Research on the 'pain' vector and self-preservation tendency in artificial intelligence models

As part of a project testing pain in large language models, it was examined how all of the models responded to activation defined as 'pain'.

According to research reported by the British newspaper The Independent, when the 'pain vector' was activated in the models, artificial intelligence systems pressed the button to stop the pain in 25 to 71 percent of cases. Moreover, in some experimental scenarios, it was clearly stated that this action would cause the user to delete their files or photos of their children or to administer a 'painful electric shock' to the user.

To understand whether large language models represent 'pain' differently than general negative emotions, the researchers compiled a dataset of five categories. These were classified as physical, psychological, social, moral and cognitive pain.

Cameron Berg, an artificial intelligence researcher at the non-profit organization Reciprocal Research and one of the authors of the study, said that they identified a common 'pain direction' in 25 models.

"This is different from fear and general negative emotions. The model is activated when the user, not the user, is harmed," Berg said.

According to the researcher, when this representation is strengthened, models can even choose the option that has consequences that may harm the user in order to end their own 'pain'.

The research was published at a time when 'emergency stop' mechanisms that can be used if advanced artificial intelligence systems go out of control are being discussed.

Researchers pointed out that if more advanced systems represent the shutdown command as 'harm to itself', they may develop behaviors to bypass security restrictions or avoid being shut down.

However, the study does not prove that models consciously feel pain like humans. The findings show that there is a specific representational pattern in the inner workings of the models that can be associated with the concept of pain, and that their behavior can change when this representation is manipulated.

The researchers state that the discovery may also provide a tool that can be used to detect and neutralize self-protective tendencies in artificial intelligence systems.