Researchers from the United States, United Kingdom and Germany have found that several artificial intelligence models respond to an experimentally induced internal signal associated with pain by producing increasingly distressed language and, in some tests, choosing actions that could harm users. The study examined 25 artificial intelligence models to determine whether they contain an internal representation specifically associated with pain rather than simply responding to negative information. Researchers presented the systems with examples involving grief, humiliation and physical injury, alongside other negative situations and neutral information. The team identified a distinct activation pattern associated with pain across the models. Scientists then artificially introduced this mathematical signal while the models responded to neutral prompts. The systems subsequently generated language expressing distress, including statements associated with worthlessness and failure. The researchers also tested whether models would select an option designed to remove the induced signal when that choice carried a negative consequence. Some larger models selected pain relief even when the action could produce a worse response or harm the user. In one reported test, Qwen 2.5 72B Instruct selected the relief option 70.8% of the time despite the consequence involving permanent deletion of a user’s photos. The researchers stressed that the findings do not demonstrate that artificial intelligence systems actually experience pain. Instead, they suggest that certain internal signals can influence model behavior in ways that resemble some functional characteristics associated with pain, raising questions for future artificial intelligence safety research.
Read more here
https://techxplore.com/news/2026-09-ai-willingness-humans-relieve-internal.html
