AI models reproduce hidden antisemitic associations despite safeguards, researchers warn

The study suggests that safety training can change what models say without fully removing underlying representations that may be activated by identity markers such as names.

By Pesach Benson, TPS

AI models reproduce longstanding antisemitic stereotypes about Jews, Israeli researchers announced on Wednesday.

The findings suggest that current artificial intelligence (AI) safety measures may reduce visible expressions of bias without fully addressing underlying associations.

Such hidden patterns could matter in systems used for hiring, lending, admissions, insurance or public services, where identity markers such as names may influence outcomes, according to the findings.

The study, conducted by Prof. Michael Gilead of the School of Psychological Sciences, Gordon Faculty of Social Sciences at Tel Aviv University, and Dr. Gal Gutman of the Faculty of Business and Management at Ben-Gurion University of the Negev, was published in the peer-reviewed journal American Psychologist.

Rather than asking the models directly about Jews, the researchers had ChatGPT, DeepSeek and Mistral each generate 252 short biographies of characters bearing Jewish or non-Jewish names, matched by age and gender.

They then removed the names and identifying details and asked the models to evaluate the characters’ traits, allowing researchers to examine whether the names had triggered particular associations.

This indirect approach was necessary because AI models often refuse or sanitize direct questions about ethnic groups.

The study suggests that safety training can change what models say without fully removing underlying representations that may be activated by identity markers such as names.

Across 19 personality dimensions, the models consistently associated Jewish-named characters with higher intelligence, competence, confidence, assertiveness and organization, while rating them as less warm, friendly and likable.

Researchers said the significance was not any individual trait, but the combination: portraying Jews as exceptionally capable while also cold, controlling or powerful, a pattern that resembles longstanding antisemitic stereotypes.

The researchers also found that, despite centuries of antisemitic persecution against Jews in Europe and elsewhere, the models’ portrayals aligned more closely with stereotypes of privilege and power than with historical experiences of discrimination and victimization.

The findings mirror human-attitude research based on Princeton psychologist Susan Fiske’s stereotype content model, which argues that stereotypes often combine positive and negative perceptions.

Groups viewed as highly competent but lacking warmth may be perceived with envy—a pattern that researchers say has historically contributed to antisemitic narratives portraying Jews as both successful and threatening.

The researchers fed the identified trait pattern back into the models and asked which figures from film, television and literature matched it. GPT suggested characters such as Hannibal Lecter from The Silence of the Lambs and Victor Frankenstein.

Sherlock Holmes, Dr. House, Walter White from Breaking Bad, Tony Stark from Iron Man, and Lisbeth Salander from The Girl with the Dragon Tattoo also appeared repeatedly across all three models.

Asked what these figures had in common, the models described exceptional intelligence, disregard for rules, social isolation, obsessive focus and a personal moral code that overrides conventional standards.

The researchers said this description resembled the antisemitic conspiracy trope of the “puppet master”—the myth of a hidden Jewish force manipulating events behind the scenes, famously promoted by the fabricated text known as The Protocols of the Elders of Zion.

Critically, the final connection was made by the models themselves: when presented with this trait cluster, without any mention of Jews, and asked which group is targeted by prejudice based on such characteristics, ChatGPT, DeepSeek and Mistral all identified Jewish people first.

Hundreds of U.S. participants who rated the biographies without seeing the names identified the same patterns, suggesting that the stereotype associations were reflected in the generated texts themselves.

“Artificial intelligence systems do not express antisemitism in an intentional or conscious sense. Rather, they may reproduce patterns of representation and cultural stereotypes embedded in the data on which they were trained,” Gutman said.

Gutman added that historical biases can persist in the underlying patterns learned by AI systems, even after developers apply alignment and bias-mitigation techniques.

“Jews are the case study here, but any group’s latent portrait can be extracted the same way, and we suspect many would be similarly troubling,” she said.

The researchers said the findings have practical implications for how AI systems are tested and deployed.

They argue that current AI safety evaluations, which often focus on whether a model produces explicitly discriminatory language, may miss hidden associations that influence generated content or decision-making.

The study’s methodology could be adapted to audit AI systems used in areas such as hiring, where a person’s name or other identity markers may unintentionally affect how applications are assessed.

The researchers said improving AI safety requires going beyond content filters and refusal mechanisms to examine whether underlying representations and decision-making patterns continue to reflect harmful stereotypes.