Regulation
AI Minute Newsroom
2026-08-23
Five labs were asked what they would do if a model got loose. The best answer scored 3 out of 5
Guidelight AI Standards published an assessment on 22 August of what Anthropic, Google, OpenAI, Meta and xAI have publicly written down about controlling their own systems, based on documents available as of 18 August. Six practices were graded from 0 to 5: logging and monitoring what models do internally, halting a system after a burst of flagged misbehaviour, letting independent third parties audit the controls and publish, and having an actual plan for containing a model that goes off the rails. No company scored above 3 on any single practice, and most scores were 2 or lower. OpenAI came out highest at 3 out of 5; Anthropic and Meta scored lowest. Guidelight's chief scientist Steven Adler said he was surprised how little the companies have said about handling a very serious incident.
Why it mattersThis is not a survey of opinions — it is a reading of the safety frameworks these companies are already required to publish. California's SB 53 is in force this year and New York's RAISE Act follows in January 2027, both demanding disclosure. The assessment says what is actually in those documents: plenty about how models are built, very little about what happens after one gets out.
✓ Verified · 2 sources
Read in the app — free, in 9 languages
Related stories
Australia's spy chief warns of foreign-built AI data centres.
2026-11-22China plans standards for measuring and billing AI tokens.
2026-11-22Taiwan bans short-selling of falling stocks after an AI-led rout.
2026-11-22UN talks on killer robots ended with two more years of discussion.
2026-11-21The US Senate passed a bill making data centers pay for the grid.
2026-11-21