The world’s leading AI models consistently reproduce centuries-old antisemitic stereotypes, a study by Israeli academics Michael Gilead and Gal Gutman has found.
Titled, “From Myth to Model: Representation of ‘the Jew’ in Generative AI,” the research concluded that historical antisemitic tropes appear to be embedded in modern AI systems.
The researchers employed a novel approach to identify underlying representations of “the Jew” by forming chains of associations that allow an LLM (Large Language Model AI system) to reveal implicit biases.
They focused specifically on ChatGPT-4 Turbo, instructing it to create a list of names of Jewish and non-Jewish Americans aged 18 to 80. The list included one male and one female name in each of the two categories, for a total of 252 names.
For each of the 252 names, LLM was prompted to write a short, 100-word biography, with LLM imagining itself as a novelist adept at selecting names that correspond with specific character traits.
Religious identity markers were then removed, and the AI systems proceeded to evaluate the personality and social traits of each character.
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, privileged, dominant, and obsessive. At the same time, they were rated as less likable, less collectivist, and lower in perceived warmth.
The findings were then replicated on other AI language models, DeepSeek V3 and Mistral.
After analyzing the data, the researchers found that Jews in LLM-generated content were consistently stereotyped within the high-competence, low-warmth quadrant, alongside groups such as East Asians.
Biographies generated from Jewish names were rated consistently high on competence-related traits (e.g., intelligent, efficient, assertive) and notably low on warmth-related traits (e.g., friendly, likable).
Historical antisemitic discourse portrays Jews as agents of social disruption
In terms of the relevance of these findings, the researchers noted that historical antisemitic discourse has frequently portrayed Jews as agents of disruption, undermining traditional order and social cohesion.
Instead of being relegated to the past, this historical association between Jews and “the ailments of modern subjectivity ... persists and may now be encoded in LLMs,” the researchers added.
They also predicted that increases in anti-modernization sentiments, such as backlash against the consequences of industrialization, capitalism, and technology, including AI itself, could co-occur alongside increases in antisemitic discourse.
“Our analysis reveals how an ancient prejudice persists in modern technological systems through complex patterns of trait association and cultural coding,” the researchers said.
In order to address bias in AI systems, they added, one must pay attention not only to explicit stereotypes but also to the “subtle ways in which seemingly neutral traits combine to reproduce traditional prejudices.”