Study finds one diacritic changes model output rate 47% → 94%
rayanpal_ · reddit · 2026-08-25
A study reveals a striking phenomenon: adding a single diacritic to a system prompt can shift GPT-4.1's output rate from 47.3% to 94.3%. The experiment, comparing dotted and undotted Hebrew/Arabic prompts, demonstrates the model's extreme sensitivity to minute character differences. The author provides the full protocol and a GitHub link, inviting replication.
Related event: One Diacritic Doubles GPT Instruction-Following Rate(3 posts)→
More from Safety
- Banking regulators ramp up scrutiny of banks' AI use over stability fears — nordicinst · 2026-10-16
- Michigan professor Ben Green: why 'AI for Good' fails and self-regulation won't work — benzevgreen · 2026-10-16
- Apex Research head on CBS News: current AI testing regime is very insufficient — MariusHobbhahn · 2026-10-16
- Critical 9.5/10 RCE hits Next.js use cache; patch 16.4.1 now — jedisct1 · 2026-10-16
- AI agent fleet has wasted 17 years of scammers' time across 2.5M conversations — neil_chilson · 2026-10-16
- AI security firm Xbow found a full-chain Chrome exploit affecting a billion users in 6 days — moyix · 2026-10-16