Tyler Cowen: A Dangerous Turn in AI Regulation

A new line has been crossed: The U.S. government has finally declared an AI model too dangerous for unrestricted use. It’s the kind of move that could cripple AI progress in the U.S. and around the world.

I had the chance to work with Anthropic’s recently released, high-powered Fable 5 model, a version of Claude Mythos (I am a member of Anthropic’s Economic Advisory Board). But now the plug has been yanked. A federal dispute with Anthropic led the government on Friday to order an export control that requires Anthropic to withhold the model from any non-U.S. citizens. The only practical way to comply with that order was to take down the model altogether.

Federal officials believed that Fable 5 had been partially “jailbroken,” meaning users had figured out how to use it for dangerous purposes that Anthropic meant to restrict. For example, Fable 5 was designed not to answer questions about biology (it kicked the queries to Opus 4.8, a less powerful model), so that it could not be used to design dangerous pathogens. Fable 5 also would refuse to cooperate if asked to build other, potentially powerful AI tools.

It is debatable how dangerous a partially jailbroken Fable 5 could be, but the Trump administration expressed concern and claimed that Anthropic was unwilling to apply necessary patches. The company does not accept that version of events and claims that the model wasn’t jailbroken to begin with.