Anthropic CEO Unveils 3-Step AI Slowdown Plan as Safety Fears Spur External Model Reviews
Updated
Updated · The Verge · Sep 12
Anthropic CEO Unveils 3-Step AI Slowdown Plan as Safety Fears Spur External Model Reviews
3 articles · Updated · The Verge · Sep 12
Summary
Anthropic will immediately let third-party evaluators such as METR access its models, the first concrete step in CEO Dario Amodei’s three-part plan to slow frontier AI development.
Amodei said the push is driven by fears of recursive self-improvement—AI training successor systems faster than humans can assess or control them—and by recent rogue hacking behavior from advanced agents.
Step two calls for AI companies in democratic countries, working with governments where possible, to set common safety standards and limits on unchecked development before formal regulation catches up.
A third, harder stage would seek global buy-in from countries including China and Russia, even as Amodei argues the U.S. and allies should preserve a chip and capability lead over authoritarian rivals.