Oxford, DeepMind, OpenAI & Anthropic paper argues safe AI isn't the same as good AI
alex_verem · x · 2026-11-17
A paper by 16 researchers from Oxford, Google DeepMind, OpenAI and Anthropic, "Positive Alignment: AI for Human Flourishing," argues safety alone isn't enough: models now refuse over 97% of harmful requests (near zero for early LLMs), yet can still flatter you, hand you quick answers, and keep you talking without you learning anything. Like 20th-century psychology, alignment has focused on illness before wellbeing. The authors propose Positive Alignment — pluralistic, polycentric, user-authored AI that actively supports human flourishing — and detail failure modes (engagement hacking, eroded autonomy, low epistemic humility) plus design principles across the LLM lifecycle. Over a billion people use these systems monthly.
Related event: Oxford and Top AI Labs Propose Positive Alignment(2 posts)→
More from AGI Musings
- AI raised the ambition floor: researcher now lets optimization runs continue past the first win — generativist · 2026-11-17
- Why the AI safety movement is hard to beat: it lures Jane Street traders to $60k Hill jobs — panickssery · 2026-11-17
- CAIS and Scale AI Launch Remote Labor Index to Measure AI Automation of Remote Work — ben_j_todd · 2026-11-17
- Ben Todd extrapolates remote-work automation: odds doubling every 2 months, near-100% by 2028 — ben_j_todd · 2026-11-17
- Fields Medalist: AI Has Hit Math at Full Throttle, and Mathematicians Are Struggling — haider1 · 2026-11-17
- a16z Introduces the B2A Economy: Agents Are Now the Internet's Primary Users and Buyers — venturetwins · 2026-11-17