward justice.

There’s been a steady stream of stories about gender and racial bias in AI in recent years, a point highlighted by news headlines this week. The Wall Street Journal reported Friday that researchers found Facebook’s algorithm shows different job ads to men and women and is discriminatory under U.S. law, while Vice reported on research that found facial recognition used by Proctorio remote proctoring software does not work well for people with dark skin over half of the time. This follows VentureBeat’s coverage of racial bias in ExamSoft’s facial recognition-based remote proctoring software, which was used in state bar exams in 2020.

Investigations by The Markup this week found advertising bans hidden behind an algorithm for a number of terms on YouTube, including “Black in tech,” “antiracism,” and “Black excellence,” but it’s still possible to advertise to white supremacists on the video platform.

Case study: Timnit Gebru and Google

Google’s treatment of Gebru illustrates each step of the playbook. Her status quo-disrupting contribution, Turner told VentureBeat, was an AI research paper about the dangers of using large language models that perpetuate racism or stereotypes and carry an environmental impact that may unduly burden marginalized communities. Other perceived disruptions, Turner said, included Gebru building one of the most diverse teams within Google Research and sending a critical email to the Google Brain Women and Allies internal listserv that was leaked to Platformer.

Shortly after she was fired, Gebru said she was asked to retract the paper or remove the names of Google employees. That was step two from the Misogynoir Playbook. In academia, Turner said, retraction is taken very seriously. It’s generally reserved for scientific falsehood and can end careers, so asking Gebru to remove her name from a valid piece of research was unreasonable and part of efforts to make Gebru herself seem unreasonable.

Evidence of step three, disbelief or discredit, can be found in an email AI chief Jeff Dean sent that calls into question the validity of the paper’s findings. Days later, CEO Sundar Pichai sent a memo to Google employees in which he said the firing of Gebru had prompted the company to explore improvements to its employee de-escalation policy. In an interview with VentureBeat, Gebru characterized that memo as “dehumanizing” and an attempt to fit her into an “angry Black woman” trope.

Despite Dean’s critique, a point that seems lost amid allegations of abuse, racism, and corporate efforts to interfere with academic publication is that the team of researchers behind the stochastic parrots research paper in question was exceptionally well-qualified to deliver critical analysis of large language models. A version of the paper VentureBeat obtained lists Google research scientists Ben Hutchinson, Mark Diaz, and Vinodkumar Prabhakaran as coauthors, as well as then-Ethical AI team co-leads Gebru and Margaret Mitchell. While Mitchell is well known for her work in AI ethics, she is most heavily cited for research involving language models. Diaz, Hutchinson, and Prabhakaran have backgrounds in assessing language or NLP for ageism, discrimination against people with disabilities, and racism, respectively. Linguist Emily Bender, a lead coauthor of the paper alongside Gebru, received an award from organizers of a major NLP conference in mid-2020 for work critical of large language models, which VentureBeat also reported.

Gebru is coauthor of the Gender Shades research paper that found commercially available facial analysis models perform particularly poorly for women with dark skin. That project, spearheaded by Buolamwini in 2018 and continued with Raji in a subsequent paper published in early 2019, has helped shape legislative policy in the U.S and is also a central part of Coded Bias, a documentary now streaming on Netflix. And Gebru has been a major supporter of AI documentation standards like datasheets for datasets and model cards, an approach Google has adopted.

Finally, Turner said, steps four and five of the playbook, erasure and revisionism, can be seen in the departmental reorganization and diversity policy changes Google made in February. As a result of those changes, Google VP Marian Croak was appointed to head up 10 of the Google teams that consider how technology impacts people. She reports directly to AI chief Jeff Dean.

On Tuesday, Google research manager Samy Bengio resigned from his role at the company, according to news first reported by Bloomberg. Prior to the restructuring, Bengio was the direct report manager for the Ethical AI team.

VentureBeat obtained a copy of a letter Ethical AI team members sent to Google leadership in the weeks following Gebru’s dismissal that specifically requested Bengio remain the direct report for the team and that the company not implement any reorganization. A person familiar with ethics and policy matters at Google told VentureBeat that reorganization had been discussed previously, but this source described an environment of fear after Gebru’s dismissal that prevented people from speaking out.

Before being named to her new position, Croak appeared alongside the AI chief in a meeting with Black Google employees in the days following Gebru’s dismissal. Google declined to make Croak available for comment, but Google released a video in which she called for more “diplomatic” conversations about definitions of fairness or safety.

Turner pointed out that the reorganization fits neatly into the playbook.

“I think that revisionism and erasure is important. It serves a function of allowing both people and the news cycle to believe that the narrative arc has happened, like there was some bad thing that was taken care of — ‘Don’t worry about this anymore.’ [It’s] like, ‘Here’s this new thing,’ and that’s really effective,” Turner said.

Origins of the playbook

The playbook’s coauthors said it was constructed following conversations with Gebru. Earlier in the year, Gebru spoke at MIT at Turner and Wood’s invitation as part of an antiracism tech design research seminar series. When the news broke that Gebru had been fired, D’Ignazio described feelings of anger, shock, and outrage. Wood said she experienced a sense of grieving and loss. She also felt frustrated by the fact that Gebru was targeted despite having attempted to address harm through channels that are considered legitimate.

“It’s a really discouraging feeling of being stuck,” Wood said. “If you follow the rules, you’re supposed to see the outcome, so I think part of the reality here is just thinking, ‘Well, if Black women try to follow all the rules and the result is we’re still not able to communicate our urgent concerns, what other options do we have?’”

Wood said she and Turner found connections between historical figures and Gebru in their work in the Space Enabled Lab at MIT examining complex sociotechnical systems through the lens of critical race studies and queer Black feminist groups like the Combahee River Collective.

