AI Language Models Replicate Centuries-Old Antisemitic Stereotypes, Study Finds

Historical antisemitic discourse has frequently portrayed Jews as agents of disruption, emphasizing low warmth and high competence. The researchers note that AI systems can reflect such entrenched prejudice through subtle trait associations rather than explicit language.

The Jerusalem Post

Israeli scholars Michael Gilead and Gal Gutman authored “From Myth to Model: Representation of ‘The Jew’ in Generative AI”. They examined three leading large-language models—ChatGPT-4 Turbo, DeepSeek-V3 and Mistral—using a multi-step indirect approach. The team generated 252 fictional American profiles (126 Jewish, 126 non-Jewish) aged 18-80, each with a 100-word biography. After removing religious identifiers, the model outputs and 378 human raters were asked to evaluate warmth (friendliness, likability) and competence (intelligence, success).

Study finds LLMs reproduce antisemitic stereotypes

Evaluated profiles: 252. Human raters: 378. Traits measured: warmth-related (likability, friendliness, collectivism) and competence-related (intelligence, efficiency, dominance). Across all models, Jewish-linked profiles scored higher on competence, privilege, dominance and obsessiveness, and lower on warmth, likability and collectivism. The evaluation covered dozens of personality and social traits.

Study finds LLMs reproduce antisemitic stereotypes

The analysis shows Jewish-associated biographies consistently fall into a “high-competence, low-warmth” quadrant, echoing historic antisemitic tropes that portray Jews as socially disruptive. The pattern mirrors stereotypes previously applied to East Asian groups.

The Jerusalem Post

The authors argue that bias can persist through subtle trait associations rather than overt slurs, urging bias-detection tools to target structural patterns embedded in model behavior. The findings highlight the risk that generative AI may reinforce longstanding prejudices.

Israeli Study Finds Major AI Models Reproduce Longstanding Antisemitic StereotypesStudy finds LLMs reproduce antisemitic stereotypesThe Jerusalem Post

The study covers only three models and includes no statements from their developers. No independent replication has been reported, and the work does not compare results with baseline bias-mitigation models, leaving the broader applicability uncertain. The study does not discuss mitigation techniques.

Study finds LLMs reproduce antisemitic stereotypes

• “The researchers' analysis found that Jews in LLM-generated content are consistently stereotyped within the high-competence, low-warmth quadrant, alongside groups such as East Asians.” — Michael Gilead, researcher

The Jerusalem Post

• “LLM-generated content stereotypes Jews as low on warmth-related traits The researchers found that characters associated with Jewish names were consistently rated as more competent, more privileged, more dominant, and more obsessive.” — Gal Gutman, researcher

The Jerusalem Post

• “the ailments of modern subjectivity [...] persists and may now be encoded in LLMs,” — Michael Gilead & Gal Gutman

The Jerusalem Post

• “The researchers conclude that their findings demonstrate how entrenched prejudice can persist within AI systems through complex trait associations rather than explicit language, emphasizing that bias detection must go beyond overt stereotypes to include more subtle structural patterns embedded in model behavior.” — Study conclusion

Israeli Study Finds Major AI Models Reproduce Longstanding Antisemitic Stereotypes

The authors recommend expanding bias-detection methods to target subtle structural patterns embedded in model behavior.

Israeli Study Finds Major AI Models Reproduce Longstanding Antisemitic Stereotypes