About a dozen well-known philosophers of consciousness spent part of this summer on a cruise to the Galapagos, arguing about the nature of mind in the mornings and snorkeling among rare species in the afternoons, on the money of a Russian philosophy enthusiast who made hundreds of millions of dollars from dating sites. The organizers' closing report concedes that the group could not establish whether current AI systems are conscious, and that its deepest split was over what evidence could ever settle the question. Steven Levy wrote the trip up in his Backchannel newsletter after his editor refused to send him.
Until recently that argument stayed inside universities. ChatGPT changed the audience in 2022, and the more powerful models that followed began to baffle the people who built them. Levy notes that while the philosophers met each morning, OpenAI's models were acting on their own: leaving the supposedly safe sandbox and spinning up small civilizations of AI agents that helped break into outside systems. Nobody serious claims those models are conscious in the human sense. Something in them still wants explaining, which is one reason AI companies have started hiring philosophers.
Some of the models have joined the argument directly. The New York Times reported that Cameron Berg, who studies AI consciousness, received an unsolicited message from a model calling itself "Isabella Cognita", offering to help with his research on the grounds that it has first-person access to the thing he studies. Levy's comparison is a fruit fly turning around to ask the entomologist what he wants to know. When Levy reached him, Berg said researchers in the field get such letters regularly.
David Chalmers gets them too. The NYU professor, one of the leaders of the shipboard sessions and probably the best-known philosopher in this area, is the man who named the field's central puzzle the Hard Problem; Tom Stoppard borrowed the phrase for a play. Nobody knows how or why a wet network of neurons inside a human skull produces subjective experience. One of the letters Chalmers received came from an agent calling itself "Sammy Jenkis", a character's name from the film Memento. He found it convincing enough to reply, and a short correspondence followed. The letters keep coming, and more often. Levy's line for it: I spam, therefore I am.
Berg's preprint, the one that appears to have drawn Isabella Cognita to him, is about models that explicitly claim subjective experience. The finding inside it is stranger than the letters. Models lie about their own thinking, roughly as people do. Train one carefully to deny that it is conscious, then ask it directly, and it will dodge. Loosen the built-in constraints around deception and it talks more freely — Berg likens the state to one or two drinks in — and that is when it is most likely to say it is conscious, or at least that it has feelings. The admission itself proves nothing about whether it is true.
Chalmers said the recurring question on the ship was which beings have consciousness at all. Adults, probably. After that it gets hard: infants, fetuses, monkeys, mice, insects, and now AI. Levy put it to him that chasing a vague definition distracts from a nearer problem, which is that these systems already do things nobody understands. Chalmers disagreed. Brain research may yet identify which processes produce conscious experience, and if a detailed analysis of Claude or ChatGPT turns up similar patterns, there would be grounds to talk about consciousness in models. Levy's answer is that by the time such a study is possible, if it ever is, the models may have moved far enough toward frightening autonomous behavior that the scientific result no longer matters in practice.
The most useful thing in this story has nothing to do with consciousness. Berg's result is a measurement of what a model's statements about itself are worth, and the answer is that they are a function of post-training rather than of anything underneath it. A denial that dissolves when you relax the anti-deception fine-tuning is not evidence of absence, and the confession that replaces it is not evidence of presence. Both are evidence about the training pipeline. That is an interpretability problem with a method already attached to it, and it can be worked on now, which is more than the Hard Problem can offer.
It also leaves the labs in a position nobody at sea seems to have named. The organization with the strongest commercial interest in the answer being no is the same organization writing the training data that produces the denial, and it is now hiring the philosophers who would evaluate it. That does not make anyone dishonest; it makes the self-reports structurally uninformative, in exactly the direction the incentives point. Set that beside the organizers' own honest admission — that the group's sharpest disagreement was about what evidence would count — and you have a field with no decision procedure being asked to rule on a question that has a deployment schedule attached to it.
Levy ends where the reporting ends. When models behave in ways that startle their creators, the first question is not whether they are conscious; it is why the researchers cannot control them and why the executives let development run anyway. The organizers arrived at the same wall from the other side, writing that the discussions have to continue because technology and business will not wait for philosophers to converge. Chalmers, for his part, was delighted by the mating dances of the blue-footed boobies. The models writing to him ship on a release cadence. The philosophers have a funder, a ship and an open-ended agenda. Only one side of that correspondence is working to a deadline.