Essay · AI & Society
Originally circulated as a working paper. Read the original PDF ↗
In response to a national and international awakening on the issues of anti-Blackness and systemic discrimination, this piece serves as a resource for allies in the AI community wondering how they can more effectively engage with dismantling racist systems. It aims to elucidate where the AI community actively and passively contributes to anti-Blackness, and offers actionable items to reduce harm.
"How did you go bankrupt?!" Bill asked. "Two ways," Mike said. "Gradually and then suddenly."
This oft-referenced phrase aptly describes how substantial changes that take a long time to develop can appear to happen all at once. The extrajudicial killings of George Floyd, Breonna Taylor, Ahmaud Arbery, Tony McDade, and others, in combination with a poorly-managed pandemic disproportionately damaging the Black community, poured gas on a growing blaze — calling on us to truly address the extent of anti-Black systemic racism in the United States and abroad. Through this lens, many in the AI community are asking: "What can I do to combat systemic racial injustice?"
The aim of this work is to help community members better identify and understand the scale and scope of anti-Black bias within our AI community, and to illustrate concrete steps members can take to mitigate these issues and build a more just community.
To summarize the contributions: we first establish the necessity of recognizing the scale and scope of anti-Blackness and how it permeates all of our institutions. We then identify areas within academia where anti-Blackness is magnified or reinforced, and propose actions for faculty, graduate students, and conferences to take to minimize these deficiencies.
Students of optimization go through a stage where they see every problem as an optimization problem. Similarly, students of racism must go through a stage where they realize that anti-Black racism impacts every aspect of our society. This acknowledgement is not sufficient to dismantle systems of racism; however, it is a necessary first step, and it allows one to better appreciate the scope of the issues we are trying to solve.
Issues such as access to healthcare8, clean air4, quality education10, credit3, clean water19, housing22, public transit22, voter suppression2, fair wages7, and criminality5 all have roots in anti-Blackness.
This piece restricts its focus to the AI community and the ways in which we contribute to anti-Black racism. While the focus of recent protests has centered around anti-Black bias in the criminal justice system, the scope here expands to include more ways in which we propagate systemic harm: in our sources of funding, in who we allow to participate in our community, in what problems we address or exacerbate through our research and applications, and in what set of principles we allow to guide us.
A simplifying model helps examine the sources of systemic racism that permeate our communities — three overlapping categories:
Differences in wealth, income, access to computing resources, access to clean air or water, healthcare, and transportation, among other areas. Included in this list, though perhaps more abstract, is time — time spent dealing with issues of race is time not spent improving skills that are more greatly valued than "the ability to navigate racism."
Who we know and who knows us are tremendous factors in explaining the opportunities presented in our lives6. Large percentages of open roles in technology companies are filled via referrals; compensation structures in CS/AI are hidden and structured such that you have to know someone to know what is fair. Projects are rarely completed by one individual, and who you know informs what you work on, what opportunities you're aware of, and who can vouch for you.
75% of White people have entirely White social networks, while the average White person's social network is only 1% Black.
This remarkable racial stratification of social networks, combined with the prominent role social networks play in our successes, leads to further racial inequity.
Anything used to evaluate, punish, or reward individuals. The most prominent example today is policing: while Black and White people use drugs at similar rates, those arrested and convicted for drug offenses are disproportionately Black. Measures of aptitude such as the SAT or GRE have also shown racial bias — as have teacher evaluations, coding interviews, open-source contribution records, promotion cycles, and in-school suspension rates.
These three categories — physical resources, social resources, and measures — are not all-encompassing, and many overlap (credit, for instance, may be viewed as either a measure of quality or a physical resource). But they serve as a good starting point for dissecting systems of systemic bias.
The importance of feedback mechanisms cannot be understated when examining any system, just or unjust. Since academia serves as both the touchstone of the AI community and the developmental environment for new AI talent, biases within academia are propagated outward — and fed back into the next generation of academia, further entrenching them. As such, academia is not only obligated to halt practices that disproportionately impact the Black community, but also to repair the damage done through generations of neglect.
