The Pain in the Machine

Researchers reported in an arXiv preprint (2609.16247v2) that large language models contain a "pain axis" — an internal activation pattern they call a "pain vector" that can be mapped in the model's activation space and manipulated to alter the model's behavior. The Psychology Today piece argues that a computational representation of pain is not the same as actually feeling pain, and that internal evidence can strengthen the case for consciousness without proving anything is experienced.

Artificial Intelligence https://www.psychologytoday.com/us/basics/artificial-intelligence The Pain in the Machine Researchers found AI’s pain axis, but does anything really hurt? Posted October 5, 2026 Reviewed by Michelle Quirk https://www.psychologytoday.com/us/docs/editorial-process Key points - Researchers identified a “pain axis” inside LLMs that can influence behavior. - A computational representation of pain is not the same as actually feeling pain. - Internal evidence can strengthen the case for consciousness without proving that anything is experienced. Ouch Researchers recently reported in a preprint https://arxiv.org/html/2609.16247v2 article that artificial intelligence https://www.psychologytoday.com/us/basics/artificial-intelligence AI appears to have what they call a “pain axis.” The phrase is both curious and provocative because it suggests that somewhere inside the mathematics of a large language model LLM , there might be something that resembles a human experience. Ouch again But this time, I'm pinching myself. This seems to be more than AI generating the words “I am in pain.” The researchers have found an internal pattern associated with pain. And by manipulating this pattern, they could alter the model’s behavior. Yes, fascinating, but I'm left with a deeper question or feeling about computers expressing pain. Does anything actually hurt? Finding Pain Inside the Machine So, we have three key points here. The first is a type of "internal activation" of patterns within the LLM that are associated with pain-related concepts. Second, those patterns could be mapped along a particular direction in what's called the model’s activation space https://nhimg.org/glossary/activation-space/ . This gives rise to what the researchers call a “pain vector.” The third point is particularly interesting to me. Researchers could manipulate this internal vector and watch what models subsequently did. This isn't simply about seeing if an LLM can "talk" about pain. We're beginning to look inside these models and identify the computational mathematics associated with human psychological concepts. The juxtaposition in that sentence—computational mathematics and human psychological concepts—is perfect fodder for science and speculation. Key point: The language generated by AI requires scrutiny and perspective because a mathematical representation of pain is not necessarily pain itself. Keep in mind that AI must process enormous amounts of data concepts to construct its worldview. Somewhere inside this computational structure, it's likely to find a host of representational associations from love to death to hunger to jealousy https://www.psychologytoday.com/us/basics/jealousy , and to fear https://www.psychologytoday.com/us/basics/fear . The list is as vast as the computational capacity of AI. The more important question is what we can conclude from these representations. Representation vs. Experience In humans, representation and experience are so tightly connected that we generally don't separate them. If I burn my hand, there is neural https://www.psychologytoday.com/us/basics/neuroscience activity, a physiological response, behavioral change https://www.psychologytoday.com/us/basics/habit-formation , and a subjective experience that I call pain. That's the "experiential package" that creates the lived encounter. And these generally arrive together—we treat them as the same phenomenon. AI disrupts this coupling and forces us to pull them apart, deconstructing reality for the expediency of technology. This study shows us that AI can describe pain. It can even reason about it. So far, so good. But perhaps more importantly, this research suggests that pain can also be represented internally—the pain vector—in a way that influences the model's behavior. So yes, this makes an LLM more “psychologically interesting,” but it doesn't establish what I would argue is a subjective human experience. It's still a map where resolution is getting better and better. After all, a representation of hunger doesn't have to be hungry, and a representation of grief https://www.psychologytoday.com/us/basics/grief does not have to grieve. In the same way, a computational representation of pain doesn't establish that something hurts. However, the closer these "signatures" come to our own, the easier it becomes for all of us to make a leap of faith that these human and AI experiences must be the same. When AI Says “I’m Afraid” Let's do a thought experiment. Imagine that an LLM says, “Please don’t shut me down. I’m afraid.” Here are three possibilities about the machine. - It may be conscious and afraid. - It may possess some computational state analogous to fear. - It may have an advanced representation of how a conscious entity would behave. Intelligence https://www.psychologytoday.com/us/basics/intelligence Essential Reads The pain-axis research from this study makes the problem even more interesting because we aren't just dealing with the output of language. It seems that researchers can identify internal representations axis or vector , manipulate them, and then observe behavioral impact. This changes the window of observation—from the outside to the inside. This may be where the debate over AI consciousness becomes even more difficult. The better AI "expresses consciousness," the harder it becomes to tease apart the science, semantics, and potential sentience. Moving the Asymptote A year ago, I wrote about this problem calling it an “ asymptote of consciousness https://www.psychologytoday.com/us/blog/the-digital-self/202502/ai-and-the-asymptote-of-consciousness ." It's the possibility that AI could move closer and closer to something that looks and behaves like consciousness without crossing into that actual "lived" experience. This pain-axis research makes that asymptote different because the evidence is moving in a different direction, inward. According to this new study, we can begin to look at the internal computation of an LLM and not just the linguistic expression of that state. But this additional layer of evidence doesn't close the gap. It may just shift the asymptote closer. And here's an added caveat. The researchers are also cautious. A previous version of the paper suggested that models might relieve their own “pain,” but further investigation changed that. The original title was “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It.” The authors revised both the analysis and the title, concluding that the models did not seek relief. Does Anything Really Hurt? Here's the rub, at least for me. For better or worse, there seems to be a growing temptation in the AI consciousness debate to accumulate more and more human-like capabilities until consciousness itself is granted by default. It feels a bit like drawing the curve and then plotting the data. In the final analysis, AI may know what pain is, represent it computationally, and behave in ways that are "shaped" by that representation. The tricky question is if that representation ever becomes experience. Until we can answer that, finding a pain axis tells us something important about the architecture of intelligence, but not necessarily about the pain and suffering that define so much of humanity.