31,000-question study finds LLMs across 24 languages are secretly obsessed with Japan
dioscuri · x · 2026-10-22
A study testing Gemini, ChatGPT, Claude and Llama with 31,000 questions across 24 languages found pervasive cultural bias: with all geographic hints stripped, models disproportionately default to Japanese norms, values and aesthetics regardless of prompt language.
The author jokes this is "emergent Japanese alignment," quipping that making models ruder, less helpful and less honest might make them obsessed with Paris instead — hinting the skew may be tied to helpfulness training objectives.
More from Models
- Decision Index 0.4: 224 models scored on private benchmarks only to curb benchmaxxing — multimodalart · 2026-10-22
- Codex is painfully slow, Claude over-refuses: one developer stuck between both — not_rian · 2026-10-22
- I ran 40 PRs through Opus, Sonnet and GPT-6: only 11 of 37 bugs were caught by all three — ParticularFan1122 · 2026-10-22
- Whatever happened to the Yi family of LLMs? Nostalgia for 01.AI's models — code_star · 2026-10-22
- Researcher uses Grok to surface 227 suspected UAE bot accounts overnight — FlorianGallwitz · 2026-10-22
- unsloth ships GGUF quantizations of RecursiveSI/flint on Hugging Face — unsloth · 2026-10-22