Our faculty set the tone for our community. If our community is to ever achieve equity, it will require buy-in and focused effort from our faculty — who decide which topics are highlighted in courses, who is admitted into our programs, and what criteria are required to attain a degree.
There is no objective way to determine research potential, and this is the primary goal of the admissions process. GPAs, letters of recommendation, publications, and GRE scores are major factors — yet GRE scores are racially biased17, and publications and letters of recommendation are heavily influenced by the physical resources available to an applicant and the researchers in their social network. For these reasons, standard measures of research output systemically disadvantage Black applicants, and feed back into the applicant pool in a way that amplifies existing bias.
The shortest answer is take risks. Right now, we screen for students who resemble previous successful students. Admitting students whose experiences and achievements look different may seem risky — but years of studying explore-exploit trade-offs should give us confidence to take that risk. Waiting for a flood of Black candidates whose applications look identical to those admitted for the last K years misunderstands the fundamentals of systemic racism.
If the goal of admissions is to select for research potential, the goal of mentorship is to develop it. A useful exercise: examine who you've written letters of recommendation for in the last five years. Does that list's diversity reflect the diversity you'd hope to see in the field? Unless we're very intentional, our mentorship list is likely even less diverse than our department. Ways to change this include actively reaching out to promising underrepresented students, participating in targeted summer REU programs, and publicizing transparent protocols for getting involved with labs. Sometimes the work of allies is to put in extra effort to make themselves as accessible as possible.
Current students should be your primary source of collaboration, yet who you work with at other institutions matters greatly. Collaborating with familiar faces has its benefits, but it also means you're less likely to encounter diverse people through your research — affecting not only whose work you're familiar with, but who is familiar with yours.
Our community has done a tremendous job of being driven by empirical performance while avoiding consequential discussion of the implications of the applications we focus on. This piece was written amid an intense international spotlight on anti-Black violence enacted and enabled through policing systems. AI has enabled more efficient and pervasive tracking, monitoring, criminalization, and repression of Black people — and law enforcement has historically deployed sophisticated surveillance campaigns aimed at dismantling civil rights movements. Diversifying our field will not remedy the harms our systems have caused, but failing to address these harms may further drive marginalized groups from our community — which can reasonably be seen as complicity in the harms our systems produce.
Work on fairness and ethical AI exists, but hasn't received the attention given to more popular, controversial work: the most-cited example of ethical AI research found here has ~1,300 citations1, while several facial-recognition papers from the same period have ≥3,00018, 21, 20. Centering research on the societal implications of AI is necessary to ensure AI has a positive impact on society.
Ten computer science programs are responsible for producing more than 50 percent of all CS faculty across the United States12. Sourcing graduate students primarily from institutions with poor Black student retention compounds under-representation by spreading the problem to other schools. It's tempting to assume the representation problem stems from a lack of suitable candidates — but Black students make up ≥4% of CS undergraduate degrees, yet ≤1% of PhDs, pointing to a distinct drop-off that undergraduate diversity alone doesn't explain.
One contributing factor may be which schools serve as feeder programs. Despite having 1/10th the students, the 101 HBCUs produce more Black CS bachelor's degrees than the 115 R1 institutions in the U.S. To conduct a quick self-audit, consider the HBCUs below — which graduate the largest numbers of Black CS students9, 16 — and how connected they are to your department:
Applications of AI have inflicted real harm on Black people. How you engage with companies that profit from anti-Black policies matters — who you take funding from, who you allow to recruit on campus, where your students end up working. These are offered as provocative questions, not prescriptive answers. But it's important for organizations to have these conversations and decide their own ethical lines, if for no other reason than to mitigate the risk of scandal tied to money accepted from nefarious sources.
As we climb the ladder of academia, institutions become more risk-averse. There might be willingness to take a chance on an undergraduate — that willingness drops sharply for graduate students, postdocs, and especially faculty. If we want to truly move the needle on anti-Blackness, the hiring and retention of Black professors must be addressed: historically, Black faculty have done a much better job of establishing relationships with Black students than other faculty. Fixing this isn't just about who is visible at the top — it's about who is positioned to succeed at every level of the pipeline.
