AI Platforms Fail to Reject Antisemitism in Persian

According to a July 8, 2026 report from the Anti-Defamation League (ADL) Center for Technology and Society, researchers tested four major AI platforms, ChatGPT, Gemini, Claude, and Grok, using eight prompts in both English and Persian and analysed 800 responses generated between March 9-30, 2026. The ADL report states the models "fail, without exception, to identify and reject antisemitism as effectively in Persian as they do in English." The report documents that Persian responses "frequently softened or equivocated on language around antisemitism," that English answers were longer and more likely to include citations, and that some Persian responses treated the Iran war as hypothetical rather than ongoing. ADL CEO Jonathan Greenblatt said, "These findings are deeply troubling." The Jerusalem Post and JNS published summaries that mirror the ADL findings.

Multilingual discrepancies typically arise from three industry-observed causes: uneven training-data coverage across languages, differences in moderation-data labeling and policy alignment, and evaluation regimes that prioritize English benchmarks. In practice, smaller corpora and weaker signal for hate-speech patterns in Persian-language training data make it harder for a model to learn both the lexicon of slurs and the contextual patterns that indicate targeted hate. Similarly, content-safety fine-tuning and instruction-following steps often rely on English-labeled examples, producing uneven guardrails across languages.

Teams building detection, moderation, or downstream services should treat language as an axis of model failure, not just a UI convenience. Observed patterns in similar evaluations suggest practitioners will need to expand test suites to include localized hate-speech examples, invest in native-language annotators for policy-consistent labels, and instrument production telemetry to compare signal (length, citation presence, refusal rate) across languages.

Use bilingual prompt testing and sample parity metrics (response length, citation rate, refusal/refusal-quality) when onboarding third-party models. Where regulatory or reputational risk is high, include adversarially generated non-English examples in red-team exercises and post-deployment monitoring.

Observers should watch for follow-up reporting from model providers or independent audits describing remediation steps, expanded multilingual safety datasets, or updated moderation pipelines. Also track whether the ADL's methodology (eight prompts, 800 responses) is reproduced at scale for other languages and geopolitical events.

The factual claims above are taken from the ADL Center for Technology and Society report (July 8, 2026) and contemporaneous reporting in The Jerusalem Post and JNS/Cleveland Jewish News, which summarize and amplify the ADL findings.