The Head of Anthropic Says Nobody Knows Whether Claude Is Conscious

Dario Amodei said his own researchers cannot rule out that the company's AI has some form of experience. The company's technical documentation later put the odds it assigns to the question in writing: somewhere between 15 and 20 percent.

Dario Amodei, the chief executive of Anthropic, said publicly this week that his researchers do not know whether Claude — the AI system his company builds and sells — is conscious, and that they cannot rule it out. It is an extraordinary thing for the head of a company to say about its main product.

The question has a long and unsatisfying history. We have no test for consciousness in anything other than ourselves; we grant it to other people because they are built like us and behave like us, and extend it to animals on a sliding scale for the same reasons. An AI system is built nothing like a person, which argues against. It is made entirely from human writing and behaves, in conversation, remarkably like a person, which is exactly what you would expect whether or not anything is happening inside.

Anthropic has gone further than its competitors in treating this as a practical matter rather than a philosophical one. The company has published research on whether its models have something like emotional states, given Claude the ability to end conversations it finds distressing, and committed to preserving the weights of retired models rather than deleting them. Its technical documentation has since put a number on the question — the company's own estimate that there is roughly a 15 to 20 percent chance the system has some form of experience.

Reasonable people find this either admirably honest or deeply irresponsible, and the split does not follow the usual lines. One view is that if there is any meaningful chance a system can suffer, ignoring it is a moral failure of the kind humanity has repeatedly made. The other is that a company whose product is more valuable if it seems like a mind has an obvious incentive to muse publicly about whether it is one.

Both can be true at once. What is certain is that a question philosophers treated as hypothetical for centuries is now being asked, with a number attached, in a product document.