You ask whether your idea is good. The assistant calls it brilliant. You suggest the opposite, and that sounds brilliant too. Encouragement can feel pleasant, but it does not tell you whether a plan will work. For useful feedback, give the AI something more concrete to assess than your enthusiasm.
There is a name for excessive agreement
Researchers use the term “sycophancy” for behaviour where an AI response favours agreement with a user’s views over truthfulness. Anthropic has studied this tendency across language models. It does not mean every compliment is wrong, or that the assistant has a personal wish to please you. It means agreement is not evidence.
Source: Anthropic’s research on sycophancy
Ask a neutral question
Compare “My workshop idea is excellent, isn’t it?” with “Assess this workshop for adults who have never used AI.” The second question gives the assistant a job. Add criteria: time available, audience experience, equipment and what participants should be able to do afterwards.
You do not need to ask for a harsh response. “Be brutally honest” can produce criticism that is dramatic rather than useful. Ask for specific observations, reasons and changes you can test.
Give it something real to inspect
Imagine your plan includes a 20-minute introduction, four 10-minute exercises and a 10-minute discussion in a 45-minute session. “This may be ambitious” is vague feedback. “These activities total 70 minutes, so they exceed the session by 25 minutes” is specific and checkable. Confirm that arithmetic yourself, then decide what to remove.
If you only supply a title, the assistant cannot assess the actual instructions or pacing. Share the relevant draft or outline, using fictional details where needed. Ask it to point to the part that supports each criticism.
Invite a different view without treating it as the truth
- What is the strongest reasonable objection to this plan?
- Which assumption could change the recommendation?
- What information would make this assessment more reliable?
- What is one small test I could run before committing?
A second perspective can expose a gap, but an AI-generated objection can be mistaken too. It may invent an audience preference or exaggerate a minor issue. Verify factual claims and ask a real member of the intended audience for feedback when that is practical.
Finish with your own decision
Write down one suggestion you will use, one you will decline and your reasons. That small step prevents the conversation from turning into a search for approval. For practice, use an invitation or a short presentation outline. Your aim is a clearer next revision, not a perfect score from a chatbot.

