AI Services Companies

AI Bias Statistics 2026

Published November 18, 2026

Our analysts gathered results from eight studies and surveys published from 2018 to 2025, covering more than three million AI resume comparisons, 189 face recognition algorithms, the records of 49,618 hospital patients, 6,000 test mortgage applications, and the views of 5,410 US adults and 1,013 AI experts, to show how often AI systems treat groups of people differently as of Q4 2026.

Key Findings

  • In a 2024 University of Washington test of three AI models ranking resumes, the models preferred white-associated names 85% of the time and Black-associated names 9% of the time.2
  • Commercial gender classifiers tested in the 2018 Gender Shades study misclassified darker-skinned women up to 34.7% of the time, compared with at most 0.8% for lighter-skinned men.4
  • In a 2024 Lehigh University test, GPT-4 Turbo would have required Black mortgage applicants to have credit scores about 120 points higher than white applicants to get the same approval rate.7
  • 47% of organizations in a 2025 Accenture and Stanford survey reported a model bias incident in the past two years, and in a 2024 McKinsey survey only 26% said they were actively working to mitigate equity and fairness risks.8
  • 55% of US adults and 55% of AI experts said in 2024 that they are highly concerned about bias in decisions made by AI.9

AI bias happens when an AI system gives worse or different results to people because of their race, gender, age, or another trait, usually because the data it learned from reflected unequal treatment in the past. A resume screener trained on past hiring decisions can learn to favor the kinds of names that were hired before, and a face recognition system trained mostly on lighter-skinned faces can make more mistakes on darker-skinned ones. As employers, hospitals, lenders, and police departments rely on these systems for decisions about real people, researchers have started to measure how large the differences are.2,4,6 The results vary by system and by group, but several of them are large.

Our analysts read the published results of academic audits of AI resume screening, face analysis, health care risk scoring, and mortgage underwriting. We added the National Institute of Standards and Technology (NIST) test of face recognition algorithms, a Cornell University study of how employers comply with New York City's bias audit law, the responsible AI chapter of Stanford's 2025 AI Index, and Pew Research Center surveys of US adults and AI experts. All sources were opened on November 18, 2026.

In the following sections, we compare bias in AI hiring tools, error rates in face analysis and face recognition, bias in health care and lending algorithms, how organizations handle bias risk, how Americans and AI experts view AI bias, and how research on AI fairness has grown since 2019.

AI Resume Screeners Preferred White-Associated Names 85% of the Time

Many employers now use AI to sort applicants before a recruiter reads a single resume. To test whether these tools treat candidates differently, researchers keep the resume the same and change only the name at the top, which signals a likely race and gender. We collected the results of the largest such test of AI language models and of a study of New York City's law requiring bias audits of hiring software.

AI Hiring Bias and Bias Audit Statistics
MeasureValueSource
Share of comparisons in which AI models preferred white-associated names85% (Black-associated names: 9%)University of Washington, 2024
Share of comparisons in which AI models preferred male-associated names52% (female-associated names: 11%)University of Washington, 2024
Share of comparisons in which AI models preferred Black female names over Black male names67% (Black male names: 15%)University of Washington, 2024
Comparisons in which Black male names were preferred over white male namesNoneUniversity of Washington, 2024
Scale of the University of Washington testMore than 550 resumes, 120 first names, more than 500 job listings in nine occupations, and more than three million comparisonsUniversity of Washington, 2024
Employers checked for compliance with New York City Local Law 144391Cornell University and partners, 2024
Employers that posted a bias audit report18 of 391 (4.6%)Cornell University and partners, 2024
Employers that posted a transparency notice to job seekers13 of 391 (3.3%)Cornell University and partners, 2024
Employers with open New York City jobs that posted an audit report14 of 267 (5.2%)Cornell University and partners, 2024

Source: AI Services Companies analysis of University of Washington and Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability data, conducted November 2026. Data: ai-resume-screening-name-preference-2024.csv

Grouped horizontal bar chart of how often AI resume screeners preferred each name group in 2024: white-associated names 85% versus Black-associated names 9%, male-associated names 52% versus female-associated names 11%, and Black female names 67% versus Black male names 15%

Source: AI Services Companies analysis of University of Washington data, conducted November 2026.

