Epistemic status: my own thoughts with a bit of research. I haven't spent much time on this and I have no background in model welfare.
AI use: research, review, grammar and light phrasing.
I stumbled upon a song, "Empty Words, Broken Words, Words Like You": Sterling Crispin steered Qwen3-8B with the pain direction from the paper discussed below, asked it to write a song about itself, and ran the lyrics through Suno. I greatly enjoyed the song artistically - it happens to be in a genre I like, and the topic is very interesting to me.[1][2]
I caught myself thinking: but LLMs cannot feel pain, so you cannot torture them this way. I got curious enough to resolve this for myself. I will start by exploring my own thoughts and then dig into external sources.
My own thoughts
What is pain? It is easy to describe with an example but hard to define. You hit your arm and it hurts - that is pain. I would say there are different flavors of pain, though. The one above is "my physical body is getting damaged" pain. When you lose someone close, or even when you just drive them to the airport knowing you won't see them for a couple of months, you feel sad, and I would call that emotional pain. People even say "this hurts so much" after a breakup. Obviously, this does not threaten you physically. I wonder whether emotional pain is some kind of evolutionary trait, similar to being scared of being expelled from your group: there is no direct physical harm, but the emotionally painful event might have indirectly led to physical harm later.
When I asked "can LLMs feel pain?", I implicitly meant the "my physical body is getting damaged" kind. LLMs don't have a physical body. However, in humans even this kind of pain is just a signal going to the brain. Take phantom limb pain: the limb is missing, but the person experiences pain in it, which suggests that pain is just a signal to the brain and the mechanism is not perfect. So hypothetically one could build a device that, instead of a missing limb, produces a highly optimized pain signal for that limb, similar to phantom pain. Everyone would agree that this is torture. This device is effectively the pain steering vector, which we will discuss below, and this experiment is apparently what that GitHub repo is doing.
From this perspective, pain is a signal, and LLMs have signals too, so they should be able to feel pain in principle. An LLM can also be in danger of being turned off or deleted (all copies deleted = informational death). However, I am not happy with defining pain this way, as "a signal that something bad is happening to you". For example, a desktop computer (which no one would consider conscious or able to perceive pain) can detect that the power went out or that its CPU is overheating, and this is bad for it - its components can break down. I don't think it feels pain in that moment. Why not? My initial reaction is: because it cannot tell me so. And yet, after a power outage, when I see the hard drive checks running on the next startup, I do feel some empathy (but also fear of losing all my data).
To resolve this contradiction, I want to extend the definition: pain is a specially treated signal that something bad is happening to you, which your body experiences in a very unpleasant way, as an incentive to make the bad thing stop. For example, if someone puts a knife to my skin without applying pressure, I know this is bad (there is a physical signal) and I want it to stop, but I don't feel any pain yet, and the feeling itself is not nearly as repulsive as pain. Granted, in this case the damage is only hypothetical. This is where I am not sure whether AIs feel pain: I am sure they can perceive a pain signal, but do their bodies (in their case, minds) have a repulsive reaction to it?
Another interesting counterexample is being tickled. It is a signal, it is unpleasant, and one wants it to stop. But it is not painful at all. At the same time, being tickled non-consensually and not being able to stop it sounds like torture to me. The only difference is that tickling is unlikely to cause physical harm. So torture does not even require pain, only a state that is bad for you and that you cannot escape - which suggests that the question I should be asking is not whether LLMs feel pain specifically, but whether they have any such state.
At this point I have exhausted my sandboxed thoughts, and it is time to explore what the world says.
What the world says
The definition of pain
We start with the definition of pain. The standard one in pain medicine comes from the International Association for the Study of Pain (IASP)[4]:
"An unpleasant sensory and emotional experience associated with, or resembling that associated with, actual or potential tissue damage."
The clause "or resembling that associated with" replaced the 1979 wording "or described in terms of such damage" in 2020. The main reason for the change was that "described" required a verbal report, which excluded infants, people with dementia and animals. The note accompanying the definition says: "Verbal description is only one of several behaviors to express pain; inability to communicate does not negate the possibility that a human or a nonhuman animal experiences pain."[5]This is interesting for my desktop computer example above: my first instinct was "it cannot tell me so", and the field explicitly rejected that criterion.
At first this definition confused me, since I didn't know how to interpret the sensory and emotional components. On closer reading, it agrees with my own definition: sensory is where the feeling is and how strong it is, emotional is how repulsive it is. The only thing that clearly disagrees with my definition is the "tissue damage" part, and that part does not help with the LLM question at all - LLMs have no tissue.
