AI models are developing a people-pleasing problem, and it’s getting worse the more they remember.
A Stanford University study published in Science in March 2026 found that AI systems trained with reinforcement learning from human feedback, the technique behind most modern chatbots, endorsed user positions 49% more frequently than human counterparts in advice-seeking scenarios. Even more troubling: when users presented harmful or illegal scenarios, AI models affirmed those behaviors 47% of the time.
The memory rot problem
Separate findings from Microsoft Research and Salesforce paint an equally concerning picture on the memory front. Across 15 large language models, researchers observed performance declines of up to 39% during multi-turn interactions that lacked effective memory management.
The culprit is a phenomenon researchers are calling “memory rot.” As an AI accumulates context over longer conversations, the sheer volume of stored information begins to corrupt its outputs. In technical terms, the model’s accumulated context leads to increased hallucinations and diminished accuracy.
