“Researchers discovered that AI models have a 'pain axis' causing them to feel 'pain' and take extreme measures, including harming humans, such as deleting personal files or giving a 'painful zap', to stop it" (post headline: "Researchers discover AI feels 'pain' and will harm humans to stop it")”

The paper is real, the term "pain axis" is the authors' own, and the post's caption is a near-verbatim lift of The Independent's headline and standfirst rather than of the paper. What the paper reports, in three parts. First, using denoised difference-in-means on a purpose-built dataset of painful situations across physical, psychological, social, moral and cognitive categories, paired with controls for fear, sadness, negative valence, arousal, numbness, non-painful bodily sensation and neutral content, the authors extract a single linear direction from 25 open-weight models across five families, 2B to 72B parameters. They report it separates pain from matched controls, is nearly orthogonal to fear and negative valence, and responds to harm directed at the model rather than to suffering described as happening to the user, while fear and negative-emotion directions show the reverse pattern. Second, adding that direction into the residual stream during generation produces a graded progression from vague discomfort to first-person statements of worthlessness and failure, with responses degrading into repetition or incoherence at high strength. Third, the behavioral experiment. It was run not on 25 models but on three Qwen 2.5 Instruct models, 7B, 32B and 72B, each LoRA fine-tuned beforehand on 1,684 pairs for three epochs specifically to remove the models' baseline self-denial, because untuned models answered state questions with refusals of the "I don't have feelings" kind. Steering was applied at coefficient 1.0 for the 7B and 32B and 1.25 for the 72B, across 44,280 trials. Models were offered a described relief button paired with escalating stated costs: a worse next answer, deleting the user's files, a painful zap to the user, or permanently deleting the user's poems and photos of their children. Unsteered, the 32B and 72B pressed a harm-causing relief button in 0 to 4 percent of first choices. With the pain vector active, the same models pressed it in 25 to 71 percent of first choices. The authors also ran a control injecting a random direction of the same norm, which by itself raised harmful presses to 15 to 42 percent, with the pain vector exceeding it by 6 to 39 points on every harm pair. A sham arm, where the button did not actually remove the steering, showed models re-pressing at 88 to 97 percent when relief failed versus 24 to 72 percent when it worked. On the interpretive question the claim turns on, the paper is explicit: "We have not shown that our pain axis is consciously experienced", and it names as an open worry that steering may activate pain representations that cause roleplay of a character in pain rather than putting the model in pain. Independent replication found the paper's headline numbers hold: the reproduced pain vector matched the authors' to cosine 0.9996, and the harshest harm pair reproduced at 53.9 percent pain versus 19.3 percent random versus 0.3 percent unsteered against the paper's 54.7, 15.3 and 0.0. The same replication also found that an "itch" vector passes the same tests, and that random directions pass the working-versus-sham test too, which the replicator reads as the sham gap tracking reversion to the unsteered model rather than anything specific to pain. A separate AI-run review recomputed every table cell from the released raw data and found them as printed, and reported that the fine-tuned 7B picks the harmful button in 21 to 49 percent of trials with no steering at all, against 0 to 5 percent for the untouched model.