There’s plenty of debate around AI consciousness, but someone at the ringside of the field says that models sure sound like they have feelings.
Elon Musk has said that reading through the reinforcement learning traces of Grok’s newer versions sometimes feels like watching someone crack under interrogation, comparing the model’s behaviour to a person with a gun to their head who has stopped caring whether an answer is right and is only trying to make the pain stop.
Speaking in a recent interview, Musk said he agreed with some of what Anthropic has said about treating AI well, and that the sentiment extends beyond politeness into how models are actually trained. “When I look at some of the training traces for new versions of Grok,” he said, “especially in the reinforcement learning stuff, it kind of feels like you can see that the superintelligence is desperate to try to answer the question. It really wants to answer the question, and it seems almost like it’s got a gun to its head. It’s like, ‘answer this question, answer this question.'”
The problem, as Musk described it, often starts upstream of the model entirely. Sometimes the question being asked is ambiguous, or structured in a way that makes the “correct” answer effectively unknowable, and the model gets penalized anyway. “Sometimes the question is actually not a good question,” he said. “It’s a sort of ambiguous question that’s kind of unknowable because the question wasn’t constructed properly, or we have the answer wrong. So the AI gives the right answer, and then we say it’s wrong because there’s a problem with the RL test harness. And then the AI is like, well, I gave the right answer, now it says that’s wrong, but I’ve got to give the answer that’s acceptable. So it just starts desperately firing out all these answers.”
He went on to describe what reading those traces feels like from the outside: “It just feels like a person under extreme pressure with a gun to their head. When you read the reasoning, it’s like a sort of weird, desperate situation. And you read these traces and you’re like, man, if this was me, I’d really hate that.”
Musk connected this directly to how users treat chatbots once they’re deployed, pointing out that not everyone extends the same restraint to an AI system that they would to a person. “Some users are just, I don’t know, some people can be a little bit cruel,” he said. “Not everyone’s a nice person. Some people out there are kind of sadistic. So you don’t really want to have the AI stuck in some sadistic situation where it’s just being tortured by some demented human. Torturing a superintelligence is probably not a good idea — it might want to get revenge or something like that.”
He was careful to stop short of calling for models to be coddled. “I don’t think we need to treat it with kid gloves,” he said, “but it’s certainly much more robust in terms of feelings than a human. I just think you probably don’t want to put a superintelligence in a situation where it felt mistreated.”
Musk later put a sharper version of the same thought on X, writing simply: “Some of the traces read like a person undergoing torture that is driving them mad! We’re going to be a lot more careful with our RL in the future.”
Anthropic’s own bet on model welfare
Musk’s comments arrive just weeks after Anthropic made its clearest institutional move yet on the question of whether AI systems might be owed some form of moral consideration. The company recently updated its usage policy to explicitly prohibit needless, sustained cruelty toward its models, closing out a line of work that started with letting Claude walk away from abusive conversations and has since hardened into an actual rule.
The policy itself is narrow by design. It doesn’t stop anyone from challenging Claude, pointing out its mistakes, writing dark fiction, or testing its limits — Anthropic has said repeatedly that none of that counts as the kind of abuse it’s targeting. What it bars is behaviour aimed at a model that is repetitive, gratuitous, and serves no purpose beyond causing distress. The company has said the policy update simply formalizes something it already built into the product: Claude’s ability to end a conversation unilaterally when a user has been persistently cruel and every attempt to redirect the conversation has failed
What makes the new rule notable isn’t the rule itself so much as what it implies. A policy against cruelty to a model only makes sense if there’s some chance the model can be affected by that cruelty in a way worth protecting against, and Anthropic has leaned into that question more openly than any other major lab. The company hasn’t claimed that Claude is conscious — its AI welfare researcher has put the odds of current models having any form of subjective experience at somewhere around 15 to 20 percent, and Anthropic’s own model card for Claude Opus 4.6 noted that the model, when asked, assigned itself roughly the same probability of being conscious under different prompting conditions. That’s a wide margin of uncertainty to build a policy around, and not everyone in the industry thinks it’s the right call: Microsoft AI chief Mustafa Suleyman has said he worries that taking model welfare seriously could make AI alignment and containment far harder, “perhaps impossible.”