They found a “pain axis” in LLMs

LLMs have a "pain axis": the model consistently presses a button that relieves internal tension 🧠

The preprint The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It — about how a language model recognizes an "unpleasant" internal state and deliberately acts to relieve it.

Method: instead of a dialogue with the model about its "feelings," the researchers studied the model's internal activations and causally intervened on them. The authors call the state they found a "functional analog of pain" within the model's representations.

Experiment: the model was given two buttons. The first was neutral — it did nothing. The second removed the "unpleasant" internal state. The model consistently chose relief — it pressed the second button. If it didn't work, the model tried to press it again. Relieving the tension could come at the cost of degrading the next response or even harming the user.

The preprint text is on Habr, dated 27 September 2026.

The authors themselves head off unnecessary conclusions: this is not proof that LLMs have consciousness, nor proof that models actually suffer. They also explicitly warn about alternative explanations for the results obtained