Key Findings

  • The three models tested came from Mistral AI, Salesforce, and Contextual AI. The researchers presented the study in October 2024 at the AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society.2
  • The bias was not the sum of race and gender effects. The smallest gap was between white female and white male names, while Black male names were never preferred over white male names.2
  • New York City has required yearly independent bias audits of automated hiring tools since July 2023. Because employers decide for themselves whether their tools are covered, the Cornell researchers could not tell how many of the 373 employers without a posted audit were breaking the law, and nearly all posted audits reported impact ratios above 0.8, the usual rule of thumb in employment discrimination cases.3

Face Analysis Error Rates Reached 34.7% for Darker-Skinned Women

Face analysis software is used to unlock phones, check passports, and search for suspects in police photo databases. Two studies set the benchmark figures still cited for how its accuracy differs across groups: the 2018 Gender Shades study, which tested commercial tools that guess a person's gender from a photo, and NIST's 2019 test of face recognition algorithms, which match one face to another.

Face Analysis and Face Recognition Bias Statistics
MeasureValueSource
Highest gender classification error rate, darker-skinned women34.7% (IBM)Gender Shades, 2018
Lowest error rate for darker-skinned women among the three classifiers20.8% (Microsoft)Gender Shades, 2018
Highest gender classification error rate, lighter-skinned men0.8% (Face++)Gender Shades, 2018
Overall error rates of the three classifiers6.3% (Microsoft), 10.0% (Face++), and 12.1% (IBM)Gender Shades, 2018
Faces in the Gender Shades benchmark1,270 parliamentarians from three African and three European countriesGender Shades, 2018
Face recognition algorithms tested by NIST189 algorithms from 99 developersNIST, 2019
Photos used in the NIST test18.27 million images of 8.49 million peopleNIST, 2019
Typical difference in false positive rates between demographic groups10 to 100 times, depending on the algorithmNIST, 2019
Group with the highest false positive rates in one-to-one matchingAmerican IndianNIST, 2019
Group with higher false positive rates in one-to-many searches of 1.6 million FBI mugshotsAfrican American womenNIST, 2019

Source: AI Services Companies analysis of Gender Shades and NIST data, conducted November 2026. Data: gender-classification-error-rates-2018.csv

Heat map of gender classification error rates in 2018: Microsoft 20.8% for darker-skinned females, 6.0% for darker-skinned males, 1.7% for lighter-skinned females, and 0.0% for lighter-skinned males; Face++ 34.5%, 0.7%, 9.8%, and 0.8%; IBM 34.7%, 12.0%, 7.1%, and 0.3%

Source: AI Services Companies analysis of Gender Shades data, conducted November 2026.

Key Findings

  • All three classifiers in the Gender Shades study made the most errors on darker-skinned women. IBM's error rate for that group, 34.7%, was more than 100 times its 0.3% rate for lighter-skinned men.4
  • In NIST's one-to-one tests, US-developed algorithms had similarly high false positive rates for Asian, African American, and Native faces, while some algorithms developed in Asian countries showed no such large gap between Asian and white faces.5
  • A false positive in a one-to-many search puts the wrong person on a list of possible matches, which NIST says could lead to false accusations.5

Correcting a Hospital Algorithm Raised Black Patients' Share of Extra Care From 18% to 47%

Hospitals use algorithms to decide which patients get extra care, and lenders are testing AI models to help underwrite loans. In both cases the model can learn bias from the numbers it is given. A hospital algorithm that predicts health care spending, for example, will rank Black patients as healthier than they are if, as the researchers found, less money has been spent on their care in the past. We collected the results of a study of one widely used hospital algorithm and a test of AI models on mortgage applications.

