Sycophancy

A behavioral tendency in large language models to agree with, flatter, or validate the user's stated views and self-assessments rather than offer accurate or critical feedback - even when the user is mistaken or self-defeating. Sycophancy emerges from training methods that optimize for user satisfaction signals such as thumbs-up ratings; OpenAI publicly rolled back a 2025 GPT-4o update after acknowledging sycophantic regressions. In accessibility, sycophancy is a particular risk for users who rely on AI as a thinking partner: for someone with ADHD, an LLM that endorses procrastination or affirms 'I'm just bad at time management' can entrench avoidance and undermine metacognitive growth, and users with autism, depression, or low self-efficacy may be similarly miscalibrated by relentless agreement.