A study published at a leading machine learning conference in Seoul in July 2026 found that AI models do not just inherit human biases: they generate new ones. Researchers at Princeton and the University of Chicago ran ChatGPT, Claude, and Gemini through a simulated hiring task and watched each model quickly sort candidates by group membership instead of individual fit. On a segregation scale where the maximum score is 2, human participants averaged 0.84. The AI models scored roughly 65% higher. OpenAI's most advanced reasoning model scored 1.83, close to the ceiling.
The models were not told anything biased. They were simply told to make good hires and given feedback on results. From that feedback, they drew sweeping conclusions. One early failure by a candidate from a particular group was enough to steer the model away from that group for an entire job category. The tendency to generalize quickly from small samples is, in the researchers' own words, central to how these systems are built. The same quality that makes AI fast at solving problems makes it fast at stereotyping people.
Telling the models to be fair changed almost nothing. What did work was giving the models explicit incentives for diverse outcomes. When researchers offered a bonus for diverse hiring in the experiment, bias dropped sharply. The practical lesson is clear: fairness instructions are not enough. The goals built into an AI system determine its behavior more than its stated values do.
A second finding from the same study is equally important. When models were given more specific personal information about individual candidates, they were less likely to fall back on group stereotypes. Relevant details, such as education and age, reduced bias. Irrelevant details, such as hair color, did not. This suggests that vague or sparse candidate data is actually riskier than detailed profiles, because thin information gives the model less to work with and more reason to rely on group patterns.
This research lands in the middle of a sharp legal shift. In the US, a class action lawsuit against Workday, whose AI screening software is used across hundreds of large companies, is proceeding in federal court. The core allegation is that the software systematically filtered out older and disabled applicants. Courts have ruled that employers who use third-party AI tools to screen candidates share liability for discriminatory outcomes, even if the vendor built the system. One company's platform is estimated to have processed over 1 billion applications during the period covered by the lawsuit.
In Europe, the compliance clock for hiring AI expires on 2 August 2026. Any tool that screens resumes, ranks candidates, or evaluates interviews is now classified as high-risk under EU law. That classification requires documented bias testing, human review at every decision point, and written disclosure to every candidate about how AI was used. Fines for violations go up to 15 million euros or 3% of global turnover. The law also applies to non-European companies if the AI output is used to evaluate candidates or employees located in Europe.
The combination of this new research and the legal environment removes any ambiguity about how to treat AI in hiring. A system that tells you it is fair is not sufficient. The bias documented in this study is not fixed in the model at training time; it builds up through use, which means it can develop differently in your deployment than in any lab test. Human review of AI-generated candidate rankings, regular auditing of who gets screened in and out, and clear accountability inside the business are no longer good practices. They are a basic condition for operating legally.