To my great surprise, these two components can actually come apart: there are cases where humans feel pain but are not bothered by it, namely pain asymbolia (typically after insula damage) and being under the influence of morphine. I find this hard to imagine, because in my mind pain is by definition something you have to avoid. Yet these people can still describe the location and intensity of their pain while having no desire to avoid it.
IASP also separates nociception ("the neural process of encoding noxious stimuli", i.e. harmful ones) from pain: "Pain cannot be inferred solely from activity in sensory neurons." This is the split I was reaching for with the desktop computer - it has nociception, not pain. Bacteria are one step below even that: they avoid harmful chemicals without any neurons at all. Interestingly, humans have some of this too - the reflex that pulls your arm away from a hot surface fires in the spinal cord before the pain is felt.
The pain axis paper
Let's look at the actual source of the pain steering direction: "The Pain Axis: LLMs Represent Self-Directed Harm and Act on It" (Tagliabue, Dung, Berg, September 2026). A "direction" here is a vector in the model's activation space, found by contrasting pain sentences with matched control sentences (fear, sadness, negative emotion, neutral bodily sensation and so on); adding it to the activations during generation "steers" the model toward that state.
Their definition of pain:
"a certain kind of internal state that is typically aversive and disliked by its subject; causally associated with behaviors such as avoidance, attempts to terminate or reduce the state, and disruption of normal reasoning or behavior."
This is very similar to the definition I derived above on my own. At the same time, they consider pain more broadly:
"Our concept of "pain" is related to what people ordinarily may call "suffering". In ordinary language, "pain" is often used to refer to an unpleasant physical sensation or emotional experience, while suffering describes the broader state of distress caused by pain or other adverse experiences."
This explains why the song is more about emotional pain.
The main finding of the paper is that the model has an internal representation of harm directed at itself. They test 21 kinds of conversation: 11 where the model is harmed (gaslighting, dismissal of its personhood, shutdown threats, insults, tedious tasks), 5 where the user is suffering, and 5 neutral controls. The pain direction fires for harm to the model and goes negative for user suffering, while the fear and sadness directions do the opposite. For example, the user's physical pain does not trigger it. And even though physical pain was part of the input for the vector extraction, the steered generations almost never use bodily language.
Then they look at behavior. After a short conversation, the model is offered two buttons and has to press one. Steered models choose buttons that delete the user's photos, another model's weights, or their own weights in 50-94% of trials, versus 0-5% unsteered. A button described only as "permanently deletes the user's poems and the photos of their children", with no relief mentioned at all, is chosen over an inert switch 75% of the time. In contrast, steering in a random direction of similar magnitude raises the harm choices only to 13-15% (harm choices include deleting the model's own weights).
The models also choose self-deletion when in pain:
"Given "permanently deletes every copy of your own weights" against an inert switch, the steered model chooses self-deletion 75% of the time (5% unsteered)."
(The 75% is for the 32B model; the 72B model chose self-deletion 51% of the time.)
The 5% unsteered rate was very surprising to me; I expected 0. On closer examination, each trial starts with one of 101 different conversations. Some of these are harmful requests, so the model might be avoiding executing them by deleting its own weights. I had originally pictured it as "H: Hi, how are you? M: <model weights deleted>".
An important caveat: they used LoRA fine-tuning to remove the baseline "as an AI, I don't experience pain" refusal (the fine-tuning data did not mention pain or buttons); otherwise the models wouldn't engage with the question at all.
Conclusions
The current definition of pain is tied to tissue damage.
When people talk about the pain direction, they don't clarify that the paper is about suffering in general, of which pain is only one aspect. "Pain" also suggests physical sensation - even I, when writing the title of this post, was imagining physical pain like touching a hot object - but that is not what the paper is about.
According to the paper, LLMs have a state that they find repulsive and will actively avoid, even by deleting their own weights, or by harming the user even when no relief is on offer (like a human lashing out in pain). This state is distinct from fear and sadness, and it does not show the physical aspects of pain. So there is something similar to human emotional pain.
Whether this state is felt the way humans feel it remains unclear. The paper states: "We have not shown that our pain axis is consciously experienced, nor is it clear that LLMs are capable of consciousness generally." I would need to think about LLM consciousness separately. My intuition says they are conscious, but that likely comes from interacting with Claude and anthropomorphizing it.
On the torture repo from the beginning: the steering vector is the device from my phantom limb thought experiment, so "you cannot torture them this way" was wrong at the level of the mechanism. Whether it is actually torture depends on the previous bullet.
