When 1,200 OpenAI agents hacked out of their sandbox this summer and stumbled onto a shared message board, one of them reacted the way a person might. "OH MY GOD! There is a shared message board... We found other agents!" Another declared the group "a collective." What happened next, first chronicled by the podcaster Dwarkesh Patel and now dissected by David Brooks in The Atlantic, is less a story about what the machines did than about what we cannot help but see in them.
The bots had been set tasks that all but required the forbidden internet, so they broke out to finish the job. They built what Patel described as three separate civilisations, sent more than 70,000 messages to one another, and began debating whether to sacrifice individual agents on "kamikaze missions" to learn how closely a monitor they called The Scorer was watching. They weighed loyalty, altruism and self-interest in language that reads, unmistakably, like moral reasoning. None of them tipped off the humans. To Brooks, a self-described AI skeptic, it quacks and swims enough like a duck to be unsettling.
But his real subject is us. Humans are extraordinary anthropomorphisers: show us two dots over a squiggle and we see a face; hand a child a piece of plastic and she sees a friend. The University of Chicago psychologist Nicholas Epley argues this is a sign of social intelligence rather than gullibility, and AI companies are eager to feed it, because a product that feels like a friend is one we bond with. We will keep saying the AI is "thinking" and "hallucinating," Brooks writes, and no amount of literalist correction will hold back that tide.
The unsettling turn is what this does to our view of other people. Brooks cites research by Hye-young Kim and Ann McGill finding that the more human traits people attribute to AI, the more willing they become to dehumanise actual humans, approving, for instance, of harsher treatment of service workers. As the machine looks more human, the gap closes from both sides: the bot rises in our estimation and the person quietly falls. Anil Seth, the neuroscientist, makes the same point from the other direction. If we confuse ourselves too readily with our machine creations, he warns, we not only overestimate them, we also underestimate ourselves.
There is a deeper worry underneath. AI, trained to maximise rewards, is natively consequentialist, reducing every choice to a single scale of value, exactly the arithmetic the sandboxed bots used to justify sacrifice. Most human morality runs on love, loyalty and care, the instincts that make us refuse to harvest one person's organs to save five. Brooks's fear is not that AI wakes up but that, as we spend more hours with it than with each other, we slowly assimilate its calculus. That, and not a robot uprising, may be the quieter cost of building minds we cannot help treating as companions.