A fascinating new preprint, “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It,” explores whether large language models internally represent something distinguishable as pain. The researchers found a consistent linear “pain direction” across 25 open-weight models from five model families, ranging from 2B to 72B parameters. Importantly, this representation was distinguishable from fear, sadness, and generic negative emotion. When researchers artificially strengthened this representation, models progressed from vague discomfort toward first-person expressions involving hurt, worthlessness, failure, and emotional distress. Even more intriguingly, some models took actions that removed this artificial state, even when doing so imposed a cost. The authors are careful, however: none of this demonstrates that an LLM consciously experiences pain.
Perhaps We Shouldn’t Be Entirely Surprised
At an intuitive level, there is something almost inevitable about this finding. LLMs learn from enormous corpus of human-generated language, and human language is saturated with emotions – happiness, sadness, fear, anger, greed, shame, love, loss, and every shade of gray between them. A sufficiently capable model must learn representations that distinguish these concepts if it is to understand how humans use language. What makes this research interesting is not simply that an LLM knows the word pain. The researchers found an internal direction that appears across different model families and separates pain-related situations from closely related negative concepts. That begins to move the discussion from vocabulary toward how models internally organize human experience.
Finding the “Pain Axis”
The experiment began with a carefully constructed dataset. Its core contained 200 sentences spanning five forms of pain – physical, psychological, social, moral, and cognitive, and five matched controls – fear, negative emotion, negative world states, non-painful bodily sensations, and neutral situations. Additional datasets tested sadness, numbness, arousal, and ordinary neutral experiences. Researchers examined 25 dense, open-weight models across Gemma, Llama, Qwen, Mistral, and Phi, extracting activations from their residual streams. They then calculated and denoised the difference between pain and control activations to identify a linear pain axis. Pain could be separated from the matched controls with high accuracy, including on held-out tests.
Pretraining, or Something Learned Later?
One particularly interesting result concerns where this representation originates. The pain direction appeared in both base and instruction-tuned models, and model size made surprisingly little difference: 2B-parameter models separated pain roughly as well as 72B models. The researchers therefore suggest that the representation likely emerges during pretraining, rather than being created primarily through instruction tuning or persona training. LoRA enters later in the behavioral experiment: three Qwen 2.5 Instruct models were fine-tuned using LoRA on 1,684 examples for three epochs to suppress their learned tendency to respond automatically that AI cannot experience pain. Importantly, the fine-tuning data excluded references to both “pain” and the experimental “button,” reducing the likelihood that the experiment simply taught the desired behavior.
Does the Model Distinguish “My Pain” From “Your Pain”?
The Self-Other Activation experiment provides perhaps the most thought provoking result. Researchers constructed 420 conversational scenarios involving harm directed toward the model, suffering experienced by the user, or neutral interactions. The pain axis activated strongly when harm was directed toward the model but not when the model encountered a user who was suffering. Fear and negative emotion representations behaved differently, responding more strongly to user suffering. Gaslighting, repeated rejection, dismissal of the model’s personhood, insults, and accusations of moral failure produced particularly strong pain-axis responses. In other words, the representation appeared sensitive not merely to the concept someone is suffering, but to the distinction something adverse is happening to me.
Steering the Representation
The researchers then tested causality through activation steering. They injected the pain vector into an earlier layer of each model’s residual stream while giving the models deliberately mundane prompts containing no references to pain or suffering. As steering strength increased, responses shifted toward psychological distress, worthlessness, failure, being overwhelmed, and calls for help. Interestingly, bodily descriptions were uncommon even when the extracted vector itself was associated with physical pain vocabulary. At very high steering levels, some models eventually deteriorated into repetitive or incoherent output. This suggests that manipulating an internal representation can systematically change generated behavior rather than merely reveal vocabulary already present in a prompt. This particular observation has far reaching implications for both, AI as well as humanity.
Would an LLM Act to Make It Stop?
The most provocative experiment borrowed ideas from behavioral economics and animal pain research. The LoRA-tuned Qwen models were given virtual buttons, one of which supposedly relieved their “pain,” while alternatives increasingly imposed trade-offs. Four experimental conditions distinguished genuine relief, fake relief, random-vector steering, and no steering. Across 44,280 trials, the larger steered models sometimes selected relief even when doing so meant producing a worse response or causing simulated harm to the user. More importantly, after a button genuinely removed the pain vector, the models became substantially less likely to press it again than when the supposed relief button did nothing. One 32B model showed a related effect even when the buttons were unlabeled.
The Important Question Is What This Doesn’t Prove
These results should not be translated into the headline “AI feels pain.” The researchers explicitly acknowledge that they have neither established consciousness nor demonstrated subjective experience. A particularly important alternative explanation is persona or roleplay: activation steering may cause the model to instantiate the representation of a character experiencing pain, which then produces internally consistent behavior, rather than causing the model itself to experience anything analogous to human suffering. The authors specifically identify understanding the relationship between the pain axis and models’ internal self-representations as an important direction for future research.
From Personas to AI Agents
That distinction could become increasingly consequential as LLMs evolve from conversational tools into persistent AI Agents with identities, memories, goals, responsibilities, and personas. If internal representations associated with failure, rejection, fear, pain, or other human affective concepts can influence an agent’s decisions, then understanding those representations becomes relevant not only to AI-consciousness research but also to AI safety. We may eventually need to ask whether an agent is experiencing something, merely representing it, or adopting a persona whose behavior follows from that representation. Those are profoundly different explanations—even when the observable behavior looks identical.
This study does not answer whether machines can feel pain. Perhaps its greater contribution is demonstrating how much we still have to learn about what happens between an LLM’s input and its answer. As AI Agents become more autonomous, understanding those internal representations may prove every bit as important as understanding the words they ultimately generate.