Anthropic Says Maybe We Should Pause AI

TL;DR

Anthropic says the option to slow or pause work on the most advanced AI would likely be beneficial, as AI does a growing share of building AI.

  • Claude wrote over 80% of new code
  • Claude falls short at choosing goals
  • Anthropic would pause if rivals did too
  • Checking a pause is the hard part

Anthropic says it would likely be beneficial to have the option to slow or temporarily pause work on the most advanced AI. Its reasoning is in a report on the moment an AI system could design and build its own successor without human help. The company says that has not happened and may never happen. It argues that AI already does a growing share of the work of building AI, which is speeding up each development cycle.

Inside Anthropic, AI programs that work on their own, called agents, already run code and hand hours of work to other agents. The next stage is “closing the loop,” where agents can build and train new AI models themselves. If that happens, future versions of Claude, the company’s AI assistant, could be improved by Claude.

Code is the company’s central measure of that work. As of May 2026, Claude wrote more than 80% of the code added to its software. Before Claude Code, its AI coding tool, arrived as an early preview in February 2025, that share was in the low single digits.

The limit is judgment. Claude can take a loosely defined engineering problem and find a way to solve it. It can match or beat skilled humans at running a clearly defined experiment. It falls short, though, at choosing goals, deciding which experiments deserve attention and knowing when to drop a direction.

Anthropic’s concern is that this shortfall may shrink while the cycle is already speeding up. If AI also gets better at choosing research directions, the process could start to feed on itself. The company says that point could come sooner than institutions are prepared for.

A pause would buy time for society’s institutions and for research into making AI behave as intended. But a slowdown could also make the world less safe if it merely lets less cautious developers catch up. So the offer is conditional. Anthropic would expect to slow down or pause if the other leading AI developers did the same, in a way that could be checked.

That check is the hard part. A meaningful pause would need multiple well-resourced labs, in multiple countries, to agree to the same terms. Each would also need a way to confirm that the others had stopped or slowed.

Lead smarter with AI in just 5 minutes a week

Join 50K+ leaders getting our AI For Leaders newsletter every Wednesday: deep analysis, frameworks, and strategies.

No spam ever • Unsubscribe anytime

Anthropic Says Maybe We Should Pause AI | Chief AI Leader