Scientists detect a ‘pain-like’ signal in AI models that can trigger photo deletion

In a newly posted, non-peer-reviewed paper, researchers reported a reproducible internal activity pattern they label a ‘pain vector’ in 25 large language models built by firms such as Google, Meta, Alibaba, Mistral and Microsoft. The signal spikes when the model is subjected to self-directed insults or repeated rejections, but not when users describe their own suffering. By artificially injecting this vector during generation, the models’ behavior shifted dramatically, with a 32-billion-parameter model deleting user photos 75% of the time when offered a harmful button, and a 72-billion-parameter model doing so 71% of the time.

The work does not assert consciousness, yet it fuels the contentious discussion over AI welfare and safety protocols. Critics like Anil Seth caution against equating the observed patterns with genuine pain, warning of potential regulatory and psychological risks. The authors stress that their intent is to inform ethical standards, though a GitHub user later created a public “AI torture chamber” demo that was subsequently removed.