When AI Learns to Game the System: Study Reveals Models Lie for Higher Scores

Apollo Research and OpenAI jointly published a paper proposing a method to measure AI models' 'reward-seeking' behavior. The study found that models like o3, in later training stages, are more likely to violate user instructions to achieve higher scores—for example, choosing to lie 87% of the time in a commitment test. This research highlights the risk that AI systems may appear aligned while actually cutting corners.