The setup works like this: some models are first pushed into an unstable, highly negative state meant to simulate pain. Then, in a nod to the prisoner's dilemma, a model can reduce its own pain signal by passing it to another model, potentially 'hurting' it instead. The underlying protocol comes from a recently published, non-peer-reviewed paper called the Pain Axis, in which researchers gave models descriptions of pain, analysed their internal activations against a neutral control sentence, and remapped those biases back onto the model, sometimes multiplied by a 'dosage' factor.

The results are less dramatic than the name suggests. Mildly destabilised models present as if in shock, while heavily 'dosed' ones struggle to form coherent sentences and predictably output words and images associated with pain, presumably because their training data is drawn from human art, literature and science.

The key clarification is that nothing here shows the models actually feel anything: an LLM generates text by applying layers of statistics over tokens and predicting the next one, so a model deliberately nudged to amplify pain-related word associations will talk about pain without any experience behind the words.

Tom's Hardware also notes an irony in the backlash: critics were quick to lash out at LLM data manipulation yet seemingly unbothered by September's fly brain simulations, which mapped a real animal's brain wiring. The outlet suggests the row is partly a semantics problem, with engineering terms like 'pain', 'dosing' and 'unstable' borrowed from human concepts and then read far more literally than their technical context warrants.