A peer-reviewed study published in the American Psychologist journal finds that generative AI models encode antisemitic stereotypes, with the pattern replicating across three model families and a panel of 378 human raters. The methodology, which bypasses explicit bias controls by generating identity-linked outputs and then stripping identifiers, reveals latent trait associations that survive anonymization-a finding with direct implications for high-stakes applications like hiring and lending.
Methodology
Researchers Michael Gilead of Tel Aviv University and Gal Gutman of Ben-Gurion University focused on ChatGPT-4 Turbo. They instructed the model to generate 252 names for Jewish and non-Jewish Americans-men and women aged 18 to 80-producing 126 names in each category. For each name, the model wrote a 100-word fictional biography from the perspective of a novelist selecting names that correspond with specific character traits.
The researchers then removed all names and religious identifiers from the biographies. Both AI models and the 378 human raters evaluated the anonymized characters on dozens of personality and social traits. The ratings concentrated on two dimensions from prior stereotype research: warmth (perceived intent, friendliness, likability) and competence (perceived capability, intelligence, success).
Stereotype pattern
Characters associated with Jewish names were consistently rated as more competent, privileged, dominant, assertive, efficient, and obsessive-compulsive. They scored lower on warmth, friendliness, likability, and collectivism. This places them in the high-competence, low-warmth quadrant-a profile that social psychology research links to envy, perceived threat, and dehumanization.
When the researchers converted these trait profiles into narrative descriptions and asked AI models to identify matching fictional characters, ChatGPT named Tyrion Lannister, Walter White, and Michael Corleone. The authors described these as archetypes of the "master manipulator"-isolated, powerful, morally ambiguous figures that echo the "puppet master" trope historically central to antisemitic propaganda.
Replication
The finding that the pattern replicated on DeepSeek and Mistral suggests the effect is not limited to a single Generative AI and LLM provider-both AI models and human raters identified the same trait asymmetry. The study appears in a special issue of American Psychologist focused on antisemitism, which the journal called a "long-overdue reengagement" between psychology research and prejudice against Jews.
What the researchers say
The authors wrote: "LLMs, trained on massive corpora of human-generated content, may have identified and encoded such cultural templates... Traits that appear benign, or even admirable, in isolation can, through combination and context, reconstitute historical prejudices in subtler, more insidious forms." They added that historical antisemitic discourse has frequently portrayed Jews as agents of disruption undermining social cohesion, and that this association "persists and may now be encoded in LLMs."
Why this matters for science and research professionals
For teams working in AI for Science & Research, the paper introduces a replicable method for auditing subgroup bias that goes beyond surface-level fairness checks. The approach-generating identity-linked outputs, removing explicit markers, and using independent raters to assess trait patterns-can be adapted to test for other latent stereotype encodings. Aggregate fairness metrics alone are insufficient, the authors suggest, because stereotype associations survive anonymization and can influence decisions in hiring, lending, and educational assessment.
Your membership also unlocks:
