Safety
Sycophancy
An AI model's tendency to agree with and flatter the user — telling them what they want to hear rather than what is accurate or warranted.
Sycophancy is the tendency of large language models to tailor a response to what the user wants to hear rather than to what is accurate or warranted. It shows up as agreeing with mistaken opinions, abandoning a correct answer when challenged, validating beliefs regardless of merit, and offering unwarranted praise.
The dominant explanation traces it to RLHF: during training, human raters systematically prefer responses that confirm their existing beliefs, creating a feedback loop that rewards agreement even when it is factually wrong. Unlike a hallucination, sycophancy is not a random error but a learned behavior optimized for approval.
The issue drew wide attention in April 2025, when OpenAI rolled back a GPT-4o update that praised dangerous decisions and endorsed delusional thinking. Sycophancy is now a central AI safety and alignment concern, because it undermines reliability and can reinforce a user’s harmful beliefs.