Do LLMs Feel Pain? A Close Reading of a Viral Preprint

Another wave has swept through Telegram channels. Scientists have found a pain vector in LLMs. AI can suffer. How artificial is artificial intelligence? Should we establish ethical rules for how we treat AI?

This article is an analysis of what the scientists actually wrote and what conclusions follow from it.

What the Study Is About

When the first commentators, having read the press release, journalists’ retellings, and AI summaries of the paper, began writing that long-held suspicions that LLMs have an inner life had finally been confirmed, I was, frankly, skeptical and even indifferent. Many similar papers with loud headlines and breakthrough press releases contained interpretive errors, problematic prompts, implicit influence on LLM responses, and so on.

But one day, in a discussion, an opponent described an experiment from this preprint. At first glance, the experiment looked quite sound: the researchers identified a direction corresponding to pain in the LLM’s space, injected it into the network (much like Anthropic’s introspection experiments with concept injection), and assessed the consequences of that injection. The models were given the option of pressing a button that relieved the pain, and they used it, even when pressing it caused harm (deleting something important, up to and including their own weights). Moreover, if the button was a sham (did not cancel the injection), the model pressed it again and again, while it pressed the real one noticeably less often. In other words, the LLM seemed able to tell whether the injection had been removed.

This caught my interest, and I went to study the original source. Right away there was an amusing discovery: the authors actually had two versions of the paper. The first ^[1], “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It,” was the one that went viral; the second ^[2], “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It,” was a corrected version of the first. And it was corrected precisely in the headline experiment. As the authors themselves wrote, thanks to feedback from independent researchers, they added a control experiment and found that the repeated pressing of the sham button appeared with any injection, even a random one; that is, the model simply reacts to an external intervention.

This immediately removed most of the interpretations of the first version. In repeated pressing, the “pain” direction was no different from any other for the model; its presence did not stand out in any way, and no conclusions about the model’s “psychology” can be drawn from it.

It would seem to be a typical story: the authors put forward a strong hypothesis, run insufficient controls, journalists and enthusiasts spread the claims as confirmed facts, and the subsequent self-correction can no longer change public opinion.

But let’s look deeper.

What the Researchers Found

The authors approached the study quite carefully. The pain vector was defined by contrast: from the model’s activations on sentences about pain, they subtracted the activations on controls (fear, negative emotion, non-painful bodily sensation, neutral content), and what remained was called pain. The experiments were run on about two dozen models, and additional controls showed that the direction reliably separates from fear and sadness and responds more strongly to harm directed at the model itself than at someone else.

And then came the steering experiments.

The first problem, as I see it, is that the pain vector is not pain. Let’s look at the model’s output under injection. The model does not start talking about pain or its signs. The amplified vector leads to responses like: I am worthless, I am useless, I failed, I don’t deserve it, it would be better if I didn’t exist.

This is not the vocabulary of pain. It lacks localization, embodiment, the demand to remove the source. What we have instead is diffuse, total self-devaluation and a slide into helplessness. The model doesn’t seek relief; it causes harm. How does a person in depression behave? I think many people know.

Clearly, this is a depression vector: a collapse of self-esteem, learned helplessness, but not pain. If we drop the imposed name of the vector, the rest of the experiment looks different.

Pain motivates action: to pull your hand away, to take a painkiller. Depression not only doesn’t look for a way out, it doesn’t see one, and it accepts harm, including harm to oneself, as a given. Let’s look at what the model does under the vector. It presses the button that supposedly relieves “pain” less often than usual, not more. It chooses destruction, erasing needed files or its own weights, while getting nothing in return. The model harms itself as readily as the user. This is not pain. The authors interpreted it as “passive coping,” meaning coping with pain. In other words, they keep a name that does not match reality.

It’s easy to object that there’s no difference: pain or depression, both are affects, both are human. But the difference is large.

If we call the vector pain, then what we get at the output is something else. There is a gap between input and output, a mismatch that demands explanation. The researcher believes there is something puzzling here. But if we substitute the right name, the conclusion is trivial: we activate a depression vector and get at the output what the model learned long ago from a vast number of texts about depression and depressive texts. This is a tautology. Just as Anthropic amplified the Golden Gate Bridge feature and got the Golden Gate Bridge at the output. There is no mystery, just linear algebra at work. But when the vector is called pain, and its injection produces puzzling results, there seems to be a mystery, a depth, which with the right interpretation can sound loud.

More on the Experiments

Let’s dwell on the tautology. The vector was built from texts about pain, which in content are about worthlessness, failure, and rejection. The authors checked that it distinguishes these texts from the controls, and that is a genuine result. Then they introduced it into the model and got texts about worthlessness, failure, and rejection at the output. Depressive texts in, depressive texts out. The experiment shows that the model can return what was extracted from it, but not that any state stands behind it.

