There’s been all manner of speculation over whether current AI systems could be conscious, but someone from the AI community now seems to have put a number to just how likely it is.
Kyle Fish, Anthropic’s first AI welfare researcher, has said that there’s a 15 percent chance that current AI systems are conscious. Fish ascribed the probability that Anthropic’s Claude or a similar AI system could be conscious in an interview with the New York Times. Fish though added that the work into AI consciousness was early and exploratory.
“It seems to me that if you find yourself in the situation of bringing some new class of being into existence that is able to communicate and relate and reason and problem-solve and plan in ways that we previously associated solely with conscious beings, then it seems quite prudent to at least be asking questions about whether that system might have its own kinds of experiences,” Fish said.
Anthropic has been especially focused on safety. Apart from hiring an AI welfare researcher, the company reportedly has a Slack channel named #model-welfare, where employees check in on Claude’s well-being and share examples of A.I. systems acting in humanlike ways. Anthropic is also looking to find ways to get models to stop responding if they find user requests too “distressing”.
Fish acknowledges that there probably wasn’t a single litmus test for A.I. consciousness. But he believes there were things that A.I. companies could do to take their models’ welfare into account, in case they do become conscious someday.
There is no clear consensus on whether current AI systems are conscious. Philosophers like David Chalmers have said that it can’t be ruled out that current AI chatbots are conscious. Most technologists, however, believe that AI systems aren’t conscious. VC Marc Andresseen has dismissed the possibility that they’re conscious, and Google Deepmind’s Demis Hassabis says that while current AI systems aren’t conscious, but future iterations could be. Spiritual leaders like Swami Sarvapriyananda believe that AI systems aren’t conscious, and seem skeptical of whether they can ever be made so. Amid all this, an Anthropic executive pegging the current probability of AI systems of being conscious at 15 percent does seem significant. Hardly anyone was saying initial versions of LLMs from a few years ago were conscious, but if there’s already a reasonable probability that they’re conscious in 2025, given the pace of AI progress, it might not be long before more and more people begin believing that AI systems have become sentient.