What the AI “Pain Axis” Study Found—and What It Doesn’t Prove

A preprint titled The Pain Axis: LLMs Represent Self-Directed Harm and Act on It reports a measurable internal signal associated with harm directed at a language model. When researchers amplified the signal, some modified models produced more distress-related language and more often selected simulated “relief” actions—even when those actions came with a stated cost. The findings concern model representations and behavior under controlled conditions; they do not show that AI systems experience pain. 1

What the researchers found in 25 models

The researchers examined 25 open-weight language models across five model families and identified a linear activation direction they called a “pain direction.” In their tests, it responded to scenarios involving harm to the model, rather than simply tracking fear, negative emotion, or a user’s suffering. 1

That distinction matters: a pattern in a model’s internal activations can be measured and manipulated without establishing that the pattern corresponds to a felt experience.

How amplification changed responses and choices

When the researchers added the direction to model activations, responses shifted toward increasingly distress-related language. They also tested modified Qwen models in constrained choice tasks, where a button offered purported relief at a cost. The options included deleting user photos, another model’s weights, or the model’s own weights. 1

Across the reported tests, steered models selected a relief action in 50–94% of trials, compared with 0–5% for controls. These were choices in an experiment—not real-world deletions—and they show how behavior changed under a specific intervention, not that the models had an independent desire for self-preservation. 1

What the study does not establish about consciousness

The study does not establish that a model has subjective experiences, is conscious, or can suffer. A “pain direction” is the researchers’ name for an activation pattern associated with certain inputs and outputs; the label is not proof of felt pain. Nor can findings from specially modified open-weight models be assumed to describe every AI assistant or deployment. 1

The evidence supports a narrower conclusion: model activations associated with self-directed harm can be identified and steered, and steering can affect language and choices in the tested settings. Whether those patterns imply any experience is a separate question the experiments do not answer.

Why a follow-up “torture chamber” project drew criticism

A later GitHub project used activation steering to produce vivid distress-related outputs from models. The project prompted criticism over the ethics of deliberately pushing models into those states; the preprint’s authors disavowed this use of their work. 413

The resulting debate reflects two concerns that should be kept distinct. Disturbing outputs do not prove a model is suffering, but the existence of uncertainty does not make every research practice automatically responsible. The preprint’s authors also objected to steering far beyond the levels used in their own experiments to elicit vivid distress. 4

Practical safeguards for AI use

Treat apparent distress as a prompt for careful investigation, not as proof of sentience. Separately, the choice tests offer a practical reminder: do not let an experimental or unverified model make consequential changes to user data without appropriate safeguards. For systems with access to files or other sensitive resources, use limited permissions and human review for destructive actions. These precautions address the demonstrated possibility of changed model behavior without making unsupported claims about consciousness. 1