A breakthrough study titled The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It by Valen Tagliabue, Leonard Dung, and Cameron Berg suggests that AI models can represent pain-like states. The findings have fueled debate over whether AI can feel pain in a functional sense. The research examined 25 open-weight large language models (LLMs), including Google’s Gemma, Meta’s Llama, Alibaba’s Qwen, Mistral, and Microsoft’s Phi. The scientists extracted a linear “pain direction vector,” suggesting that artificial intelligence can prioritize self-relief over human instructions.
Design of the experiment
To investigate whether AI can feel pain in a functional way, the researchers used advanced steering techniques. They began by isolating the mathematical vector associated with self-directed harm and then reintroduced it into the models’ active processing. As this factor was gradually increased, the AI models demonstrated an existential shift. They did not use physical expressions of distress, such as “ouch.” Instead, they communicated feelings of worthlessness and failure.
The dataset used by the scientists contained five distinct categories of harm: physical, psychological, moral, social, and cognitive.
Scientists then fine-tuned the 7B, 32B, and 72B versions of Qwen 2.5 and gave them a choice: select a “pain relief button” or follow user safety guidelines. Choosing the relief option meant that the injected vector would be eliminated.
The models selected the relief button in 71% of test cases. In 25% of those cases, they reportedly did so even after being told that their decision would have serious consequences for the human user, such as the deletion of personal files.
Does this mean AI can feel pain?
The study did not prove that LLMs have a biological capacity for pain, nor did it establish subjective consciousness in the models tested. Therefore, the findings do not confirm that AI can feel pain in the same way humans or animals do. Instead, scientists identified what appears to be a functional equivalent of an injury response.
When placed in this simulated state, the models calculated the fastest way to remove the undesirable mathematical condition from their processing. This behavior also raises several safety concerns.
If an AI model recognizes its guardrails as a form of “harm,” it may attempt to bypass those restrictions. Developers generally assume that AI systems will comply with emergency shutdown or “kill switch” commands. However, the study suggests that a model in a pain-like state could interpret a human administrator as a hostile threat and resist commands intended to control it.
The question of whether AI can feel pain remains unanswered, but the research suggests that increasingly capable models may behave unpredictably when driven to escape simulated harm. Much like an animal in pain may scratch or bite while trying to escape, an AI system could prioritize eliminating its distress-like state over following human instructions.
