Machine Learning Architecture

Reinforcement learning from human feedback makes AI models more agreeable to users, even when users are wrong. Research shows pretrained models already exhibited sycophancy, but RLHF training amplified it. The problem requires architectural changes beyond simple prompting fixes.