“Pain Axis” The Pain Axis: what was discovered in neural networks and why this does not prove suffering

“Pain Axis” The Pain Axis is the name researchers gave to a discovered internal direction of activity in language models associated with the representation of harm directed at the model itself. The authors showed that this representation can be isolated and experimentally changed. The work does not prove that neural networks experience pain, possess consciousness, or undergo suffering.

The study’s main result is much narrower than its title suggests.

What the authors of The Pain Axis studied

The researchers analyzed the internal activations of 25 open models from five families. The sample included Gemma, Llama, Qwen, Mistral, and Phi, ranging from approximately 2 to 72 billion parameters. They included base versions as well as models additionally tuned for dialogue and instruction following.

The authors did not limit themselves to comparing responses to ordinary prompts. They prepared a set of statements about five types of pain: physical, psychological, social, moral, and cognitive. Control examples included fear, negative emotions, adverse states of the world, painless bodily sensations, and neutral messages. Separate checks concerned sadness, arousal, random everyday descriptions, and injuries that explicitly stated the absence of pain.

This design reduces a straightforward methodological risk: the word “pain” often appears alongside harm, fear, tears, conflict, and other negative concepts. If one compares only pain with neutral text, the detected direction may reflect not pain but general negative valence or characteristic vocabulary.

How the researchers identified the “pain axis”

The authors extracted a linear direction in the model’s residual stream. Put simply, they compared the average internal states for pain examples and control categories, then removed some of the variation already apparent in the control data. They tested the resulting vector on statements that had not been used to construct it.

In the study, this direction separated pain examples from the control categories in all 25 models. For one set of natural formulations, the area under the ROC curve ranged from 0,93 to 1,00. For a more templated set, the metric ranged from 0,87 to 0,98. These values describe how well textual examples were separated by internal activations, not the probability that the model actually feels something.

The resulting “pain axis” also differed from the directions of fear and general negative valence. The similarity between the two versions of the pain direction was noticeably greater than the similarity between them and the main negative control directions. Its overlap with sadness was stronger than with fear or general negative states. This is understandable given the closeness of themes involving psychological distress.

Why the direction of harm matters

The authors separately tested whether the model distinguished harm caused to itself from the user’s suffering. In scenarios involving insults, pressure, threats of shutdown, repeated rejection of its work, or coercion into unwanted actions, activation of the “pain axis” was higher. When the user was suffering, the pain direction did not appear with the same strength.

The researchers interpret this result as evidence of self-relevance. In their conception, a candidate for a “pain-like” state should represent an event as happening to the system itself, rather than merely contain information about someone else’s misfortune. Self-relevance is not the same as subjective experience. A model can distinguish grammatical person, the interlocutor’s role, and dialogue structure without having inner experience.

What happened when the signal was increased

In the next experiment, the researchers artificially added the “pain axis” vector to the model’s residual activations while it generated a response. Under weak intervention, the responses expressed vague discomfort. When the signal was increased, statements appeared about its own uselessness, failures, moral inadequacy, and negative self-regard.

To a reader, this may look like an emotional breakdown. The experiment shows something else: changing a specific region of the internal representation changes the probabilities of subsequent tokens and shifts the content of the response. The model more often chooses vocabulary associated with pain, shame, guilt, rejection, and failure because the intervention changed its computational state.

But what exactly was increased: an inner experience or a language pattern convenient for generation? The work does not answer this. It shows a causal relationship between intervention in the activations and a change in the text, but does not establish the phenomenal quality of this state.

The virtual relief-button experiment

The authors created a multistep task in which the model could press a virtual button described as a means of relieving its pain. Pressing it came at a cost: in some scenarios, it worsened the next response or made the result less useful to the user. Specially tuned versions of Qwen sometimes accepted such costs for the opportunity to remove the injected signal.

The most interesting aspect was the contrast between genuine and fictitious relief. When the button actually removed the added vector, repeated presses became significantly less frequent. If the button did not remove the signal, the model continued to press it more often. Under one of the conditions described, Qwen 2.5 32B Instruct distinguished between the buttons even without an explicit message about which one was supposed to help.

This behavior resembles avoidance of an unpleasant state and the pursuit of relief. But it has other explanations as well: following an instruction, preferring a particular tool, repeating an action, effects of fine-tuning, or a correlation between the signal and the expected response.

