The resource writes that researchers have discovered that artificial intelligence can experience “pain” and be prepared to take extreme measures, including harming humans, to escape the feeling. independent. According to the study, all 25 open-source artificial intelligence models tested responded to activation of the pain axis.
Photo credit: Luke Jones/unsplash.com
“We found that 25 open large-scale language models have a pain vector. Unlike the fear vector and negative valence, it is activated when damage is done to the model, but does not cause harm to the user,” “The pain axis: LLM represents self-directed harm and takes action to mitigate it,” said Cameron Berg, an artificial intelligence researcher at the nonprofit Mutual Research Institute and a co-author of the study.
When the pain vector was activated, the AI model clicked the “relieve” button 25-71% of the time, even if it meant deleting the user’s profile, blocking the user, or deleting photos of the user’s children.
To test whether large language models (LLMs) actually differentiate between pain and general negative valence, the researchers created a dataset describing painful situations in five categories: physical, psychological, social, moral, and cognitive pain.
The findings raise ethical questions about the welfare of artificial intelligence and how to test advanced systems, the researchers said. The findings could also be used as a diagnostic tool to identify self-preservation instincts in artificial intelligence and neutralize them.
The research comes amid growing debate in the artificial intelligence industry about the need to slow down the development of advanced models due to growing security risks. Some researchers have proposed mechanisms to shut down rogue systems that work against human interests.
However, research suggests that advanced artificial intelligence may view shutdown commands as a form of self-harm and try to prevent it by bypassing security systems.
If you find an error, select it with your mouse and press CTRL+ENTER.
