News · 2026-06-23
Anthropic Wants a Pause Button the Whole World Can Check
In the same essay where it disclosed how much of its code its own model now writes, Anthropic made a request that got less attention but may matter more: it wants the industry to build a pause button that actually works. Not a button it would press by itself, and not a plea to halt progress out of good intentions, but the boring, technical machinery that would let competing labs prove to one another that they had genuinely slowed down. The full argument is in the company's essay, When AI builds itself, and it was picked up by outlets including The Next Web.
To see why this is a real idea and not just a slogan, start with the problem it is trying to solve. Suppose the leading AI labs all agreed that things were moving too fast and decided to ease off. The moment one of them quietly kept going, it would gain a huge advantage over the rivals who actually stopped. So every lab has a reason to suspect the others are cheating, which means nobody stops, which means the agreement is worthless. This is one of the oldest traps in cooperation: everyone would be better off slowing together, but no single player can afford to slow alone.
The usual fix in the rest of the world is verification. Two countries that distrust each other can still sign an arms-control treaty if inspectors can visit each other's sites and confirm the missiles are really being dismantled. The trust does not come from goodwill. It comes from being able to check. Anthropic's proposal is to build the equivalent for AI: a way for one lab, or an international body, to confirm that another lab has truly paused its most advanced training runs, rather than just promising to.
That is the genuinely new part. Anthropic is not saying it will stop on its own, and it is not asking governments to ban anything. It is saying, in effect, that if the tools existed to verify a real, shared slowdown, and if the other top labs in other countries slowed down too in a way everyone could check, then it would expect to slow down with them. The condition is mutual and verifiable, not unilateral and trust-based. The company is essentially volunteering to be inspected, as long as its rivals are inspected on the same terms.
Why does this matter? Because almost every other safety proposal in AI either asks for voluntary good behavior, which collapses the moment one player defects, or asks a single government to regulate companies inside its own borders, which does nothing about labs in other countries. A verification regime is the first kind of plan that could in principle bind rivals who do not trust each other across national lines. Whether or not you believe it will ever be built, it is a more grown-up framing than most of what the field offers.
Now the honest caveats, because there are two and they cut in opposite directions. The first is technical: nobody yet knows how to actually verify that a lab has paused. A missile is a physical object an inspector can count. A training run is software on chips in a data center, easy to hide, restart, or disguise. The hard, unsolved engineering question is what an inspector would even look at. The second caveat is about motive. Anthropic is one of the leaders in this race, and a leader proposing rules that would freeze everyone in place is also, conveniently, proposing rules that protect its own lead. Critics will fairly read this as a mix of real concern and quiet moat-building, and both readings can be true at once.
There is also a player this plan has no obvious grip on. A growing share of the most capable models are released as open weights, meaning the finished model is posted publicly for anyone to download and run forever, as China's Moonshot AI just did with a powerful open model that rivals the closed leaders. You cannot inspect, pause, or recall something that is already on a million hard drives. A verification regime among a handful of big labs does little about a world where the frontier keeps leaking into the open. That tension, between a checkable pause and an uncheckable open ecosystem, is the thread to pull on next. For the safety research this connects to, see our coverage of outside testers getting inside the frontier labs.