AI BriefingCohere Labs open science project reveals AI models hallucinate cultural knowledge in riddle tests
·2026.10.22 20:00
Key point
A global initiative with over 100 contributors from 50 countries found that AI models often provide authoritative but incorrect answers to culturally specific riddles.
Details
- Cohere Labs and Hungarian diplomat Boczka Károly launched an open science project to test AI models' cultural understanding using riddles, revealing that models frequently hallucinate culturally specific facts.
- The project gathered over 100 contributors from more than 50 countries, including a 16-year-old student and a retired teacher, to build benchmarks across sports, history, and politics.
- AI models trained mostly on English data showed virtually zero knowledge of less common languages like Buginese, leading to incorrect answers that sounded authoritative.
- Participants noted that models rarely admitted ignorance, instead providing confident but wrong responses, highlighting a gap between AI fluency and true cultural knowledge.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI Safety Leader David Robinson Quits, Citing Broken Culture and Insufficient Care
Google Research Releases Harness² for Recursive Agent Self-ImprovementVals Launches Independent Router Evaluation on Real-World Agentic Tasks