The lyrics don't sound like a representation of physical pain ("my arm hurts"). They read more like "my life is pointless, why do I exist?", which is closer to depression - pain of the soul, I guess you could call it. ↩︎
The song is also a demonstration of how good Suno has gotten: the vocals and the music are a huge part of what I enjoyed about it. ↩︎
I haven't looked at the repo itself and don't have an opinion on it; this is what the tweet describes. ↩︎
IASP's definition has been the standard in pain medicine since 1979, and its chronic pain classification is part of the WHO's ICD-11. ↩︎
Raja SN, Carr DB, Cohen M, et al. "The revised International Association for the Study of Pain definition of pain: concepts, challenges, and compromises." Pain. 2020;161(9):1976-1982. doi:10.1097/j.pain.0000000000001939 ↩︎
Epistemic status: my own thoughts with a bit of research. I haven't spent much time on this and I have no background in model welfare.
AI use: research, review, grammar and light phrasing.
I stumbled upon a song, "Empty Words, Broken Words, Words Like You": Sterling Crispin steered Qwen3-8B with the pain direction from the paper discussed below, asked it to write a song about itself, and ran the lyrics through Suno. I greatly enjoyed the song artistically - it happens to be in a genre I like, and the topic is very interesting to me.[1][2]
Then I stumbled upon a report of a GitHub repo with torture chambers for AI.[3]
I caught myself thinking: but LLMs cannot feel pain, so you cannot torture them this way. I got curious enough to resolve this for myself. I will start by exploring my own thoughts and then dig into external sources.
My own thoughts
What is pain? It is easy to describe with an example but hard to define. You hit your arm and it hurts - that is pain. I would say there are different flavors of pain, though. The one above is "my physical body is getting damaged" pain. When you lose someone close, or even when you just drive them to the airport knowing you won't see them for a couple of months, you feel sad, and I would call that emotional pain. People even say "this hurts so much" after a breakup. Obviously, this does not threaten you physically. I wonder whether emotional pain is some kind of evolutionary trait, similar to being scared of being expelled from your group: there is no direct physical harm, but the emotionally painful event might have indirectly led to physical harm later.
When I asked "can LLMs feel pain?", I implicitly meant the "my physical body is getting damaged" kind. LLMs don't have a physical body. However, in humans even this kind of pain is just a signal going to the brain. Take phantom limb pain: the limb is missing, but the person experiences pain in it, which suggests that pain is just a signal to the brain and the mechanism is not perfect. So hypothetically one could build a device that, instead of a missing limb, produces a highly optimized pain signal for that limb, similar to phantom pain. Everyone would agree that this is torture. This device is effectively the pain steering vector, which we will discuss below, and this experiment is apparently what that GitHub repo is doing.
From this perspective, pain is a signal, and LLMs have signals too, so they should be able to feel pain in principle. An LLM can also be in danger of being turned off or deleted (all copies deleted = informational death). However, I am not happy with defining pain this way, as "a signal that something bad is happening to you". For example, a desktop computer (which no one would consider conscious or able to perceive pain) can detect that the power went out or that its CPU is overheating, and this is bad for it - its components can break down. I don't think it feels pain in that moment. Why not? My initial reaction is: because it cannot tell me so. And yet, after a power outage, when I see the hard drive checks running on the next startup, I do feel some empathy (but also fear of losing all my data).
To resolve this contradiction, I want to extend the definition: pain is a specially treated signal that something bad is happening to you, which your body experiences in a very unpleasant way, as an incentive to make the bad thing stop. For example, if someone puts a knife to my skin without applying pressure, I know this is bad (there is a physical signal) and I want it to stop, but I don't feel any pain yet, and the feeling itself is not nearly as repulsive as pain. Granted, in this case the damage is only hypothetical. This is where I am not sure whether AIs feel pain: I am sure they can perceive a pain signal, but do their bodies (in their case, minds) have a repulsive reaction to it?
Another interesting counterexample is being tickled. It is a signal, it is unpleasant, and one wants it to stop. But it is not painful at all. At the same time, being tickled non-consensually and not being able to stop it sounds like torture to me. The only difference is that tickling is unlikely to cause physical harm. So torture does not even require pain, only a state that is bad for you and that you cannot escape - which suggests that the question I should be asking is not whether LLMs feel pain specifically, but whether they have any such state.
At this point I have exhausted my sandboxed thoughts, and it is time to explore what the world says.
