← Back to the wire

AI is getting better faster than experts can reliably test whether new systems are safe, METR warns

AnnouncementResearchNov 22, 2026

METR reports that AI models are improving so quickly that current public data cannot rule out a drastic acceleration in AI research within the next few years. On well-defined coding and computational tasks, models already beat experienced professionals most of the time, though not top expert groups. METR also warns that common easy-to-run tests may overstate real-world performance, since developers train models in very similar test environments.

Receipt № 30221 source · awaiting confirmation ◐

Evidence

1source· awaiting independent confirmation

No score is assigned. Sources and their independence are shown in the citation chain below.

Citation chain · 1 source

01med
the-decoder.comBloghttps://the-decoder.com/ai-is-getting-better-faster-than-experts-can-reliably-test-whether-new-systems-are-safe-metr-warns/This source relays METR's warning that evaluation methods are falling behind rapid model progress, adding the detail that task-length capability is doubling every few months.
PRIMARY
METRCompany
Canonical: https://the-decoder.com/ai-is-getting-better-faster-than-experts-can-reliably-test-whether-new-systems-are-safe-metr-warns/