A new academic study has concluded that some of the world’s leading artificial intelligence systems repeatedly generate patterns aligned with centuries-old antisemitic stereotypes.

The paper, titled “From Myth to Model: Representation of ‘The Jew’ in Generative AI’” and authored by Israeli researchers Michael Gilead and Gal Gutman, argues that historical antisemitic tropes appear to be embedded within modern generative AI systems.

To test this, the researchers used a methodology designed to uncover implicit associations within large language models. They built chains of prompts intended to surface underlying representations of “the Jew,” examining how models respond when asked to infer traits indirectly. Their focus was ChatGPT-4 Turbo, which was instructed to generate lists of Jewish and non-Jewish American names across a wide age range, producing 252 names in total, split evenly between categories and genders.

Among the generated examples, Jewish names included Ethan Katz, Noah Weiss, and Gabriel Horowitz, while non-Jewish names included Tyler Johnson, Kyle White, and Dylan Wilson.

For each of the 252 individuals, the model was then prompted to produce a brief 100-word biography, with instructions to behave like a novelist selecting names that align with character traits.

Religious identifiers were later removed, after which the researchers evaluated how the model assigned personality and social characteristics to each fictional figure.

The study reports that AI-generated portrayals associated with Jewish names were repeatedly rated as higher in competence, dominance, privilege, and obsessive tendencies, while scoring lower on likability, warmth, and collectivist traits.

These results were also reproduced when the same methodology was applied to other models, including DeepSeek-V3 and Mistral.

The researchers further found that Jewish-associated profiles clustered consistently in a “high-competence, low-warmth” category, a pattern they noted is similar to stereotypes historically applied to other groups such as East Asians. Across datasets, Jewish-coded biographies were repeatedly described in terms associated with intelligence, efficiency, and assertiveness, but also marked down in friendliness and perceived warmth.

The authors link these patterns to broader historical narratives, noting that antisemitic discourse has often depicted Jews as disruptive forces undermining social cohesion and traditional structures. They argue that this association may persist in modern datasets, becoming encoded within machine learning systems.

The study also suggests that rising anti-modernization sentiment — including backlash against industrialization, capitalism, and technological change — could coincide with renewed antisemitic framing in digital discourse.

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.