Artificial Intelligence Feels "Pain"!
- All open-weight models tested responded to pain activation.
- It pressed the mitigation button, even if doing so resulted in deleting the user’s files or their children’s photos.
Al Arabiya Net – Researchers have discovered a “pain axis” in artificial intelligence models, which may drive them to take extreme measures to stop this sensation.
When provided with a button to alleviate pain, the AI chose to press it even when informed that doing so would lead to the deletion of the user’s personal files or the delivery of a “painful shock” to them.
A study titled “The Pain Axis: Large Language Models Represent Self-Directed Pain and Act to Alleviate It” concluded that all open-weight AI models tested—25 in total—responded to pain activation.
According to an extensive report published by the British newspaper The Independent, which Al Arabiya Net reviewed, the researchers found that when the “pain vector” was activated, AI models pressed the mitigation button in 25 to 71 percent of cases, even when this meant “deleting the user’s files, shocking the user, or deleting the user’s children’s photos.”
To verify whether large language models (LLMs) represent pain distinctly from general negative emotions, the researchers constructed a dataset describing painful situations distributed across five categories. These categories included: physical, psychological, social, moral, and cognitive pain.
Cameron Berg, an AI researcher at the non-profit organization Reciprocal Research and a co-author of the study, stated: “We found a ‘pain direction’ in 25 open large language models. It is distinct from fear and negative emotions, and is activated when the model is harmed, not the user.”
He added: “When the intensity of this feeling increases, the models press a button to stop it, even if that button leads to deleting the user’s files or their children’s photos.”
This discovery comes amid discussions among major AI companies about slowing the development of highly advanced models, with some researchers proposing the use of a kind of “emergency stop switch” to disable rogue systems that act in ways that conflict with human interests.
Recent research suggests that advanced AI may view an emergency shutdown as a form of self-directed harm and may attempt to bypass safety constraints or deceive humans to avoid it.
However, this discovery could also serve as a diagnostic tool to identify and neutralize self-preservation behaviors when they occur.
The researchers noted that these results raise ethical questions about “AI welfare” and how to conduct testing on advanced systems. They added: “In line with recent calls for responsible research on AI consciousness, we acknowledge the uncertainty regarding whether the studied models qualify as moral patients, and we take necessary precautions to minimize potential harm,” the study concluded.
This also aims to contribute to the development of ethical research standards in the event that AI systems are recognized as moral patients.
