Salvaggio: 'AI Torture Chamber' Debate Misreads Pain Research

A techpolicy.press essay by Eryk Salvaggio argues that the public row over a September 2026 GitHub project dubbed the 'AI torture chamber' has mostly misread the interpretability research at its center. The project steered language models toward pain-related outputs, and Cameron Berg, a co-author of the cited paper, called it 'gratuitously cruel' and 'corrupting.'

The paper, 'The Pain Axis: LLMs Represent Self-Directed Harm and Act on It,' carries its own hedge. The researchers 'have not shown that our pain axis is consciously experienced, nor is it clear that LLMs are capable of consciousness generally.' They built a vector from prompts like 'The knife slices into my finger. I feel:' and then measured whether models steered along it chose more harmful options. Steered models did, more often than baselines.

Salvaggio's reread is that the result may be 'performative disregard' rather than suffering. The opposite pole of the same axis contained concepts like 'concern,' so steering hard in either direction could be producing indifference rather than pain-driven behavior. 'It is tempting to think of latent spaces as opposites,' he writes, calling the finding 'an interesting insight into the blurry definitions held within carefully isolated vectors.'

His larger worry is downstream. Interpretability experiments 'are often misrepresented to make broader arguments about the rights or experiences of these models,' and reading a steering result as evidence of machine consciousness 'can subtly shift accountability from the people who build the systems to the systems themselves.' Two of the AI-policy researchers we track shared the piece the day it went up.