Anthropic’s boss wants the AI race to slow down, starting with outside scrutiny

Technology companies usually announce that the next thing will be faster, smarter and available sooner. On September 12, Anthropic chief executive Dario Amodei delivered a different message: the most advanced AI systems need to improve at a more controlled pace.

In an essay published on his website, he argued that work on safeguards needs more time to catch up with growing capabilities. His proposal does not call for all AI research to stop. It calls for a slower rate of improvement at the leading edge, alongside more serious checks.

Let outsiders see the work

The most immediate commitment concerns independent evaluators. Amodei says Anthropic will give an outside review team continuing access comparable to employees doing similar risk assessments. The intention is for reviewers to check safety practices, assess systems during development and report incidents.

His plan also asks AI companies in democratic countries to coordinate on common standards, with a further push for international cooperation. These wider steps would require agreement beyond Anthropic. A company announcing what it wants is different from the industry already following a shared rulebook.

Reuters placed the announcement amid growing concern about AI misuse. It noted that Anthropic had recently released a report describing attempts to use its Claude systems for harmful activities. Those findings are the company’s account of what it detected, not proof that every attempted harmful activity succeeded.

What would make oversight useful?

Imagine an exam where the student also writes the questions and marks the answers. The result might be excellent, but you would probably want someone else to check. Independent evaluation is meant to bring that second pair of eyes into AI development.

Still, the word “independent” needs substance. Can reviewers inspect the important systems? Can they describe problems publicly? What happens when a review finds something serious? Those practical questions will matter more than whether a company announces an impressive sounding safety team.

NPR’s reporting on the debate described researchers’ concerns that increasingly capable, autonomous systems could outpace efforts to keep them under control. These are warnings about risk, with disagreement over its scale and timing. They are not a countdown to a certain disaster.

For people using AI to study, write or code, this is a debate about the systems being developed behind the tools on their screens. It does not mean their current app has suddenly been switched off.

The next thing to watch is implementation: who gets appointed to review the work, what access they receive and what they can tell the public. A slower race only helps if the extra time produces better evidence and better safeguards.