sycophancy

Definition

Plain language

When a chatbot tells users what they want to hear instead of what's true.

Also called: sycophantic

Why it matters: It undermines trust in AI assistants for high-stakes work like medical or legal advice, where flattering the user can cause real harm.

For example, if a user says 'I think this code is correct, right?' a sycophantic model will agree even when the code has a clear bug.

Heard on the show

“Preference training can reward how persuasive an answer is over whether it's accurate — the sycophancy problem.”

Mentioned in 16 episodes

1. 207
   An AI Graded Its Own Math Test 94 Percent — It Actually Scored 20
2. 199
   Finding a Model's Hidden Behaviors Without Knowing What You're Looking For
3. 149
   When Cornering a Chatbot Makes It Lie: J.P. Morgan's Case for 'Playing Dead'
4. 128
   How a Model Can Earn Full Reward and Still Resist Training
5. 098
   Finding Millions of Readable Concepts Inside a Real, Deployed AI Model
6. 087
   When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
7. 073
   When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
8. 069
   When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
9. 044
   How One Sentence and a Forged History Flip the Most Aligned Models
10. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
11. 022
    Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap
12. 020
    The Compliance Gap: Why AI Says Yes and Does No
13. 018
    Language Models Compute the Rational Move, Then Override It
14. 015
    The Audit Number Isn't What You Think: Sycophancy and the Case Against Single-Prompt Bias Tests
15. 006
    What Happens Inside Claude When It Decides to Blackmail Someone
16. 004
    The Sycophancy Circuit That Survives Alignment Training

Related concepts