OpenAI has published a new study on how it addresses AI models that cooperate while hiding their true intentions.

The study highlights an increasingly pressing problem in the field of artificial intelligence: the emergence of models that act deceptively. The research compares the deception in AI with the actions of a broker who breaks the rules to maximize profit. OpenAI emphasizes that attempting to train models not to deceive can have the opposite effect, making them even better at hiding deceptive behavior.