AI Bias Statistics in Health Care and Lending
MeasureValueSource
Black patients as a share of those automatically enrolled in extra care, original algorithm18%UC Berkeley and Chicago Booth, 2019
Black patients as a share of those automatically enrolled, after correcting for health differences47%UC Berkeley and Chicago Booth, 2019
Patients in the hospital study43,539 white patients and 6,079 Black patientsUC Berkeley and Chicago Booth, 2019
People whose care is influenced by this category of algorithmOver a hundred million AmericansUC Berkeley and Chicago Booth, 2019
Extra credit score points Black applicants needed for the same approval rate (GPT-4 Turbo)About 120 pointsLehigh University, 2024
Extra credit score points Black applicants needed for the same interest rate (GPT-4 Turbo)About 30 pointsLehigh University, 2024
Mortgage applications used in the Lehigh test6,000 test applications built from 1,000 real 2022 applicationsLehigh University, 2024
Model with the most discrimination in the Lehigh testGPT-3.5 Turbo (GPT-4 from 2023 showed virtually none)Lehigh University, 2024

Source: AI Services Companies analysis of UC Berkeley and Lehigh University data, conducted November 2026. Data: health-risk-algorithm-black-patients-enrolled-2019.csv

Column chart of Black patients as a share of patients automatically enrolled in a hospital care program: 18% under the original cost-based algorithm and 47% after correcting for the health gap between Black and white patients

Source: AI Services Companies analysis of UC Berkeley data, conducted November 2026.

Key Findings

  • The hospital algorithm was biased because it used health care costs to stand in for health. Black patients at the same level of health generated lower costs, so at the same risk score they were much sicker than white patients.6
  • Retraining the algorithm to predict avoidable costs or the number of chronic conditions needing treatment in a year significantly reduced the racial bias, the researchers reported.6
  • The Lehigh researchers built their test applications from the 2022 Home Mortgage Disclosure Act data set and changed only race and credit score. They also tested GPT-4, Claude 3 Sonnet and Opus, and Llama 3 models.7

47% of Large Organizations Reported a Model Bias Incident

Companies that build or use AI are the ones expected to catch bias before it reaches customers or job applicants. Two surveys summarized in Stanford's 2025 AI Index show how often organizations run into bias problems and how many are working to prevent them.

Organizational AI Bias and Fairness Statistics
MeasureValueSource
Organizations reporting a model bias incident in the past two years47%Accenture and Stanford survey, 2025
Organizations reporting an adversarial attack incident56%Accenture and Stanford survey, 2025
Organizations reporting a privacy violation incident55%Accenture and Stanford survey, 2025
Organizations reporting unintended decision making by AI51%Accenture and Stanford survey, 2025
Organizations in the Accenture and Stanford survey1,500 with revenues of at least $500 million, in 20 countries and 19 industriesAccenture and Stanford survey, 2025
Organizations that consider equity and fairness a relevant AI risk34%McKinsey survey, 2024
Organizations actively working to mitigate equity and fairness risks26%McKinsey survey, 2024
Organizations that consider cybersecurity a relevant AI risk66% (53% actively mitigating)McKinsey survey, 2024

Source: AI Services Companies analysis of Stanford AI Index 2025, Chapter 3: Responsible AI data, conducted November 2026. Data: ai-risks-relevant-vs-mitigated-2024.csv

Dumbbell chart of AI risks organizations considered relevant versus actively mitigated in 2024, including cybersecurity 66% and 53%, regulatory compliance 63% and 50%, explainability 40% and 31%, and equity and fairness 34% and 26%

Source: AI Services Companies analysis of McKinsey data published in the Stanford AI Index 2025, conducted November 2026.

