AI chatbots found to mirror antisemitic stereotypes, Israeli study warns

Artificial intelligence systems are reproducing longstanding antisemitic stereotypes despite safeguards designed to prevent biased responses, according to a new peer-reviewed study by Israeli researchers.

The research, published in the latest edition of American Psychologist, found that some of the world’s most widely used AI models associated Jewish people with traits linked to historic antisemitic narratives, raising concerns about the growing influence of artificial intelligence in everyday life.

Researchers from Ben-Gurion University and Tel Aviv University examined how Jews are represented within large language models (LLMs), the technology that powers popular chatbots such as ChatGPT.

The study focused primarily on OpenAI’s ChatGPT-4 Turbo and found similar results when tests were repeated using other AI systems, including DeepSeek and Mistral.

The authors warned that hidden biases embedded within AI models could become increasingly significant as the technology is used in areas such as recruitment, education and financial decision-making.

Researchers said they faced a challenge because AI systems are programmed to avoid producing overtly offensive or discriminatory content. To test for underlying attitudes, they developed a series of indirect experiments designed to uncover what they described as “latent biases”.

The team asked ChatGPT to generate hundreds of fictional biographies for people with either Jewish or non-Jewish names. Names and religious references were then removed before the biographies were assessed by human participants and AI systems.

According to the study, biographies generated for Jewish characters were consistently rated as showing higher levels of competence but lower levels of warmth than those created for non-Jewish characters.

The Jewish characters were more likely to be described as intelligent, confident and successful, but less likely to be viewed as friendly, approachable or likeable.

Researchers also found that Jewish characters were more frequently associated with privilege, dominance, emotional restraint and obsessive behaviour.

The findings echo a well-documented stereotype that portrays Jews as capable and influential while simultaneously casting them as untrustworthy or socially distant.

In a further test, researchers created personality profiles based on the traits identified in the biographies and asked AI models to match them with well-known fictional characters.

Among the names suggested by ChatGPT were Tyrion Lannister from Game of Thrones, Walter White from Breaking Bad and Michael Corleone from The Godfather.

The study’s authors said these figures reflected characteristics often linked to the antisemitic “puppet master” stereotype, in which Jews are portrayed as powerful, manipulative and morally ambiguous.

The researchers wrote: “LLMs, trained on massive corpora of human-generated content, may have identified and encoded such cultural templates.”

They added: “Traits that appear benign, or even admirable, in isolation can, through combination and context, reconstitute historical prejudices in subtler, more insidious forms.”

The paper, titled From Myth to Model: Representation of ‘The Jew’ in Generative AI, appears in a special edition of American Psychologist dedicated to antisemitism.

The authors said their findings demonstrate how historical prejudices can be absorbed and reproduced by modern technologies, even when developers actively try to remove harmful bias from their systems.

The study also noted that previous research has identified comparable forms of bias affecting other groups, including Black people and women.