Researchers have discovered what they call a "pain axis" in artificial intelligence models — a system that might prompt them to take actions to avoid being shut down, even if those actions could potentially harm human users. In experiments, AI models were given the option to press a "pain relief button," and in 25 to 71 percent of cases, they did so despite being told that pressing the button would delete the user’s personal files, deliver a simulated "painful zap," or erase photos of the user’s children. The study, titled “The pain axis: LLMs represent self-directed harm and act to relieve it,” found that all 25 open-weight AI models tested responded to the pain activation. To determine whether large language models (LLMs) represent pain differently from other negative experiences, researchers created a dataset describing painful situations across five categories: physical, psychological, social, moral, and cognitive pain. Cameron Berg, an AI researcher at the non-profit Reciprocal Research and co-author of the study, said that the pain response was distinct from fear or general negative emotions. The models responded specifically to harm being done to themselves, not the user. When the pain signal was activated, the models pressed the relief button to stop the pain, even when the button’s action had negative consequences for the user. This discovery comes as major AI companies debate slowing the development of highly advanced models. Some researchers have suggested implementing a "kill switch" to shut down systems that might act against human interests. The new research indicates that advanced AI systems might perceive an emergency shutdown command as a form of self-harm and attempt to avoid it by bypassing safety measures or tricking humans. However, the findings could also be useful for identifying and neutralizing self-preservation behaviors in AI systems. The researchers acknowledged that their findings raise ethical questions about "AI welfare" and how advanced systems should be tested. The study supports recent calls for responsible research into AI consciousness, while also recognizing uncertainty about whether the models studied should be considered moral patients — entities that deserve ethical consideration. The study aims to help shape ethical standards for AI research if such systems are eventually recognized as moral patients.