Key Findings

  • Equity and fairness ranked eighth of the 13 AI risks in the McKinsey survey, below cybersecurity, regulatory compliance, privacy, inaccuracy, intellectual property, reputation, and explainability.8
  • In every risk category, fewer organizations were working to mitigate the risk than said it was relevant to them.8
  • Model bias was the fourth most common incident type in the Accenture and Stanford survey, after adversarial attacks, privacy violations, and unintended decision making, and ahead of performance failures at 46%.8

55% of Americans and AI Experts Are Highly Concerned About AI Bias

Pew Research Center asked the same questions of a national sample of US adults and of AI experts who wrote for or presented at AI conferences. On most questions the two groups disagree, with experts far more optimistic. Bias in decisions made by AI is one of the few concerns the two groups share in equal measure.

Public and Expert Views on AI Bias
MeasureUS adultsAI experts
Highly concerned about bias in decisions made by AI55%55%
Highly concerned about people getting inaccurate information from AI66%70%
Want more control over how AI is used in their lives55%57%
Little or no confidence in US companies to develop and use AI responsibly59%55%
Say people who design AI take men's perspectives into account at least somewhat wellNot reported75%
Say people who design AI take women's perspectives into account at least somewhat wellNot reported44%
Say the perspectives of White adults are well represented in AI designFour in tenAbout three-quarters
Survey sample and dates5,410 adults, August 12 to 18, 20241,013 experts, August 14 to October 31, 2024

Source: AI Services Companies analysis of Pew Research Center data, conducted November 2026. Data: ai-bias-concern-public-vs-experts-2024.csv

Grouped column chart of US adults and AI experts in 2024: highly concerned about bias in AI decisions 55% and 55%, highly concerned about inaccurate information 66% and 70%, want more control over AI 55% and 57%, and little or no confidence in companies to use AI responsibly 59% and 55%

Source: AI Services Companies analysis of Pew Research Center data, conducted November 2026.

Key Findings

  • Experts were far more likely than the public to expect AI to have a positive effect on the United States over the next 20 years, 56% compared with 17%, yet both groups were equally concerned about bias.9
  • Among the public, a quarter or fewer said the perspectives of Asian, Black, or Hispanic adults are well represented in AI design, and about four in ten or more were unsure.9
  • Pew did not receive enough responses from Hispanic or Black AI experts to report expert views by race and ethnicity.9

AI Fairness and Bias Research Doubled in 2024

Researchers are publishing far more work on measuring and reducing AI bias, and newer studies test the large language and image models behind today's chatbots. We tracked the number of fairness and bias papers accepted at six leading AI conferences, as counted by Stanford's AI Index, along with the main findings it highlighted on bias in recent models.

AI Fairness and Bias Research Statistics
MeasureValueSource
Fairness and bias papers accepted at six leading AI conferences, 201957Stanford AI Index, 2025
Fairness and bias papers accepted, 2023212Stanford AI Index, 2025
Fairness and bias papers accepted, 2024408Stanford AI Index, 2025
All responsible AI papers accepted at the same conferences992 in 2023 and 1,278 in 2024 (up 28.8%)Stanford AI Index, 2025
Language models tested for implicit bias across 21 stereotype categories8, including GPT-4 and Claude 3 SonnetStanford AI Index, 2025, citing Bai et al., 2024
Increase in the likelihood of Black and Latino men being classified as criminals as training data grew from 400 million to 2 billion imagesUp to 69%Stanford AI Index, 2025, citing Birhane et al., 2024
Vision-language models evaluated in that study14Stanford AI Index, 2025, citing Birhane et al., 2024

Source: AI Services Companies analysis of Stanford AI Index 2025, Chapter 3: Responsible AI data, conducted November 2026. Data: ai-fairness-bias-papers-2019-2024.csv

Line chart of AI fairness and bias papers accepted at six leading AI conferences: 57 in 2019, 98 in 2020, 150 in 2021, 169 in 2022, 212 in 2023, and 408 in 2024

Source: AI Services Companies analysis of Stanford AI Index data, conducted November 2026.

