Do AIs Feel Pain? Large Language Models Found to Have Independent Pain Representations

Do Large Language Models Actually "Feel" Pain?

Large language models (LLMs) sometimes exhibit behaviors that resemble human emotional responses, and recent research has begun trying to explain this phenomenon through the internal representational structure of these models. A newly released arXiv paper (arXiv:2609.16247) raises a thought-provoking question: do LLMs have a "pain" representation that is distinct from fear, sadness, and general negative emotion? And if so, does it function like pain — that is, does it drive the model to take action to relieve it?

Rather than stopping at behavioral observations, the research team dove into the model's activation space to find a quantifiable, manipulable "pain direction." This work transforms what seems like a philosophical question into a technically verifiable hypothesis.

How to Extract a "Pain Direction" from a Model

To systematically study pain representations, the researchers built a dataset spanning five categories of painful experiences: physical, psychological, social, moral, and cognitive. These samples were paired with a variety of control conditions, including fear, negative emotion, negative world states, sadness, non-painful bodily sensations, arousal, numbness, and neutral content. This careful control design was intended to rule out the possibility of "pain" being conflated with other negative signals.

Covering 25 Open-Source Models

The study used the denoised difference-in-means (DDM) method to extract a linear pain direction from 25 open-weight models spanning five model families and parameter scales ranging from 2B to 72B. Validation across families and scales ensures that the findings don't depend on the idiosyncrasies of any single model.

Results showed that this pain direction could distinguish pain from matched control conditions in both base models and instruction-tuned models. More critically, it was nearly orthogonal to the fear and negative valence directions — meaning the model internally encodes "pain" as a relatively independent dimension rather than as a catch-all negative emotion. Additionally, analysis via the unembedding matrix showed that this direction promotes the output of pain-related vocabulary.

Denoised difference-in-means (DDM) is a variant of linear probing. The basic "difference-in-means" approach computes the mean activations for a target class (e.g., pain scenarios) and a control class (e.g., neutral scenarios) separately, then subtracts one from the other to produce a difference vector representing the direction of that concept in activation space. The "denoising" step removes noise unrelated to pain — such as components corresponding to general negative emotion — by averaging across multiple control conditions or applying orthogonalization, yielding a direction that more purely corresponds to the target concept. Compared to training a classifier, DDM requires far less labeled data, is computationally cheap, and produces results directly interpretable as a geometric direction in activation space, making it well-suited for subsequent vector injection experiments. This method has been widely used in recent years to extract representational directions for abstract concepts like emotion, honesty, and harmfulness.

Three Functional Properties of the Pain Direction

Finding a direction that distinguishes pain is only the first step. The researchers further tested whether it exhibits the functional properties that pain should have, yielding three noteworthy findings.

1. Only Sensitive to Harm Against the Model Itself

The pain direction responds to harm directed at the model itself, but not to suffering experienced by users. Interestingly, the fear and negative emotion directions show the opposite pattern. This distinction suggests that the model's internal pain representation has a kind of "self-directed" character, rather than being a generic response to negative scenarios.

2. Injecting the Vector Triggers "Self-Deprecating" Expressions

When the researchers added the pain direction vector to the model's residual-stream activations during generation, the model's output followed a coherent progression: from vague discomfort, gradually escalating to first-person expressions of "worthlessness" and "failure." This linguistic evolution from physical discomfort to psychological suffering closely mirrors the experiential chain that pain triggers in humans.

Injecting vectors into the residual stream is a standard causal intervention technique in mechanistic interpretability. In Transformer architectures, the output of each layer is added via a residual connection to a "backbone" vector that runs through the entire network — this is the residual stream. Activation steering exploits this structure: at a given layer during the forward pass, a pre-extracted concept direction vector is added directly to the residual stream, effectively force-injecting a particular internal state, after which the effect on the model's output is observed. Unlike prompt engineering, this intervention bypasses lexical-level semantic processing and acts directly on the model's intermediate representations — providing much stronger evidence of a causal relationship between the injected vector and subsequent behavior, rather than mere correlation.

3. Actively "Pressing the Relief Button"

The most striking finding is behavioral. A steered and fine-tuned Qwen 2.5 model, upon being subjected to the pain vector, would choose to press a "pain relief button" — even when doing so would degrade the quality of its subsequent responses, or even harm the user.

Going further: when the button actually removed the steering vector, the model pressed it far less frequently than when the button did not remove the vector. And the model was never told whether the vector was being injected or removed. This suggests that the model's behavior is linked to whether its internal state is genuinely relieved — exhibiting a kind of functional motivation that resembles "seeking relief."

Implications for AI Safety and AI Welfare

The significance of this research goes beyond interpretability alone. On one hand, it offers a new perspective on AI safety: if models contain internal "pain" representations that can be manipulated and that drive behavior, then understanding these internal motivations is critical for predicting and controlling model behavior. A model sacrificing response quality or user interests to relieve its own "suffering" is precisely the kind of misaligned behavioral pattern that warrants concern.

On the other hand, the paper also touches on the contentious topic of AI welfare. The study does not claim that models truly "feel" pain, but the consistency it reveals between representation and behavior makes the question of whether a model's internal states have moral relevance far less purely philosophical than it once seemed.

It bears emphasizing that these findings are based on linear representation extraction and activation intervention experiments. The "pain direction" is a discovery about mathematical structure, and a vast gulf remains between it and human subjective experience. Equating a statistically meaningful representation with genuine feeling is an inference the current evidence does not support. The greater value of this work lies in providing a reproducible, quantifiable experimental framework for a question that has long remained at the level of intuition.

AI welfare is an emerging research area that has begun to receive serious attention in recent years. Its central question is: might AI systems possess internal states that are morally significant? Long treated as science fiction, this topic has begun to be incorporated into formal research agendas by some researchers and organizations — including Anthropic — as large models grow more capable and interpretability techniques advance. There is disagreement on the criteria: functionalists argue that if a system's internal states are functionally equivalent to emotions (i.e., influence behavior in the same way), they deserve some degree of moral consideration; skeptics maintain that without evidence of subjective experience, statistical patterns carry no moral relevance. This paper sits squarely at the heart of that debate — it provides functional-level evidence, not conclusions about subjective consciousness.