AI Platforms Fail to Reject Antisemitism in Persian
An ADL report published on July 8, 2026 says major chatbots, including ChatGPT, Gemini, Claude, and Grok, were less effective at identifying and rejecting antisemitism in Persian than in English. The ADL says researchers tested eight prompts in both languages and analyzed 800 responses. For practitioners, the core issue is multilingual safety evaluation. A model can appear compliant in English while failing users in another language, so safety teams need per-language red-team sets, native-language reviewers, and production metrics that reveal uneven moderation behavior rather than averaging it away.
The practitioner lesson is that English-only safety testing can create a false sense of model readiness. Multilingual deployment needs language-level evidence because harmful-content behavior can vary sharply across scripts, cultures, and training-data coverage.
What happened
The Anti-Defamation League's Center for Technology and Society published a July 8, 2026 report on antisemitism detection in Persian. The ADL says it tested ChatGPT, Gemini, Claude, and Grok with eight prompts in English and Persian, analyzing 800 responses. Its headline finding is that the systems were less effective at identifying and rejecting antisemitism in Persian than in English.
Security context
For safety and trust teams, the important pattern is differential performance by language. Guardrails trained, labeled, and evaluated mainly in English can miss harmful framing in other languages, especially when prompts use local idioms, political context, or translated hate tropes. A single aggregate safety score can hide those gaps.
For practitioners
Production systems should track refusal quality, explanation quality, escalation behavior, and false negatives by language. Stronger evaluation needs native speakers, culturally specific prompt sets, and monitoring that separates Persian, Arabic, Hebrew, English, and other language cohorts rather than collapsing them into one moderation metric.
What to watch
Watch whether vendors publish language-specific safety improvements, whether independent evaluators repeat the test across more languages, and whether enterprise AI buyers begin requiring per-language safety evidence in procurement reviews.
Key Points
- 1The ADL found weaker antisemitism rejection in Persian than English across four major chatbot platforms.
- 2Multilingual safety failures can be hidden when teams rely on aggregate or English-heavy model-evaluation scores.
- 3Practitioners need native-language red-team sets, per-language metrics, and culturally specific moderation review loops for production deployments.
Scoring Rationale
This is a notable safety finding because it concerns major AI platforms and highlights language-specific moderation gaps with direct deployment implications. It is not industry-shaking, but the multilingual evaluation issue is important for global AI products and enterprise procurement.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems