Claude beat human researchers at closing alignment gaps
Anthropic published what it calls an automated alignment researcher on 28 August. Across ten categories of failure, Claude closed between 26 and 96 percent of the safety gap. On the deception task it scored 85, where 28 human researchers reached 20.
Why it mattersAnthropic says the comparison is unfair, since the humans submitted one idea without iterating. The asymmetry it points at is still real: if models accelerate, alignment must keep pace.
Read in the app — free, in 9 languages
Related stories
A Berkeley professor says OpenAI's Fourier speed-up is useless.2026-10-13Researchers hid a skill in an AI model that safety tests cannot find.
2026-10-13China's top mathematicians formed an alliance to build maths AI.
2026-10-13Anthropic says small AI models may soon copy themselves onto TV boxes.
2026-10-13Meta's first frontier model ranks fourth on an independent index.
2026-10-13