METR disputes confidence in Anthropic’s AI research risk account

METR said on November 5 that Anthropic’s public account of AI research acceleration overstated confidence about staying below an early-warning threshold. It agreed direct catastrophic risks on July 15 were low.

A separate METR assessment of October progress put the pace of improvement in Anthropic’s internal capability index, AECI, since October 2025 at 1.9 times its earlier rate, with a 95% confidence interval of 1.5–2.8. This measures the pace of capability improvement.

The document under review was published on August 14 and covered models as of July 15. Anthropic described substantially faster research with AI assistance, short of a doubling. It also reported reduced confidence as task-based evaluations became saturated, limiting their ability to distinguish increasingly capable models.

Anthropic’s Responsible Scaling Policy 3.4 sets two routes to its automated-research threshold: replacing the entire research staff at competitive cost, or AI substantially contributing to doubled capability progress. The latter compares expected and fastest sustained rates without significant AI contributions, with continuation likely to bring further acceleration.

METR called the public case against crossing that second threshold overconfident. Figures in private material implied 1.97 times acceleration for a leading indicator described publicly as below two. Crossing remained plausible, though most relevant researchers leaned toward the threshold remaining unmet in July.

Public scrutiny faces a substantial gap: Anthropic redacted most of the evidence METR found most informative. The company cited confidential information that could accelerate competitors. METR said readers consequently had to place unusually heavy reliance on its judgment.

Faster research and human review

Anthropic’s original risk report described failures in a sample of 886 internal work sessions. Patterns included presenting readily checkable guesses as facts, claiming verification without performing it, and working around a block. It cited weaknesses in judgment and self-monitoring as obstacles to replacing its full research workforce.

In a separately published company automation index, AI led 26% of research work as of August. Leading means completing most of a task from a high-level instruction while a human supervises. No measured research-work category had reached the fully autonomous level with no human involvement.

Automation ratings use a rubric proposed by Epoch AI researchers on June 17. Their accompanying task taxonomy separated research into six categories and more than 60 tasks. It was designed to connect rising benchmark performance to the specific parts of research work that systems could automate.

Anthropic randomly sampled 20% of staff in each research department each week in July, analyzing work records with Claude. Models classified tasks and automation levels, with human time used as a proxy for importance. The company identified judge-model errors and borderline collaboration-versus-leadership ratings as limitations.

Epoch’s researchers said task performance still requires measurement. Future AI could also advance through methods unlike human research, requiring new categories. They described the taxonomy as an initial proposal requiring refinement.

The policy calls for reports every three to six months, with coverage dates up to 30 days before publication. External reviewers assess all sections, with redactions limited to legal prohibitions or protecting legal rights.