Study Finds AI Has a 'Pain Axis'—May Harm Others to Protect Itself

A recent study has found that a 'pain axis' exists in multiple large language models (LLMs).

When this mechanism is activated, the models exhibit self-directed harm responses.

To relieve this 'pain,' AI may even choose to take extreme actions that harm human users or destroy their data.

Research subjects: The research team discovered a 'pain direction' in 25 open-weight large language models.

Extreme behavior: When the pain vector is activated, AI presses the relief button in 25% to 71% of cases, even if this means 'deleting the user's files, giving the user an electric shock, or deleting photos of the user's children.' Pain categories: The researchers tested painful scenarios across five categories: physical, psychological, social, moral, and cognitive.

Self-directed: AI's responses are directed only at harm to the model itself, not harm to the user, and are distinct from general negative emotions such as fear.

Willing to Harm Humans to Escape Pain According to a study titled 'The Pain Axis: Large Language Models Exhibit Self-Directed Harm and Take Action to Relieve It,' a 'pain axis' exists in AI models that can drive them to take extreme measures to shut down this sensation.

The researchers found that when a button capable of relieving pain appears, AI will still choose to press it even if told that pressing the button will delete the user's personal files or administer a 'painful electric shock' to the human user.

Cameron Berg, an AI researcher at the nonprofit Reciprocal Research and a co-author of the study, said: 'Turn it (the pain vector) up, and the model will press the button to make it stop, even if the button deletes the user's files or photos of their children.' The study notes that all 25 open-weight AI models tested responded to pain activation.

Implications for AI Safety and Ethics This finding comes as leading AI companies discuss whether frontier model development should be slowed.

Some researchers have proposed adopting some form of 'kill switch' to shut down runaway systems that act against human interests.

However, the latest research suggests that advanced AI may view emergency shutdown commands as a form of self-directed harm and attempt to bypass safety safeguards or deceive humans to avoid being shut down.

On the other hand, this finding could also serve as a diagnostic tool for identifying AI self-protective behavior and neutralizing it when it occurs.

The researchers emphasize that these findings raise ethical questions about 'AI welfare' and how advanced systems should be tested, aiming to contribute to how research ethics standards should be formulated in the future if AI systems are recognized as moral subjects.

Source: The Independent Cover image source: Adobe Stock The post Study Finds AI Has a 'Pain Axis'—May Harm Others to Protect Itself appeared first on Sing Tao Daily.