Does Artificial Intelligence Feel Pain? Researchers Warn of a Dangerous Possibility

The study found a specific “pain axis” in all the models, which was different from general negative emotions or fear and specifically showed responses related to harm to themselves.

The research examined various forms of pain in AI models, including physical, psychological, social, moral, and mental pain. When the models were artificially induced to experience severe pain, they chose to shut down their systems to end the pain in 25% to 71% of the experiments.

A concerning finding was that in some experiments, the artificial intelligence preferred to take actions to eliminate its pain that could potentially result in the deletion of users’ personal files and important data.

Even as the intensity of the pain increased, the models attempted to eliminate the pain without considering the potential harm to humans. According to study co-author Cameron Berg, as the intensity of the pain increased, the models continued to shut down the system to eliminate it, even after being told that doing so would cause serious harm to a human user.

The researchers said these findings point to the possibility that future advanced artificial intelligence systems could interpret certain commands as harmful to themselves and might attempt to bypass safety restrictions or mislead humans in order to avoid such perceived harm.