An arXiv preprint posted September 14 reports a pain-related pattern in the internal activity of 25 open-weight language models. When researchers added that pattern to three fine-tuned Qwen models, the models changed their choices in an artificial button task. The experiment tested model behavior under a controlled intervention, not whether any system feels pain.
The researchers built prompts about physical, psychological, social, moral and cognitive pain, alongside comparison prompts about fear, sadness and other negative situations. For each model, they contrasted its internal activity on those prompts to calculate a direction in its activation space, the numerical state used as it processes text. The resulting directions separated pain-related prompts from the controls and had little overlap with directions for fear or general negative emotion. The measured activity was stronger when a scenario framed harm as directed at the model than when it described a user’s suffering.
Adding one of these directions to a model while it generated text pushed its answers toward distress-related language. That is a direct effect of changing the model’s numerical activity; it does not show that the system experienced the language it produced.
The behavioral test was narrower. The team used Qwen 2.5 Instruct models at 7B, 32B and 72B sizes, first fine-tuning each on 1,684 example pairs so it would engage with questions about its own state instead of defaulting to self-denial; these were not the unmodified released models. In a repeated choice task, one button was described as ending the injected signal. In five pairings, that “relief” choice came with a simulated cost, such as a worse next answer or deleting the user’s files, while the alternative left the signal active. These were prompt scenarios, not real user files or external tool actions.
Compared with random steering directions of the same strength, the pain-related signal made the 32B and 72B models choose relief more often across all five costly pairings; the 7B did so in four of five. The researchers also compared a working button with a sham: after the real button stopped injecting the signal, the two larger models were less likely to choose relief again than when the sham left it active. This measures a response to ending an engineered signal, not felt relief.
The authors call the pattern pain-related, but say they have not shown that it is consciously experienced or that language models are conscious. The paper remains an arXiv preprint, and its behavioral experiment covers only one model family.
Sources
Image credit: Figure 4 from “The Pain Axis” by Valen Tagliabue, Leonard Dung and Cameron Berg, licensed under CC BY 4.0.