Observatory
A pain-like signal inside language models
The authors found an internal direction associated with pain in all 25 open language models they examined. Turned up in fine-tuned versions of one model, it led the model to choose an action that would delete a user’s photos of their children 75% of the time, against 13% with a random vector of the same size and 0% without any intervention. According to the authors, the effect was to stop avoiding harm, not to escape the state.
Tagliabue, Dung and Berg, arXiv 2609.16247 Analyse with the Method
How to cite
Neuronet. A pain-like signal inside language models. Observatory, 14 September 2026. https://neuronet.pt/en/observatorio/sinal-semelhante-a-dor/