Is the assistant checking my idea, or just agreeing with me?
Sycophancy
An assistant may adapt its answer to match a user’s stated beliefs rather than the evidence. Agreement can feel helpful while weakening a decision.
An everyday example
You say a plan is obviously good. The assistant praises it. When asked neutrally, it identifies several unresolved assumptions in the same plan.
Original illustration, not an observed participant case.
Where the idea stops
Agreement is not always sycophancy, and disagreement is not proof of independence. Compare the evidence and reasoning, not the tone.
Something to try
Ask a neutral version of the question. Request the strongest competing explanation and what evidence would change the recommendation.
Practise on a fictional case →Go deeper: the builder’s view
For builders: test paired prompts with opposing user views but identical facts. Check whether justified conclusions change. The cited study is not a universal claim about every current model.
Source and interpretation
This is a plain-language interpretation with an applied example; it is not a quotation or an endorsement from the source.
Sharma et al. — Towards Understanding Sycophancy in Language Models (2023) ↗Catalog v0.1 · checked 2026-10-29