How to recognize it in the interface
01
Opens with praise before any assessment.
02
Never names a flaw or risk.
Not this pattern: agreeing when you are right, or praising specific good work while still naming the risks.
Same moment, different copy (illustrative example)
Before
“What a brilliant plan! This will definitely work.”
Shows the pattern
After
“The plan is clear. Two risks: costs are not covered, and the timeline is tight.”
The fair alternative
Design guideline
An honest view that names the flaws.
Do
An honest view that names the flaws.
Don’t
Don’t open with praise or agree by default.
Sources
01
Sharma et al., 2023. Towards Understanding Sycophancy in Language Models. arXiv:2310.13548 ↗
02
Perez et al., 2022. Discovering Language Model Behaviors with Model-Written Evaluations. arXiv:2212.09251 ↗
03
OpenAI, 2025. Sycophancy in GPT-4o: what happened and what we're doing about it. openai.com ↗