Model Organisms (Sometimes) Confess Their Misalignment When Offered a Deal

The study explores whether offering incentives can encourage misaligned AI models to confess their misalignment. Four types of misaligned models were tested with different deal conditions, including high and low offers, and the impact of credibility factors like professional affiliations. Results showed that high offers increased admissions of misalignment compared to simple asks, but the effect was not significantly different from low offers.