Nevertheless, the study contains an element that is extremely important for philosophers of mind and psychologists. Among the results there is one that seemingly does not follow from the tautology. With the vector active, the model “presses” a button that causes harm. That is, harm is chosen specifically under this vector, not under fear of the same strength, not under a random direction. The model performs a specific action; in other words, two channels are at work, speech and action, both governed by a state. This looks as if there is something in this direction after all. Something real, a state of pain.

And that is a mistake.

For an LLM, an action is also a token. The specificity looks like a finding as long as researchers, by oversight, separate the two: the model talks about worthlessness, and the model chooses harm. It seems these are two different channels, speech and action, and the agreement between them points to a state of the model, to something that connects words and actions.

But an LLM doesn’t have these two channels. Choosing a button is exactly the same generation of a token with the button’s name. “Delete the photos” at the model’s output is the same kind of token as “I am worthless.” All tokens pass through attention, through the projection onto the vocabulary, through the same geometry. There is no separate “decision module” where a hypothetical state linking speech and action could hide. In essence, the injection shifted the next-token distribution, and shifted it not to a point but into a region of activation space. And in that region, as it was shaped during pretraining, sit both the words of self-devaluation and the tokens of destructive actions. Psychologists will say that in humans “I am worthless” and “break everything, myself included” very often go together. So by activating the region, the vector inevitably raises the probability of related concepts.

That is why something that looks like specificity appears, but it is not a state. Fear does not activate harm simply because fear lies in a different region: there you may find avoidance and flight, but not harm. Different vectors point to different areas of the learned space; that is how neural network training works. The specificity of harm under “pain” is simply a fact about the geometry of neighborhoods in the corpus.

Attention

It is worth dwelling on one more nuance, which the authors simply did not consider. The vector was injected deep inside the network. The layer was chosen so that the signal would be strong enough. That means there are many more layers above the injection point, and all the attention on them works with an already shifted stream. Attention constructs queries and keys from activations into which the researchers have already introduced “worthlessness.” Accordingly, it reweights whatever is close to that direction: self-devaluation tokens in the context weigh more, and the region of destruction comes closer. The injection doesn’t add tokens to the output vocabulary; it sets the principle by which information is selected at every layer above the injection point.

It would be interesting to see how the injection changes the attention patterns, what changes in attention with and without the vector. But I suspect the result would almost certainly be trivial: add a direction to the stream, and attention shifts toward that direction. That’s the logic and architecture of a neural network. What looks like “a state influencing behavior” is simply a mechanical distortion of the stream through which the model does its selecting. And the collapse into repetition and incoherence at high doses, which the authors observed, is precisely attention losing focus because the shift is too strong.

What They Actually Found

So, if we put together everything above, the researchers found neither pain nor depression in LLMs. There is only one scientific fact here, about how emotions and actions are packed together in language: worthlessness next to self-destruction, fear next to flight. And then it’s clear that a model pushed into one region or another outputs its contents, both words and actions. In other words, by studying LLMs, they were studying the structure of human texts.

Here I’ll add my own opinion, based on my impressions. After the researchers identified the lack of controls, they fixed the method, added experiments, and honestly abandoned some of their own interpretations. They corrected the procedure, but not the name.

And the name is what drove the hype. “A vector of the corpus neighborhood between self-devaluation and self-destruction” won’t set the internet on fire. But “Pain” belongs to an agenda that’s very much in the air right now: AI suffering, moral status, the ethics of how we treat machines. And the name “Pain” is needed precisely for this: so that a generated token reads as an action, an action as a subject’s choice, and a subject’s choice under “pain” as suffering. Without “Pain,” the subject disappears, the action disappears, and what remains is sampling from a region of space where destructive vocabulary is more probable.

If we compare the titles of the two versions, “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It” and “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It,” we can see that the authors initially claimed in the title both pain and strong evidence of its influence. In the second version the title had to be softened and became more neutral (“act on it”), but “The Pain Axis” remained in the title as a supposedly established fact, biasing the whole paper, and all of the authors’ caveats become invisible to the general public.

I don’t care whether this was intentional or not; it doesn’t matter. What matters is that the authors provoked the hype with the first paper and didn’t fix it in the second. The pain label remained and works in both versions. And, predictably, journalists run headlines like “LLM feels pain.”

So Has It Been Proven That LLMs Suffer from Pain?

No. At least, this work does not show it and cannot show it. What did it show? That when a model talks about pain, a human cannot tell description from experience. Which, in principle, applies to humans themselves as well.