In addition to instances of misogynoir and abuse at Amazon and Google, coauthors say the playbook represents a historical pattern that has been used to exclude Black women authors and scholars dating back to the 1700s. These include Phillis Wheatley, the first published African American poet, journalist Ida B. Wells, and author Zora Neale Hurston. Generally, the coauthors found that the playbook tactics visit great acts of violence on Black women that can be distinguished from the harms encountered by other groups that challenge the status quo.

The coauthors said women outside of tech who have been targeted by the same playbook include New York Times journalist and 1619 Project creator Nikole Hannah-Jones and politicians like Stacey Abrams and Rep. Ayanna Pressley (D-MA).

The long shadow of history

The researchers also said they took a historical view to demonstrate that the ideas behind the Abuse and Misogynoir Playbook are centuries old. Failure to confront forces of racism and sexism at work, Turner said, can lead to the same problems in new and different tech scenarios. She went on to say that it’s important to understand that historical forces of oppression, categorization, and hierarchy are still with us and warned that “we will never actually get to an ethical AI if we don’t understand that.”

The AI field claims to excel at pattern recognition, so the industry should be able to identify tactics from the playbook, D’Ignazio said.

“I feel like that’s one of the most enormous ignorances, the places where technical fields do not go, and yet history is what would inform all of our ethical decisions today,” she said. “History helps us see structural, macro patterns in the world. In that sense, I see it as deeply related to computation and data science because it helps us scale up our vision and see how things today, like Dr. Gebru’s case, are connected to these patterns and cycles that we still haven’t been able to break out of today.”

The coauthors recognize that power plays a major role in determining what kind of behavior is considered ethical. This corresponds to the idea of privilege hazard, a term coined in the book Data Feminism, which D’Ignazio coauthored last year, to describe an inability to fully comprehend another person’s experience.

A long-term view seems to run counter to the traditional Silicon Valley dogma surrounding scale and growth, a point emphasized by Google Ethical AI team research scientist and sociologist Dr. Alex Hanna weeks before Gebru was fired. A paper Hanna coauthored with independent researcher Tina Park in October 2020 called scale thinking incompatible with addressing social inequality.

The Abuse and Misogynoir Playbook is the latest AI work to turn to history for inspiration. Your Computer Is On Fire, a collection of essays from MIT Press, and Kate Crawford’s Atlas of AI, released in March and April, respectively, examine the toll datacenter infrastructure and AI take on the environment and civil rights and reinforce colonial habits about the extraction of value from people and natural resources. Both books also investigate patterns and trends found in the history of computing.

Race After Technology author Ruha Benjamin, who coined the term “new Jim Code,” argues that an understanding of historical and social context is also necessary to safeguard engineers from being party to human rights abuses, like the IBM workers who assisted Nazis during World War II.

A new playbook

The coauthors end by calling for the creation of a new playbook and pose a challenge to the makers of artificial intelligence.

“We call on the AI ethics community to take responsibility for rooting out white supremacy and sexism in our community, as well as to eradicate their downstream effects in data products. Without this baseline in place, all other calls for AI ethics ring hollow and smack of DEI-tokenism. This work begins by recognizing and interrupting the tactics outlined in the playbook — along with the institutional apparatus — that works to disbelieve, dismiss, gaslight, discredit, silence, and erase the leadership of Black women.”

The second half of a panel discussion about the playbook in late March focused on hope and ways to build something better, because, as the coauthors say, it’s not enough to host events with the term “diversity” or “equity” in them. Once abusive patterns are recognized, old processes that led to mistreatment on the basis of gender or race must be replaced with new, liberatory practices.

The coauthors note that making technology with liberation in mind is part of the work D’Ignazio does as director of the Data + Feminism Lab at MIT, and what Turner and Wood do with the Space Enabled research group at MIT Media Lab. That group looks for ways to design complex systems that support justice and the United Nations Sustainable Development Goals.

“Our assumption is we have to show prototypes of liberatory ways of working so that people can understand those are real and then try to adopt those in place of the current processes that are in place,” Wood said. “We hope that our research labs are actually mini prototypes of the future in which we try to behave in a way that’s anticolonial and feminist and queer and colored and has lots of views from people from different backgrounds.”

D’Ignazio said change in tech — and specifically for the hyped, well-funded, and trendy field of AI — will require people considering a number of factors, including who they take money from and choose to work with. AI ethics researcher Luke Stark turned down $ 60,000 in funding from Google last month, and Rediet Abebe, who cofounded Black in AI with Gebru, has also pledged to reject funding from Google.

In other work at the intersection of AI and gender, the Alan Turing Institute’s Women in Data Science and AI project released a report last month that documents problems women in AI face in the United Kingdom. The report finds that women only hold about 1 in 5 jobs in data science and AI fields in the U.K. and calls for government officials to better track and verify the growth of women in data science and AI.

“Our research findings reveal extensive disparities in skills, status, pay, seniority, industry, job attrition, and education background, which call for effective policy responses if society is to reap the benefits of technological advances,” the report reads.

Members of Congress interested in algorithmic regulation are considering more stringent employee demographic data collection, among other legislative initiatives. Google and Facebook do not currently share diversity data specific to employees working within artificial intelligence.

The Abuse and Misogynoir Playbook is also the latest AI research from people of African descent to advocate taking a historical perspective and adopting anticolonial and antiracist practices.

In an open letter shortly after the death of George Floyd last year, a group of more than 150 Black machine learning and computing professionals outlined a set of actions to bring an end to the systemic racism that has led Black people to leave jobs in the computing field. A few weeks later, researchers from Google’s DeepMind called for reform of the AI industry based on anticolonial practices. More recently, a team of African AI researchers and data scientists have recommended implementing anticolonial data sharing practices as the datacenter industry in Africa continues growing at a rapid pace.