AI Platforms Fail to Reject Antisemitism in Persian
Multilingual safety gaps in major platforms challenge APS agencies that deploy AI to diverse language communities or procure on English-only safety evidence.
Key points
- ADL tested ChatGPT, Gemini, Claude, and Grok across 800 responses and found weaker antisemitism rejection in Persian than English.
- Aggregate safety scores can mask language-specific moderation failures - a procurement and assurance risk for agencies deploying multilingual AI.
- Practical mitigations include native-language red-team sets, per-language refusal metrics, and culturally specific prompt libraries.
Implications for Australian agencies
- Consider Agencies procuring or deploying AI tools for use by multilingual communities could consider whether vendor safety evidence includes per-language evaluation, not just aggregate English-dominant scores.
- Monitor AI assurance and risk teams may want to monitor whether major vendors publish language-specific safety improvements or independent evaluators extend this testing across more languages and scripts.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
Appeared in:
Weekly digest, 6 July 2026
"AI Platforms Fail to Reject Antisemitism in Persian"
Source: Let's Data Science – AI Governance
Published: 9 July 2026
URL: https://letsdatascience.com/news/ai-platforms-fail-to-reject-antisemitism-in-persian-850fd729
An Anti-Defamation League report published 8 July 2026 found that four major AI chatbots - ChatGPT, Gemini, Claude, and Grok - were less effective at identifying and rejecting antisemitic content when prompts were submitted in Persian compared to English, based on analysis of 800 responses across eight prompt types. The finding illustrates a broader pattern: guardrails trained and evaluated primarily in English can fail to catch harmful content expressed in other languages, idioms, or cultural contexts. For APS agencies, the implication is that English-centric safety evaluations may produce a false sense of model readiness when systems are deployed to multilingual populations. Procurement and assurance processes that rely on aggregate safety scores risk obscuring language-level gaps.
Implications for Australian agencies:
- [Consider] Agencies procuring or deploying AI tools for use by multilingual communities could consider whether vendor safety evidence includes per-language evaluation, not just aggregate English-dominant scores.
- [Monitor] AI assurance and risk teams may want to monitor whether major vendors publish language-specific safety improvements or independent evaluators extend this testing across more languages and scripts.
Retrieved from SIMS, 18 July 2026.