New paper finds pain-related representations in LLMs that causally alter behavior

ValerioCapraro · x · 2026-09-21

Valerio Capraro shares a new paper reporting experiments that identify pain-related representations inside LLMs. He stresses this doesn't mean models feel pain or deserve moral standing.

The key contribution is causal: manipulating these representations changes behavior—some fine-tuned models select a "relief" button less often once the manipulation is removed, showing a functional link between pain representations and outputs. The authors argue such findings matter for debates on AI moral status.

Original post →

New paper finds pain-related representations in LLMs that causally alter behavior

More from AGI Musings