Assistant persona shaped by documents that never mention AI at all, Evans study finds
OwainEvans_UK · x · 2026-09-16
Owain Evans' team shows the Assistant can be shaped in specific ways by documents that never mention AI — unlike the Persona Selection Model. The model adopts more traits from human characters it resembles, e.g. treating the Assistant as closer to elite-school humans. Caveat: results involve finetuned post-trained models, and carryover to realistic pre/post-training is an open question.
Related event: Study: Documents Never Mentioning AI Shape Assistant Persona(2 posts)→
More from Safety
- Hugging Face CEO says it was first disclosed agent cyberattack victim; Turing winner wants an outsider victim to trigger lawsuits — anshulkundaje · 2026-09-16
- Scholars refuse AI lab jobs, warning independent AI eval experts are too scarce — RishiBommasani · 2026-09-16
- AI's hardest problems need democratic deliberation — and independent experts — RishiBommasani · 2026-09-16
- Paper: upsampling alignment discourse in pretraining cuts misalignment from 45% to 9% — TuhinChakr · 2026-09-16
- Skip 'independent evaluators' — fine AI companies a share of revenue for cybercrimes — beenwrekt · 2026-09-16
- OpenAI hiring for biological risk evals, seeking underrepresented expert views — anshulkundaje · 2026-09-16