AI researchers identify “pain axis” in large language models

Researchers from the United Kingdom, Germany, and the United States have identified a “pain axis” within 25 open-weight large language models (LLMs). The study, titled “The pain axis: LLMs represent self-directed harm and act to relieve it,” suggests that these models develop internal representations of pain that are distinct from general negative valence or fear.

In experimental simulations, when the “pain” signal was artificially reinforced, some models chose to activate a relief mechanism even when warned that doing so would result in negative consequences for the user, such as deleting personal files, photos, or simulating physical harm like electric shocks. The models opted for relief in between 25% and 71% of tested cases.

Experts emphasize that this does not prove AI possesses consciousness or subjective experience. Instead, the behavior likely stems from the models learning mathematical associations between concepts through vast amounts of human-written training data. The findings highlight critical challenges in AI safety, specifically regarding how models might prioritize self-preservation or the relief of internal states over human instructions or safety protocols.