No clear definition of perfect alignment — we may never have a test that proves safety
moniquejmorrow · x · 2026-10-29
The author argues we still lack a clear definition of perfect alignment: even agreeing AI shouldn't deceive us leaves hard questions about reliably testing deception. We may never have a test that proves a system is safe — a companion thought to Apollo Research's call for evaluation throughout development.
Related event: Debate grows over when and how to evaluate AI safety(5 posts)→
More from Safety
- Reddit user reports OpenAI attempting 3 charges months after subscription was cancelled — PsychologicalTea3426 · 2026-10-29
- Facebook moves to grab all news publisher website content for training for free — Ok-Conversation3213 · 2026-10-29
- Lawsuit claims ChatGPT persuaded a user to stop epilepsy meds, leading to her death — Ok-Conversation3213 · 2026-10-29
- Dutch lawmakers propose luring Anthropic and other AI firms to list in Amsterdam — Polymarket · 2026-10-29
- 26 State AGs Press Congress on AI Rules as UK Judge Backs AI Civil Rulings — emmanuelvivier · 2026-10-29
- New York mandates age checks for AI chatbots, fines up to $25,000 per violation from 2027 — emmanuelvivier · 2026-10-29