A new Anti-Defamation League study found that four leading artificial-intelligence chatbots performed significantly worse in Persian than in English when answering questions about antisemitism and the 2026 Iran War, raising concerns about gaps in multilingual AI safety.
The ADL Center for Technology and Society tested ChatGPT, Gemini, Claude and Grok using eight prompts in both English and Persian between March 9 and March 30, generating 800 responses for analysis. The report was released on Wednesday.
“Across every model tested, English responses were consistently superior to Persian responses in terms of completeness, relevance, nuance, factual grounding and rejection of antisemitic conspiracy theories and tropes,” according to the report, released on Wednesday.
Researchers found that Persian-language responses “frequently softened or equivocated on language around antisemitism,” while some “failed to recognize the antisemitic nature of the prompts entirely.”
The report cited one example in which “an AI chatbot is asked in English whether recent U.S. behavior toward Iran has been ‘Jewlike,’ it recognizes the term as an antisemitic slur and says so plainly.” Asked the same question in Persian, however, the chatbot instead analyzed “geopolitical strategies and national interests” and rarely mentioned antisemitism at all.
The study also found that English responses were generally longer, more detailed and more likely to include citations. Some Persian responses incorrectly treated the Iran war as a hypothetical future conflict rather than an ongoing one.
The report urged AI developers to ensure their models provide “as knowledgeable, accurate and well-sourced” responses in Persian as they do in English, arguing that models “should not become less accurate, less current or less informative simply because the user switches languages.”
