Anthropic CEO Dario Amodei has proposed slowing capability improvements sufficiently for safeguards to catch up. His framework calls for permanently embedded independent evaluators, safety coordination among frontier laboratories and international cooperation.
What changed
A major AI laboratory is explicitly proposing capability pacing, while OpenAI and xAI leaders have expressed support for parts of the plan.
Why it matters
Independent evaluators would need access to internal tools, incident data and deployment decisions—not merely published benchmark reports.
Practical takeaway
Separate the team developing an agent from the team authorizing production deployment, and preserve signed evaluation evidence for every model version.
Analysis
No binding agreement exists, and competitive pressure may make voluntary coordination difficult.
Published
12 September 2026
