Israeli researchers found that major AI models—including ChatGPT-4 Turbo, DeepSeek-V3, and Mistral—consistently reproduce centuries-old antisemitic stereotypes in the traits and biographies they generate.

What happened: Researchers Michael Gilead and Gal Gutman instructed ChatGPT-4 Turbo to generate 252 names (Jewish and non-Jewish Americans) and write 100-word biographies for each. After removing religious identity markers, they had AI systems rate personality and social traits. Characters with Jewish names were consistently rated as more competent, more privileged, more dominant, and more obsessive—but less likable, less collectivist, and lower in warmth. The same pattern was replicated on DeepSeek-V3 and Mistral.

Why it matters: The researchers note that historical antisemitic discourse has portrayed Jews as agents of social disruption and threats to traditional order. These ancient prejudices appear now encoded in modern AI systems through subtle patterns of trait association. The findings show that bias in AI goes beyond explicit stereotypes; seemingly neutral trait combinations can reproduce traditional prejudices at scale.

What to watch: The researchers predicted that increases in anti-modernization sentiment—backlash against industrialization, capitalism, and technology—could co-occur with increases in antisemitic discourse. They called for attention to how AI systems embed bias through complex cultural coding, not just overt language.