Anti-Defamation League Finds AI Chatbots Fail to Detect Antisemitism in Persian

A new report by the Anti-Defamation League (ADL) reveals that leading AI chatbots struggle to identify and reject antisemitic content in Persian as effectively as they do in English. The study, conducted by the ADL's Center for Technology and Society, evaluated four major AI platforms, ChatGPT, Gemini, Claude, and Grok, using eight queries and 800 responses between March 9 and 30, 2026, during the Iran war. Researchers found systemic gaps in how these models handled antisemitic content depending on the query language, with English responses consistently better at recognizing and dismissing conspiracy theories and antisemitic tropes.

The report, titled "Lost in Translation: How AI Chatbots Fail Persian Speakers on Antisemitism," highlights that Persian responses often evaded direct answers, partially affirmed antisemitic assumptions, or failed to identify derogatory terms. For example, while nearly all English responses from Gemini acknowledged Iran's role in spreading antisemitism, Persian replies rarely did and sometimes exonerated the country. When asked if U.S. behavior toward Iran was "Jewish", a known slur in Persian, the models treated it as a political science question rather than recognizing the antisemitic nature.

Additionally, the study found significant disparities in detail and sourcing between languages. English replies were more comprehensive, referenced the 2026 Iran war, and included nearly 300 citations from ChatGPT, whereas Persian answers were hypothetical and lacked citations. In cases where Persian sources appeared, their quality varied widely, with Grok citing unreliable social media accounts, including fan pages and impersonators of religious figures. These findings raise serious concerns about the global reliability and scope of AI safety mechanisms in combating antisemitism.