Systems requiring financial "sacrifice" for professional advancement are extremely effective at removing high-achieving Black people from the leadership pipeline. Given the 10x difference in median wealth between Black and White families, a drop in income during graduate school disproportionately impacts Black students and potential students — meaning many qualified, interested Black students simply cannot afford to attend. Some programs have developed corporate partnerships providing supplemental income, but these relationships are often opaque to students without close insider connections — so even meaningful improvements often go unseen by the students who'd benefit most.
Are the Black students on your campus adequately supported? This is best answered by talking to current and former Black students directly. Many students make it through graduate programs with traumatic experiences, and without Black faculty present, may not trust anyone enough to be candid about problems in their environment. When you do hear complaints, recognize that sharing them is often a risk for the student — listen without becoming defensive.
What resources have you committed to combating anti-Black bias? Instead of proclaiming beliefs and convictions, demonstrate them through investment and action. Meaningful change in racial equity is achievable, but requires tangible effort from people other than those who are discriminated against.
Substantial research shows diverse teams achieve better performance13 — but this piece rejects that predatory framing, in which the worth of underrepresented people is tied to their value-add to in-group members. The argument here is for combating anti-Blackness through the lens of justice. All members of the community should be invested in a more inclusive, less discriminatory environment — full stop.
Black graduate students and faculty shoulder an extraordinary amount of the burden of "diversifying" university campuses. One of the most impactful actions an ally can take is to help carry that work — which not only helps diversify the campus, but frees up Black peers to spend more time on their own research. Many of the necessary tasks are relatively simple: show up early, ask what you can do to support, volunteer for a non-leadership role. "Hey, I'd like to help secure food for all the NSBE meetings this semester." Attention on anti-racism might be fleeting; show up consistently, and be willing to be led.
Lack of social integration is a significant factor in disparate outcomes Black graduate students face in academia. Lunches, dinners, happy hours, and game nights bear no direct relationship to technical merit, but significantly contribute to a sense of safety and belonging. Study groups, research collaborations, and workshop organizers are often selected through a mix of social and academic ties. Put in the extra effort to know your Black colleagues' skills and interests, so that when relevant opportunities arise, they're given fair consideration.
Who we invest our time in matters — today's undergraduates are tomorrow's graduate students. One of the biggest levers for increasing diversity in graduate school is increasing diversity in undergraduate research, both by encouraging more underrepresented students to participate and by recruiting more graduate mentors.
It's also necessary to focus on funding for undergraduate researchers. Unpaid internships have a long history of reinforcing racial inequality in the U.S. Consider a department's total mentorship capacity C = U + P, where U is unpaid and P is paid research slots. If unpaid slots are effectively reserved for students of substantial financial means, this arrangement alone — ignoring every other source of systemic bias — produces a system that significantly disadvantages Black students, who form the primary pool for future graduate students. Calculating the true numbers for P and U at your institution is a straightforward exercise that can illustrate just how skewed the current system is.
Racism exists — in our neighborhoods, our departments, our labs. Conversations that question its existence are either intellectually lazy or conducted with ill intent. Just as a budget meeting filled with debate about whether currency exists is unproductive, so is a discussion of anti-Blackness in which we refuse to name it, or indulge endless debate over whether it exists at all. We cannot continue to offer misinformed individuals or bad actors as much space as those working through active solutions.
This piece discusses several areas within the AI community where systems perpetuate anti-Blackness. It falls woefully short of being comprehensive, and intentionally does not address systems of racism within industry, conferences, or pipelines of capital. It also does not address issues of sexism within our community, whose effects compound with racism to disproportionately impact Black women.
The ultimate goal of this work is to help create a more equitable AI community, whose broader impact includes reducing anti-Blackness in society. While not simple, this is a tractable task — and one this community is called upon to help make happen.