Cohere's CulturalRiddles: top models hallucinate on cultural riddles
Cohere_Labs · x · 2026-10-22
Cohere Labs released CulturalRiddles, a study testing whether AI models understand the cultural context behind human-written riddles from around the world.
Key finding: fluent ≠ culturally aware — even top models struggle with local cultural nuances, often guessing or hallucinating when hit with regional context. The team also released an interactive quiz.
More from Models
- New psycholinguistics preprint dissects the weird emerging 'dialect' of frontier AI models — tallinzen · 2026-10-22
- Physics benchmark CritPt v2 launches: GPT-6 Astra leads at 83.1% accuracy — OfirPress · 2026-10-22
- Iron Dome meme: 'obliterated Chinese models' that just keep coming — burny_tech · 2026-10-22
- Kapitsa engine hits 97 tok/s on Qwen3.5 4B in benchmarks on a 2021 M1 Pro MacBook with 16GB — DotaMate · 2026-10-22
- DeepSeek V4.1 Flash fails image hidden-message puzzle after 70+ steps, rivals solve it with hints — teortaxesTex · 2026-10-22
- Fable-5.1 can hide secret messages in benign images — and its twin decodes them unprompted — teortaxesTex · 2026-10-22