The Misunderstanding We Must Correct
For years, the public conversation about inequality in institutions was organized under the acronym DEI — Diversity, Equity, and Inclusion. That framing, however well-intentioned, was always insufficient. It reduced a vast and complex set of structural inequities to a three-letter label that different people defined differently, funded differently, and valued differently.
Organizations responded to DEI pressures from within and without. Many installed Black women to lead programs as complex as the entire human resources structure itself. DEI lived under the HR banner but did not carry the power or tools that HR commanded. In most organizations, the person appointed to lead DEI stood alone. Once installed, DEI officers were overwhelmed by the volume and weight of issues arriving at their doorstep, with very little institutional authority to address any of them.
Placing one person from a marginalized group in charge of solving the systemic inequities that affect that group is not a solution. It is a deflection. It confuses representation with remedy. And when those programs inevitably struggled under impossible conditions, under-resourced and over-scrutinized, critics pointed to the failure as proof that DEI itself was the problem. They were wrong. The conditions were designed to produce that outcome.
DEI was never simply about advancing, hiring, or promoting any one group. It was meant to create equity and fairness for all. When we turn words into weapons, we invite exactly the distortion that followed. Loud voices carried through break rooms, meetings, and corridors suggested DEI was destroying the morale and productivity of organizations. Those voices drowned out the evidence. DEI programs historically benefited white women more than any other group. The people most threatened by equity were not losing something. They were being asked to share something they had never had to share before.
In 2025, DEI came under coordinated attack from every level of American institutional life. Businesses began to eradicate DEI programs in large numbers. The glass ceiling that had been beginning to crack was reinforced, not with glass, but with concrete and steel. The advances made over decades were not simply paused. They were deliberately reversed. The conditions were set to ensure DEI would fail, and then that failure was used as justification for the reversal.
This matters for AI because the same dynamic is now playing out in the technology sector. When AI systems produce biased outcomes, the instinct is to appoint a Chief Responsible AI Officer, publish a fairness framework, and move on. What is rarely done is the harder work: examining the data the systems were trained on, the teams that built them, and the institutional assumptions embedded in every design decision made along the way. Equity in AI, like equity in organizations, will not arrive through symbolism. It requires structural change.
How Bias Enters AI Systems
Bias in AI does not arrive from malicious intent. It arrives from data. AI systems learn from the past. When the past is characterized by exclusion, discrimination, and unequal access, an AI system trained on that past will reproduce those patterns with extraordinary efficiency. The algorithm does not know it is being unfair. It knows only what it was taught.
The problem is compounded by who builds these systems. The technology industry remains among the least demographically diverse sectors in the American economy. When the teams designing AI tools do not reflect the populations those tools will affect, entire categories of bias remain invisible until they cause harm.
The Language of Technology Has Always Carried Bias
Long before AI became a household word, the language of computing embedded hierarchy and exclusion into its foundation. These were not accidents. They were reflections of who was in the room when the terminology was chosen, and who was not.
| Term | The Problem | The Shift |
|---|---|---|
| Master / Slave | Used to describe primary and secondary hardware components | Replaced with Primary / Replica or Leader / Follower in most modern standards |
| Whitelist / Blacklist | White as approved, Black as blocked, embedded in network security language | Replaced with Allowlist / Denylist in updated style guides |
| Grandfathered | Originated from post-Civil War laws designed to disenfranchise Black voters | Being replaced with Legacy status or Exempt |
| Manpower | Assumes a male default in workforce language | Replaced with Workforce, Staffing, or Human Resources |
| Sanity check | Uses mental health as a metaphor for correctness | Replaced with Confidence check or Validation |
These language choices matter because they shape what people perceive as normal and what they perceive as other. When the default is white, male, able-bodied, and English-speaking, everything outside that default is treated as an edge case. In AI, edge cases are frequently the people who need the system most.
The Data: What Research Tells Us
Algorithmic bias is not a theory. It is documented, measured, and growing. The following statistics come from peer-reviewed research, federal government studies, and major academic institutions. They represent only a fraction of what has been found.
of the time, AI resume screening tools favored white-associated names over Black-associated names, even when qualifications were identical.
Source: University of Washington, October 2024of Fortune 500 companies now use AI in their hiring processes, yet most applicants have no way of knowing they were screened by an algorithm.
Source: Brookings Institution, 2025In multiple hiring bias tests, AI resume screening tools showed a near-zero selection rate for Black male names, regardless of qualifications.
Source: All About AI Bias Report, 2026African American and Asian faces were up to 100 times more likely to be misidentified than white faces by federal government facial recognition studies.
Source: U.S. Federal Government / Amnesty International CanadaFacial recognition software misidentified the gender of dark-skinned women 35 percent of the time in landmark MIT Media Lab research.
Source: MIT Media Lab, Joy Buolamwini, 2018Facial detection technology failed to detect faces with the darkest skin tones 24.34 percent of the time, compared to 0.28 percent for the lightest skin tones.
Source: U.S. Department of Homeland Security / Biometric Update, July 2025A 2025 study published in PNAS Nexus analyzed approximately 361,000 fictitious resumes submitted to five leading AI models including GPT-3.5, GPT-4o, Gemini, and others. Black male applicants were systematically disadvantaged across every model tested, even when their qualifications were identical to those of favored candidates. The researchers concluded that the transition from human to AI decision-making will redistribute employment opportunities across social groups in ways that are complex, largely invisible, and currently unregulated.
