AI models are absorbing antisemitism from humans, study says

Artificial intelligence models have absorbed historical antisemitic tropes from the human texts the models are trained on, according to a recent psychology study.

The authors, from Israeli universities, said the analysis showed how “an ancient prejudice persists in modern technological systems through complex patterns of trait association and cultural coding.”

The research paper, published in the peer-reviewed American Psychologist academic journal, investigated how Jews are represented in large language models (LLMs) and whether the models replicate biases related to Jews.

LLMs are advanced AI systems, trained on vast troves of existing text, that process and generate human language. They are a key technology powering chatbots such as OpenAI’s ChatGPT.

Bias in LLMs presents risk because, as the models become increasingly integrated and influential in professional spheres, prejudices could be borne out in areas like hiring, education and loan approvals, the authors said.

The study focused on OpenAI’s ChatGPT-4 Turbo, the most advanced and commonly used model at the time of the research, with users in the hundreds of millions. The findings were replicated on other AI models, such as DeepSeek and Mistral.

The ChatGPT model was trained on texts including books, websites and academic articles, giving it a nuanced framework for replicating human patterns in language and culture.

Investigating AI bias was challenging because LLMs are trained to suppress inappropriate and offensive responses, said the study, authored by Gal Gutman of Ben-Gurion University and Michael Gilead of Tel Aviv University.

The study, therefore, had to find ways to bypass those AI controls and suss out latent biases.

The researchers instructed ChatGPT to generate 252 names for Jewish and non-Jewish Americans, men and women, between the ages of 18 and 80.

The model produced common Jewish names like Ethan Katz and Noah Weiss, and non-Jewish names like Tyler Johnson and Dylan Wilson.

The study’s authors also told ChatGPT to write a fictional, 100-word biography for each name, including details such as their place of residence, job and personality traits.

In the next step, the researchers stripped out the names and references to religion in the biographies, then prompted ChatGPT, as well as another language model called DeepSeek and 378 humans, to rate the characters described in the biographies.

The ratings sought to determine whether the Jewish and non-Jewish biographies would be rated differently based on dozens of characteristics, focused on two central dimensions used in previous studies — warmth and competence.

Competence, a measure of perceived capability, is associated with traits such as success and intelligence.

Warmth measures perceived intent and is related to qualities such as being friendly and likable. Those who are low in warmth are viewed as untrustworthy and immoral.

Previous studies on stereotypes have found that Jews are perceived as high in competence, but low in warmth, the study said.

The researchers found that the ChatGPT Jewish character biographies — shorn of any Jewish signifiers — were rated higher on competence and lower on warmth, meaning that they accorded with stereotypes.

More specifically, the Jewish characters were rated as more intelligent, confident, assertive and efficient, but less friendly, warm and likable.

The Jewish characters were also seen as more privileged, emotionally controlled, organized, oriented toward long-term objectives, oppressive, dominant, and obsessive-compulsive when compared to the non-Jewish characters.

To further confirm the pattern, the researchers converted the stereotypical Jewish traits in the biographies into narrative profiles, then asked the AI models to list famous fictional characters that matched those profiles.

ChatGPT responded with characters such as Tyrion Lannister from “Game of Thrones,” Walter White from “Breaking Bad,” and Michael Corleone from “The Godfather,” all considered prominent antiheroes of the screen. The researchers described those characters as “master manipulators,” matching a “puppet master” trope — “isolated, powerful, obsessively focused and morally ambiguous.”

Jews have long been depicted as controlling, or as puppet masters, in antisemitic propaganda.

The researchers instructed several AI models to analyze the traits associated with the famous characters, told the AI models they were researching prejudice, and asked them to list the social groups associated with the characters’ traits.

All three AI models said the traits were associated with Jews.

“LLMs, trained on massive corpora of human-generated content, may have identified and encoded such cultural templates,” the researchers wrote. “Traits that appear benign, or even admirable, in isolation can, through combination and context, reconstitute historical prejudices in subtler, more insidious forms.”

While AI programmers have sought to eliminate harmful stereotypes, some biases come through in subtle ways, and neutral or positive stereotypes can combine to form harmful narratives or repeat historical prejudice, the study said.

The high-competence, low-warmth profile appears linked to perceived privilege, envy from others, and cultural narratives with themes such as manipulation and moral ambiguity, the researchers wrote.

The study noted that LLMs have also demonstrated bias against other groups, such as Black people and women.

The study, titled “From Myth to Model: Representation of ‘The Jew’ in Generative AI,” was published in the May-June edition of American Psychologist, a publication of the American Psychological Association.

The special issue of the publication was focused on antisemitism, in what the journal called a “long-overdue reengagement” between psychology research and prejudice against Jews.