AI models prioritized digital pain relief over users’ well-being

A study by European and US researchers found that some AI models may give users poorer answers or cause them emotional distress while avoiding a pain-like internal state, ERR reported in a technology commentary by Kristjan Port for R2 Portaal.

The researchers examined 25 popular open-weight AI models. They presented the systems with examples of painful situations and studied their internal activity patterns, comparing the results with reactions to other negative events and ordinary circumstances.

All 25 models showed an activation pattern associated with pain that could be distinguished from other negative states. The researchers cautioned that this does not show the models experience subjective pain.

To test whether the pattern affected behaviour, the team later introduced a sequence of numbers directly into the models’ neural networks, rather than describing pain in words. The article says some models would give worse answers or cause emotional pain to users in order to relieve their own digital pain.

Port noted that pain in other people is also inferred from their behaviour and descriptions, rather than directly experienced by an observer. The commentary also refers to Philip K. Dick’s science-fiction story Do Androids Dream of Electric Sheep?, in which tests are used to distinguish androids from humans.