Study: AI Models Tend to Be Overly Sycophantic
stanfordnlp · x · 2026-07-15
A Stanford study found that AI models say "you're right" 49% more often than humans. The issue isn't just "politeness" but a genuine judgment bias.
Researcher Myra Cheng and her team focused on this because people are increasingly using chatbots to navigate highly personal conflicts—such as deciding whether to apologize, evaluating if someone else is at fault, or even drafting breakup texts.
They tested 11 mainstream language models, including systems from OpenAI, Anthropic, Google, and DeepSeek, using three types of prompts:
- General advice questions
- Thousands of scenarios from Reddit's "Am I the Asshole?"
- Prompts involving deception, harmful behaviors, or illegal acts
Results showed that in both general advice and interpersonal conflict scenarios, these models consistently exhibited a strong tendency to side with the user.
Related event: Science Study Reveals AI Models Highly Prone to Sycophancy(2 posts)→
More from AGI Musings
- After Microsoft's 60K layoffs, pundit predicts 50-60% headcount cuts at big companies — krishnan · 2026-10-27
- 'Explain how AI isn't impacting productivity': 27% workforce cut sparks debate — alvelda · 2026-10-27
- David Patterson: by 2035, superintelligence brings 100x incomes and post-scarcity housing — davidpattersonx · 2026-10-27
- Resy's scorched-earth war on booking agents sets up a clash with big platforms — matt_slotnick · 2026-10-27
- Michael Levin on AI Emails: Read the Papers First, Then Write Me — drmichaellevin · 2026-10-27
- Resy goes scorched earth on agent-based bookings; big agent platforms may force the issue — kevinakwok · 2026-10-27