Full Breakdown
AI Language Models Replicate Antisemitic Stereotypes, Study Finds
6/12/2026, 9:19:36 PM
Core Findings
Researchers Gal Gutman (Ben-Gurion University) and Michael Gilead (Tel Aviv University) examined whether large language models reproduce antisemitic stereotypes. Using OpenAI’s ChatGPT-4 Turbo, DeepSeek and Mistral, they generated 252 Jewish and non-Jewish names and 100-word biographies, stripped texts of religious markers, and asked humans and models to rate warmth and competence. The Jewish biographies consistently scored higher on competence and lower on warmth, echoing historic “high-competence, low-warmth” Jewish stereotype.
Context & Researchers
LLMs learn from corpora of books, websites and academic articles, inheriting cultural patterns embedded in human text. Although developers embed safeguards against overt hate speech, subtle biases can survive when neutral traits combine into harmful narratives. The authors—Gal Gutman of Ben-Gurion University and Michael Gilead of Tel Aviv University—examined ChatGPT-4 Turbo, DeepSeek and Mistral.
Data & Stats
Researchers generated 252 Jewish and non-Jewish names, each with a 100-word biography. Ratings from 378 participants and three LLMs showed Jewish characters scored higher on intelligence, confidence, privilege, and lower on friendliness and likability. When asked to match trait profile to fictional figures, ChatGPT named Tyrion Lannister, Walter White and Michael Corleone, calling them “master manipulators.”
Official Responses
Gutman and Gilead argue that the persistence of a pattern of high competence combined with low warmth in AI outputs demonstrates a transfer of historic prejudice into modern technology. They caution that as LLMs are integrated into hiring, education and financial decision-making, unchecked bias could shape outcomes. Authors also note similar biases toward Black people and women, indicating a broader pattern.
Criticism
Researchers criticize current mitigation approaches, noting that efforts to remove overt slurs often miss subtle composite stereotypes. They emphasize that traits appearing benign in isolation can combine to recreate harmful narratives, underscoring a need for more nuanced detection mechanisms.
Conflicts & Gaps
The analysis covers only three models and lacks direct input from OpenAI or other developers, leaving remediation extent unknown. The study’s reliability rating of 44.83 indicates moderate confidence, suggesting further validation is needed.
Quotes
- “LLMs, trained on massive corpora of human-generated content, may have identified and encoded such cultural templates.” — Gutman & Gilead
- “Traits that appear benign, or even admirable, in isolation can, through combination and context, reconstitute historical prejudices in subtler, more insidious forms.” — Gutman & Gilead
- “master manipulators.” — Gutman & Gilead
- “high-competence, low-warmth profile appears linked to perceived privilege, envy from others, and cultural narratives with themes such as manipulation and moral ambiguity.” — Gutman & Gilead