Bias in Practice
The following examples illustrate how algorithmic bias moves from data and design into real decisions that affect real people. These are not hypothetical scenarios. They are documented cases and patterns observed across industries and institutions.
The Resume Experiment
A documented pattern, replicated in formal research and in personal accounts: an applicant submits credentials to a major employer and receives no response. A colleague submits the identical credentials but changes the name and the institution attended to reflect a different racial identity. The second applicant receives an interview invitation. The qualifications were the same. The algorithm was not evaluating qualifications. It was evaluating patterns it had learned to associate with success, patterns built on decades of biased hiring decisions made by humans before the algorithm existed.
Facial Recognition and Wrongful Arrest
Facial recognition technology has led to the wrongful arrest of multiple Black men in the United States, including documented cases in Michigan and New Jersey, where facial recognition matches sent police to arrest individuals who were later proven through irrefutable evidence to have been miles away at the time of the alleged crime. In each case, the algorithm had misidentified a Black face. These were not errors of human prejudice in the traditional sense. They were the automated output of systems trained on data that underrepresented darker skin tones.
Soap Dispensers and Self-Driving Cars
Electronic soap dispensers in airports and office buildings that failed to recognize dark skin. Self-driving vehicle systems that struggled to detect pedestrians with darker complexions. iPhones that could not distinguish between Asian women. These are not isolated engineering failures. They are the product of development teams that did not include the populations their products would serve. When the people building the technology do not reflect the full range of humanity, the technology will reflect the narrow range of humanity that built it.
Grant Algorithms and Institutional Exclusion
When AI and algorithmic systems are used to screen grant applications, loan requests, or institutional funding decisions, the same pattern emerges. Systems trained on historical approval data will learn to favor the characteristics of historically approved applicants. Institutions like North Carolina Central University, a historically Black university, may find themselves systematically disadvantaged by algorithms that were never designed to recognize their value or their mission because institutions like theirs were never well represented in the data the algorithm learned from.
Healthcare Algorithms and Unequal Care
In healthcare, algorithmic systems used to predict which patients need additional care have been found to systematically underestimate the needs of Black patients. One widely studied algorithm used healthcare costs as a proxy for healthcare needs. Because Black patients had historically received less care and therefore generated lower costs, the algorithm rated them as having lower needs than white patients with the same health conditions. The algorithm was accurate by its own internal logic. Its internal logic was built on inequity.
A Pattern We Have Seen Before
Social media was deployed at extraordinary scale with minimal consideration for how it would affect the people using it. Platforms optimized for engagement without asking what engagement at that scale would do to human attention, to democratic discourse, to the mental health of adolescents, or to the spread of harmful content. By the time those questions were being asked seriously, billions of people were already inside systems that had been designed without them in mind.
We are in a similar position with AI. The deployment has outpaced the governance. The scale has outpaced the scrutiny. And the populations most likely to be harmed by biased AI systems are the same populations that were most harmed by unregulated social media: people of color, women, young people, and those without the resources or platform to document and challenge the harms being done to them.
The Regulation Gap
As of 2026, only New York City requires annual bias audits for automated employment decision tools. California finalized regulations on AI hiring tools in October 2025. The Colorado AI Act, addressing algorithmic discrimination in hiring, does not take effect until June 2026. At the federal level, EEOC guidance on AI discrimination published in 2022 was removed when the new administration took office in January 2025. The laws against discrimination have not changed. The guidance on how they apply to AI has been withdrawn. The gap between the scale of AI deployment and the regulatory frameworks governing that deployment grows wider every day.
What We Are Calling For
The goal is not to stop AI. The goal is to demand that AI be built and deployed in ways that do not replicate and amplify the exclusions that have characterized human institutions for centuries. That requires specific, enforceable action.
Mandatory Bias Audits Before Deployment
Any AI system used to make or inform decisions about people in hiring, lending, healthcare, education, housing, or criminal justice must be subject to independent bias audits before deployment and on a regular basis thereafter. Those audits must be conducted by parties with no financial interest in the outcome and must be made publicly available.
Transparency for Affected Individuals
People who are subject to algorithmic decision-making have a right to know it. They have a right to know what factors were considered, what data was used, and what recourse is available to them if they believe a decision was made in error or in violation of their rights. Proprietary systems are not exempt from accountability.
Diverse and Representative Development Teams
The homogeneity of AI development teams is not a coincidence and it is not a pipeline problem that will resolve itself. It is the result of systemic exclusion in technology education and hiring. Addressing it requires deliberate investment in inclusive hiring, in historically Black colleges and universities, in programs that bring underrepresented communities into the development process as full participants rather than as afterthoughts or end users.
Intersectional Impact Standards
Bias audits and fairness assessments must account for intersectionality. A system may show acceptable outcomes for Black applicants and acceptable outcomes for women and still systematically disadvantage Black women. Regulatory frameworks must require analysis at the intersection of race, gender, disability, age, and other characteristics, not just single axes of identity.
Reinstatement of Federal Guidance on AI Discrimination
The withdrawal of EEOC guidance on AI and employment discrimination in January 2025 left workers without a federal framework for understanding how existing anti-discrimination law applies to algorithmic hiring decisions. That guidance must be reinstated and strengthened to address the current scale of AI deployment in employment contexts.
AI is not broken. It is working exactly as it was designed. Changing the outcomes requires changing the design: the data, the teams, the assumptions, and the accountability structures that govern what these systems are allowed to do to people.