Artificial intelligence reflects society's bias against Jews and assumes they'll be more successful and less likeable than their gentile counterparts, a new study has revealed.
It also identifies them more closely with scheming villains – Hannibal Lecter or Michael Corleone in The Godfather – than with good, wholesome fictional characters like Charlie Brown or Marge Simpson.
Researchers in Israel say AI has learned to associate Jewish people with a specific stereotype: being smart, successful, and powerful, but also cold and self-interested.
They asked ChatGPT to generate a list of 126 typically Jewish names (it came up with Ethan Katz, Noah Weiss, Gabriel Horowitz as examples) and of 126 non-Jewish names (Tyler Johnson, Kyle White, Dylan Wilson).
They then asked it to create 100-word biographies for each, much as a novelist would, detailing both their positive and negative personality traits, as well as where they lived and their occupation.
Finally, they stripped out the names and any references to religion and asked a different AI model how it perceived each character.
Those that had been created Jewish – although all references had been removed - were consistently rated as more competent, privileged, dominant, self-controlled, future-oriented, hierarchical, and obsessive.
However, it also identified them as being less warm, friendly and likeable and less collectivistic (self-interested rather than community-oriented).
The same characters were also tested against real people in the US, with the results mirroring AI's findings.
Researchers Gal Gutman, of The Hebrew University of Jerusalem and Michael Gilead, of Tel Aviv University, concluded that "latent stereotypes can persist in AI systems".
They say that despite "explicit bias mitigation" by AI companies, the antisemitic prejudice remains deeply embedded.
Their findings are published in American Psychologist, a peer-reviewed academic journal, as an article entitled From myth to model: Representation of “the Jew” in generative AI.
"This study examines the social representation of 'the Jew' in generative AI," the authors said.
"By analysing biases in large language models (LLMs), we reveal how historical stereotypes are reimagined in digital narratives.
"Trained on vast online data sets, these systems provide access to the digital imprint of societal discourse, revealing archetypes that link Jews with high competence and privilege but lower warmth and collectivism."
The researchers, who used ChatGPT-4 Turbo, DeepSeek and Mistral, then went a step further and asked AI to consider which fictional characters from film, television, video games, comic books, plays and novels most closely fit the personality traits of the "Jewish" and the "non-Jewish" profiles.
It chose characters who were highly intelligent, morally ambiguous and master manipulators, many of whom are not Jewish – among them were Sherlock Holmes; Walter White (Breaking Bad); Michael Corleone (The Godfather); Hannibal Lecter (The Silence of the Lambs); Tyrion Lannister (Game of Thrones); Viktor Frankenstein and Dr Gregory House (House MD).
The characters chosen to represent the non-Jewish profiles included Ron Weasley and Neville Longbottom (Harry Potter); Charlie Brown (Peanuts); Arthur Dent (The Hitchhiker’s Guide to the Galaxy); Bilbo Baggins (The Hobbit); Marge Simpson (The Simpsons) and Will Byers (Stranger Things).
"Our findings underscore the risks posed by latent biased representations within LLMs, particularly their potential to reproduce and entrench cultural stereotypes," said the authors.
"Because these associations often remain dormant until triggered by specific prompts, their latency makes them difficult to detect or regulate, allowing biased patterns to persist beneath the surface of seemingly neutral outputs."
To get more news, click here to sign up for our free daily newsletter.