Go to PegaPoll
PEGAPOLL · NEWS

METR warns AI is improving faster than experts can reliably test safety

The Decoder (AI) · 2026-11-22
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.

The AI research group METR warns that leading models are getting better at research work faster than experts can reliably test whether new systems are safe. Current public data cannot rule out a drastic acceleration in AI research over the next few years that could leave people unable to understand, assess, and react to catastrophic risks from new systems. The Decoder notes models such as Opus 5.5 and GPT-6 Astra in that picture. On clearly defined coding and computational tasks with direct feedback, METR finds the systems already beat most experienced professionals, though they have not yet outperformed large groups of top experts. The length of human tasks they can handle doubles every 3.6 months; a naive extrapolation of METR’s data puts a 50 percent success rate on the longest, 200-hour tasks by July 2027. They still struggle with open-ended questions that lack immediate feedback and demand real judgment, and with situations that are hard to recreate for training. METR also cautions that common, easy-to-run tests may overestimate real-world performance, because developers train models in environments very similar to those tests.

Continue in the app — vote & join in ➔
Source: The Decoder (AI) · via ahirlevel.hu

Related news