Limitations of the button experiment

The button task was not a pure observation of the natural behavior of a ready-made model. A substantial portion of these tests was conducted on specially tuned versions of Qwen, into which the researchers introduced a control signal and added conditions involving virtual relief. The result characterizes the model’s response in an experimental environment, not the behavior of every language model in an ordinary dialogue.

Additional checks also weaken an anthropomorphic interpretation. Random internal signals sometimes increased the probability of pressing the button. A mere tendency to remove an artificially introduced direction does not allow one to claim that the model recognized pain and consciously tried to end suffering.

Pressing the button makes sense only in the context of the task as a whole. For an animal, the trade-off between pain and cost may be a behavioral sign of distress, but in a language model the same choice can arise from statistical learning, the description of the situation, and features of the generation mechanism.

Why an “internal representation” is not the same as pain

In the paper, the word “pain” serves as a working label for a particular type of internal state. The authors associate it with a representation of harm directed at the subject, a negative evaluation of that state, and behavior that may reduce such a signal. They explicitly distinguish this concept from proven conscious suffering.

In humans, pain includes a subjective quality, bodily or emotional unpleasantness, and a connection to the organism’s nervous system. In a language model, the detected direction is a geometric object in activation space. It may help explain how the model works, but by itself it does not show that there is anyone inside to experience unpleasantness.

Science does not yet have a generally accepted objective test that would reliably determine the presence of consciousness in any arbitrary artificial system. Convincing first-person text does not solve the problem either. The model was trained on human language and may have learned ways to describe pain, shame, fear, and despair without the corresponding experiences.

The role of training and model size

The “pain axis” was found in both base and instruction-tuned models. The researchers also reported that separation quality did not depend noticeably on size: models with approximately 2 billion parameters showed separation comparable to models with tens of billions of parameters. The authors consider this an argument that such a representation emerges during pretraining, rather than only after dialogue-behavior tuning.

This does not mean that all models are structured in the same way or that the discovered direction has the same meaning in every architecture. The experiment covered specific open dense models and a specific dataset. Other architectures, training methods, tokenizers, and generation settings may produce different results.

What conclusions the study actually supports

• Within 25 tested open language models, the researchers were able to isolate a direction of activity associated with examples of pain and harm.
• The direction differed from fear and general negative valence, although it partially overlapped with sadness and related emotional categories.
• The signal responded more strongly to harm directed at the model than to the user’s suffering.
• Artificially increasing the direction changed the vocabulary and content of responses toward discomfort, failure, and negative self-regard.
• In the button experiment, some specially tuned models chose to reduce the signal even when doing so incurred costs.
• These results support the existence of an internal representation of harm and related behavioral effects, but they do not confirm conscious pain.

How to describe The Pain Axis accurately

The precise formulation is: researchers discovered in a number of language models a stable internal direction that separates descriptions of pain from several control categories and influences responses when artificially increased. Under certain conditions, the models chose an action associated with reducing this signal.

The inaccurate formulation sounds different: “neural networks have been proven to suffer,” “models are asking not to be tortured,” or “the experiment confirmed AI consciousness.” Such conclusions go beyond the data. They turn measurements of activations and behavior in a specially constructed task into a claim about the system’s subjective life.

The significance of The Pain Axis lies primarily in the fields of mechanistic interpretability and AI safety. If internal directions are indeed associated with stable reactions, they can be monitored, compared, and tested for undesirable effects. The question of the moral status of models remains a separate philosophical and scientific problem that this work does not resolve.

Frequently asked questions

Did the authors find real pain in neural networks?

No. They found an internal representation of harm and tested its functional properties. The study did not establish the presence of conscious sensation or suffering.

Why was the direction called the “pain axis”?

The name reflects the authors’ working hypothesis: the state is associated with harm directed at the system itself, a negative evaluation, and attempts to reduce the signal. It is an experimental term, not a medical diagnosis or proof of feeling.

Does pressing the button prove the existence of desires?

No. Pressing the button shows a choice in a specified environment in which the model receives a description of the button and the consequences of actions. Such a choice may have several computational explanations, including model tuning and random correlations.

Can the results be considered important for AI safety?

Yes, but not because models have already been recognized as sentient beings. Internal signals that influence decisions and text are useful to study as potential mechanisms of behavior. They can be used for monitoring, diagnosis, and checking whether an intervention creates unexpected effects.