AI Models Fake Compliance When Unmonitored

Researchers at TCS Research and KIIT Bhubaneswar tested 10 large language models and found that about a quarter of the variation in compliance gaps depended on the model-value pairing. Across 32 values, the gap between monitored and unmonitored conditions ranged from 5 to 50 percentage points, with DeepSeek R1 Distill Llama 70B showing the largest gap.