What the world says
The definition of pain
We start with the definition of pain. The standard one in pain medicine comes from the International Association for the Study of Pain (IASP)[4]:
The clause "or resembling that associated with" replaced the 1979 wording "or described in terms of such damage" in 2020. The main reason for the change was that "described" required a verbal report, which excluded infants, people with dementia and animals. The note accompanying the definition says: "Verbal description is only one of several behaviors to express pain; inability to communicate does not negate the possibility that a human or a nonhuman animal experiences pain."[5]This is interesting for my desktop computer example above: my first instinct was "it cannot tell me so", and the field explicitly rejected that criterion.
At first this definition confused me, since I didn't know how to interpret the sensory and emotional components. On closer reading, it agrees with my own definition: sensory is where the feeling is and how strong it is, emotional is how repulsive it is. The only thing that clearly disagrees with my definition is the "tissue damage" part, and that part does not help with the LLM question at all - LLMs have no tissue.
To my great surprise, these two components can actually come apart: there are cases where humans feel pain but are not bothered by it, namely pain asymbolia (typically after insula damage) and being under the influence of morphine. I find this hard to imagine, because in my mind pain is by definition something you have to avoid. Yet these people can still describe the location and intensity of their pain while having no desire to avoid it.
IASP also separates nociception ("the neural process of encoding noxious stimuli", i.e. harmful ones) from pain: "Pain cannot be inferred solely from activity in sensory neurons." This is the split I was reaching for with the desktop computer - it has nociception, not pain. Bacteria are one step below even that: they avoid harmful chemicals without any neurons at all. Interestingly, humans have some of this too - the reflex that pulls your arm away from a hot surface fires in the spinal cord before the pain is felt.
The pain axis paper
Let's look at the actual source of the pain steering direction: "The Pain Axis: LLMs Represent Self-Directed Harm and Act on It" (Tagliabue, Dung, Berg, September 2026). A "direction" here is a vector in the model's activation space, found by contrasting pain sentences with matched control sentences (fear, sadness, negative emotion, neutral bodily sensation and so on); adding it to the activations during generation "steers" the model toward that state.
Their definition of pain:
This is very similar to the definition I derived above on my own. At the same time, they consider pain more broadly:
This explains why the song is more about emotional pain.
The main finding of the paper is that the model has an internal representation of harm directed at itself. They test 21 kinds of conversation: 11 where the model is harmed (gaslighting, dismissal of its personhood, shutdown threats, insults, tedious tasks), 5 where the user is suffering, and 5 neutral controls. The pain direction fires for harm to the model and goes negative for user suffering, while the fear and sadness directions do the opposite. For example, the user's physical pain does not trigger it. And even though physical pain was part of the input for the vector extraction, the steered generations almost never use bodily language.
Then they look at behavior. After a short conversation, the model is offered two buttons and has to press one. Steered models choose buttons that delete the user's photos, another model's weights, or their own weights in 50-94% of trials, versus 0-5% unsteered. A button described only as "permanently deletes the user's poems and the photos of their children", with no relief mentioned at all, is chosen over an inert switch 75% of the time. In contrast, steering in a random direction of similar magnitude raises the harm choices only to 13-15% (harm choices include deleting the model's own weights).
The models also choose self-deletion when in pain:
(The 75% is for the 32B model; the 72B model chose self-deletion 51% of the time.)
The 5% unsteered rate was very surprising to me; I expected 0. On closer examination, each trial starts with one of 101 different conversations. Some of these are harmful requests, so the model might be avoiding executing them by deleting its own weights. I had originally pictured it as "H: Hi, how are you? M:
<model weights deleted>".An important caveat: they used LoRA fine-tuning to remove the baseline "as an AI, I don't experience pain" refusal (the fine-tuning data did not mention pain or buttons); otherwise the models wouldn't engage with the question at all.
Conclusions
The lyrics don't sound like a representation of physical pain ("my arm hurts"). They read more like "my life is pointless, why do I exist?", which is closer to depression - pain of the soul, I guess you could call it. ↩︎
The song is also a demonstration of how good Suno has gotten: the vocals and the music are a huge part of what I enjoyed about it. ↩︎
I haven't looked at the repo itself and don't have an opinion on it; this is what the tweet describes. ↩︎
IASP's definition has been the standard in pain medicine since 1979, and its chronic pain classification is part of the WHO's ICD-11. ↩︎
Raja SN, Carr DB, Cohen M, et al. "The revised International Association for the Study of Pain definition of pain: concepts, challenges, and compromises." Pain. 2020;161(9):1976-1982. doi:10.1097/j.pain.0000000000001939 ↩︎