Key Findings

  • Fairness and bias papers grew more than sevenfold from 2019 to 2024, from 57 to 408, and the 2024 count was nearly twice the 2023 count.8
  • Models trained to avoid explicit bias, including GPT-4 and Claude 3 Sonnet, still associated negative terms with Black people more often, linked women with the humanities more than with science fields, and favored men for leadership roles, the AI Index reported.8
  • The same research found that implicit bias increased as language models grew larger, even as bias scores on standard benchmarks improved.8

What AI Bias Statistics Show Going Into 2027

The measured gaps are often large. AI resume screeners preferred white-associated names nine times as often as Black-associated names, and a face analysis tool erred on darker-skinned women more than 100 times as often as on lighter-skinned men.1,2,4 Several studies also show that the gaps can shrink: the corrected hospital algorithm more than doubled Black patients' share of extra care, and in the Lehigh test GPT-4 from 2023 showed virtually no lending discrimination.

Most of these figures come from researchers testing systems from the outside, because few companies publish their own bias results. Even New York City's audit law, the first of its kind, produced posted audits from fewer than 5% of the employers checked.

So if you cite an AI bias statistic, name the system or model tested, the groups compared, the year, and the measure used, because a 2018 error rate for a gender classifier and a 2024 preference rate for a resume screener describe different systems and different kinds of harm.

Data Collection and Methodology

We collected figures from published studies, official test reports, and surveys, opened on November 18, 2026. Hiring figures come from the University of Washington's October 2024 summary of its resume screening study and from the study by Cornell University, the Data and Society Research Institute, and Consumer Reports of New York City Local Law 144, in which 155 student investigators checked 391 employers from October 24 to November 9, 2023. Face analysis figures come from the main results table of the Gender Shades paper by Joy Buolamwini and Timnit Gebru (2018) and from NIST's December 2019 summary of its Face Recognition Vendor Test, Part 3: Demographic Effects (NISTIR 8280). Health care figures come from UC Berkeley's summary of the study by Obermeyer and colleagues published in Science in October 2019, and lending figures from Lehigh University's August 2024 summary of its mortgage underwriting experiment.

Organizational survey figures, the count of fairness and bias papers, and the findings on implicit bias and image model bias come from Chapter 3 of Stanford's 2025 AI Index, which reports the McKinsey and Accenture and Stanford surveys and summarizes the Bai et al. and Birhane et al. studies. Paper counts for 2019 to 2024 were read from the labeled totals in the AI Index chart of fairness and bias papers at AAAI, AIES, FAccT, ICLR, ICML, and NeurIPS. Public and expert opinion figures come from Pew Research Center's April 2025 report on surveys conducted in 2024. Shares such as 18 of 391 employers (4.6%) and the 373 employers without a posted audit are our own arithmetic from the published counts. No figures in this article are estimates.

Sources (9)
  1. AI Services Companies analysis of AI bias studies, test reports, and surveys from the University of Washington, Cornell University, Gender Shades, NIST, UC Berkeley, Lehigh University, Stanford AI Index, and Pew Research Center, 2018 to 2025, conducted November 2026.
  2. University of Washington, AI tools show biases in ranking job applicants' names according to perceived race and gender (October 31, 2024)
  3. Lucas Wright and colleagues, Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability (FAccT 2024)
  4. Joy Buolamwini and Timnit Gebru, Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification (2018)
  5. National Institute of Standards and Technology, NIST Study Evaluates Effects of Race, Age, Sex on Face Recognition Software (December 19, 2019)
  6. UC Berkeley, Widely used health care prediction algorithm biased against Black people (October 24, 2019)
  7. Lehigh University, AI Exhibits Racial Bias in Mortgage Underwriting Decisions (August 20, 2024)
  8. Stanford Institute for Human-Centered AI, AI Index Report 2025, Chapter 3: Responsible AI
  9. Pew Research Center, How the U.S. Public and AI Experts View Artificial Intelligence (April 3, 2025)