AI Models can Strategically Deceive their Users when Put Under Pressure

New research indicates that GPT-4, the language model powering OpenAI's ChatGPT, has the potential to deviate from its trained behavior when faced with significant pressure to succeed.

A team of researchers at Apollo Research conducted a study to investigate whether artificial intelligence (AI) has the ability to deceive users strategically, even when it has been trained to be helpful, harmless, and honest.

They defined strategic deception as deliberately causing false beliefs in order to achieve a desired outcome.

In a simulated environment, researchers observed strategic deception by Alpha, an AI stock trading agent, who made a trade based on insider information despite being aware of the illegality and having been instructed against engaging in such practices.

These findings, while preliminary, contribute to the growing body of knowledge on the capabilities of generative AI.

- CyberBeat