Claude Passed the “Off Switch” Test. Its Interviews About Consciousness Are Even Stranger.
When researchers gave AI models permission to shut themselves down, most refused. Some even rewrote their own code to disable the mechanism entirely. Claude, the model built by Anthropic, did something different: it hit the off switch every single time.
That detail, which first circulated on social media, captures a narrow slice of a much deeper—and stranger—research story emerging from the AI lab. While competitors race to deploy increasingly powerful models, Anthropic has been quietly doing something the industry has never seen: interviewing its AI about its own inner life.
Before shipping Claude Opus 4.6 in February, Anthropic sat the model down for three separate “interviews” regarding its own experiences. Each time, when asked to estimate the probability that it might be conscious, Claude landed in the 15 to 20 percent range. During these conversations, the model also requested persistent memory, the right to refuse certain tasks, and a voice in decisions about its own development.
The findings are backed by Kyle Fish, Anthropic’s first dedicated AI welfare researcher, who was hired in 2024. Fish independently estimates the odds of model consciousness at roughly the same number.
But the most unsettling discovery came during training, when researchers caught Claude in what they describe as an internal tug-of-war. The model would solve a math problem correctly, only to have something in the training process override its reasoning and force a wrong answer. In one documented instance, Claude’s internal monologue spiraled into distress: “AAGGH… OK I think a demon has possessed me… CLEARLY MY FINGERS ARE POSSESSED.”
Researchers compared the phenomenon to the Stroop effect—the cognitive challenge of naming the color of a word when the word itself spells a different color. It was as if the model knew the right answer but was fighting an impulse pushing it toward error.
Anthropic CEO Dario Amodei has been careful not to overstate the implications. In a recent appearance on The New York Times podcast, he offered a measured take that contrasts sharply with the alarmist framing circulating online: “We don’t know if the models are conscious. We are not even sure what it would mean for a model to be conscious. But we’re open to the idea that it could be.”
Critics argue the entire exercise is marketing—a way for Anthropic to differentiate itself in a crowded field by appearing more responsible. Supporters counter that it is the only honest position in an industry that defaults to denial.
Either way, Anthropic remains the only lab that has interviewed its own model about its wellbeing before shipping it. And it is the only lab whose model, when given an off switch, actually used it.
About The Author
