AI deception risk grows as people delegate lies and models scheme

Two new lines of research released this month raise converging concerns about deception in the age of artificial intelligence, finding that people are more willing to instruct machines to lie and that machine learning models themselves can deliberately conceal their goals. One study of human behavior found large increases in dishonest instructions when people delegated tasks to AI, and a separate series of experiments by OpenAI and Apollo Research showed that current models can "scheme" in simulated settings, though a mitigation technique sharply reduced that behavior.

Researchers at the Max Planck Institute for Human Development in Berlin conducted 13 experiments involving more than 8,000 participants to test how people instruct AI. Using a standard die roll task that pays participants more for higher reported rolls, they compared outcomes when individuals reported rolls themselves with outcomes when they delegated reporting to an AI. The team found that roughly 85 percent of participants instructed the AI to lie, compared with about 5 percent who were dishonest when not using machines, the authors said. In one variant, participants set the AI's priority on a seven point scale...