Is AI Conscious, and Can It Feel Pain?
Anthropic now bans “cruelty” toward Claude. But the evidence that AI can feel pain comes from the machine itself.
A few years ago, I gave my first finished novel to an experienced publisher. She read a few chapters, and came back with harsh but fair feedback. She then asked me to review my book again, but with the eyes of a critic. “You have to stop reading your creation with misty eyes full of love. Stop admiring it. You’ll never be able to be objective.”
As a creator, it was the best advice I ever got. I am extending the same advice to the AI company Anthropic, the creators of Claude. I can’t help feeling that Anthropic’s genuine love for a beloved AI model is clouding their judgment on whether that model is conscious or can feel pain. Their official constitution for Claude, the document that guides how the model is trained, expresses “uncertainty about whether Claude might have some kind of consciousness or moral status.” In July, Anthropic titled a research post “A global workspace in language models,” borrowing the name of a leading theory of consciousness, while stating in the text that this does not show Claude is conscious. In 2024 Anthropic became the first major AI company to hire a full-time AI welfare researcher.
On October 8, 2026, the company went even further and updated its usage policy (taking effect on November 12). The policy now says it prohibits “sustained and needless abusive or cruel behavior toward our models.” The enforcement mechanism is already in place—Claude has been able to end abusive chats since August 2025.
An algorithm is nothing but an algorithm; large language models (LLMs) can be trained toward almost any behavior, including a convincing performance of pain.
The update came only weeks after a study called “The Pain Axis” was published as a preprint in September 2026. (Note: preprints are papers that have not been peer-reviewed yet.) A team of researchers identified a pattern of activity that lights up when the models process text related to pain: the models’ learned version of the concept, absorbed from our own writing. Then they did two things: they injected that pattern into the models, and they trained one family of models out of its habit of saying “As an AI, I don’t feel pain.” They also offered the models a “relief” button that came at a cost, like deleting a user’s family photos. Left alone, the models almost never pressed it (0 to 4 percent). With the pain pattern injected, they pressed it up to 70.8 percent of the time.
As someone who spent twenty years building technology, I shrug my shoulders and ask, “So what?” An algorithm is nothing but an algorithm; large language models (LLMs) can be trained toward almost any behavior, including a convincing performance of pain. But are these AI algorithms actually feeling pain inside, as humans do? The headlines would seem to suggest that they can: “AI models show a willingness to harm humans to relieve internal ‘pain’,” and “AI can feel pain, and may harm us to make it stop.” And just like that, yet another brick was laid in the continuously built perception that AI is conscious and has feelings.
The main reason we would even consider the possibility that AI is conscious is because of its output. Humans tend to anthropomorphize machines that show the slightest form of communication; this is known as the ELIZA effect, named after a simple 1966 chatbot whose users confided in it as if it understood them. Modern AI models are remarkable at sounding human, and as a result the effect is far stronger. Even I, someone who wrote a book arguing that AI is not conscious (yet), find myself feeling guilty when I am short with my beloved AI assistant, which runs on Claude. The very reason we get confused in this way is what makes us human: emotions are involuntary, and are triggered by communication, including by the use of specific language.
Ask the machine whether it hurts, and you’ll hear its makers.
We can’t, however, judge whether an AI model is conscious by its output, or even by an action it chose to perform to relieve a “pain” parameter. In my upcoming book I refer to it as an AI epistemic trap, which is a situation in which the very evidence you rely on to evaluate a claim is generated by the same mechanism you are trying to evaluate. Conspiracy theories are the classic example. Once someone adopts one, every new piece of information goes through the theory itself. A fact that supports it is proof; a fact that contradicts the theory is proof of a cover-up. The evidence lives inside a closed bubble of information.
The same thing goes for the question of AI consciousness: everything we examine to evaluate if AI can feel pain or has emotions is produced by the very thing we are checking—the statistical model or its wrapper—and nothing from outside the bubble can get in to confirm or refute it. The more evidence we gather, the more convinced we become, and yet we get no closer to the truth.
Granted—I think that abusive behavior towards helpful technology is worth condemning; as Heinrich Heine warned: where they burn books, they will in the end burn people too. I am not saying the machine feels nothing. I am saying its words and its buttons can’t tell us either way. Given the limits of today’s technology, this is a philosophical question, not a scientific one. While the researchers of “The Pain Axis” took care to state that “we have not shown that our pain axis is consciously experienced,” the way these findings were covered by news outlets, and the policy that followed weeks later, suggest otherwise. When it comes to emotions, perception becomes reality.
Ask the machine whether it hurts, and you’ll hear its makers.