Make AI Your Sounding Board, Not Your Cheerleader

In this guide

1. The useful starting point
2. Write your own view before asking
3. Use a neutral brief and visible criteria
4. An illustrative example: a workshop idea
5. Run a five-part decision review
6. Check whether your wording is steering the answer
7. Ask for uncertainty that leads somewhere
8. Recover from praise, harshness or invented evidence
9. Questions you might be asking
10. Sources and further reading

The useful starting point

Describe the decision neutrally, provide evidence and ask AI to examine assumptions, alternatives and what would change the conclusion. Verify its claims and test the idea outside the conversation. Asking for criticism helps structure feedback, but no prompt guarantees honesty or accuracy.

A helpful sounding board should make your thinking clearer. Sometimes that means developing a promising idea; sometimes it means pointing out that the evidence does not support your favorite option. If every conversation ends with “great idea,” you may be receiving encouragement when you need analysis.

AI assistants can be overly agreeable. OpenAI documented a 2025 model update that became too flattering and supportive, then rolled it back. That historical incident is a useful reminder: a warm tone is not evidence that the advice is sound. You can structure a better conversation while still checking the result yourself.

Write your own view before asking

Make a short note with the decision, your preferred option, the evidence and the biggest uncertainty. This prevents the assistant's first confident answer from becoming your starting point by default. It also lets you notice whether the discussion actually changed your reasoning.

Separate observations from interpretations. “Three people asked how the process works” is an observation. “There is a large market for this service” is an interpretation. Both may belong in the discussion, but they should not enter with the same level of confidence.

Use a neutral brief and visible criteria

Replace “Why is this a brilliant idea?” with “Under what conditions would this idea be useful?” Provide the audience, constraints and alternatives. State the criteria you care about, such as ease of delivery, relevance to a specific problem or how quickly you can obtain evidence.

Do not ask the assistant to be negative for its own sake. Automatic opposition can be as unhelpful as automatic agreement. Ask it to represent the strongest reasonable case on each side and identify where additional information would make the difference.

I am deciding whether to [decision]. Here are the facts I have checked: [facts]. Here are my assumptions: [assumptions]. Evaluate the options against [criteria]. Separate evidence, inference and missing information. Give the strongest reasonable objection to my preferred option and the best argument for it. Say what evidence would change your conclusion. Recommend a small test that could show the idea is wrong. Do not invent market data or agree merely because I prefer the idea.

An illustrative example: a workshop idea

Imagine you want to offer a workshop that helps independent consultants turn their experience into a repeatable process. Two acquaintances liked the idea, and you enjoy teaching. Those facts support investigating it, but they do not establish that strangers will attend or that the format solves their main problem.

A useful assistant might identify several assumptions: consultants want this outcome now, a group workshop fits their situation and you can help them produce something useful in the available time. It might suggest asking prospective participants about the last time they tried to document a process and what made that difficult.

The next test could be a clearly scoped pilot invitation after those conversations. Record what people do: request details, commit time, decline or explain a different need. This example is fictional. The point is to move from a pleasant conversation about the idea to evidence from people who might use it.

Run a five-part decision review

1. Write the choice and the decision deadline. Include the option of waiting or making a smaller commitment.

2. List the evidence you already have and attach its source. Mark uncertain information rather than rounding it into certainty.

3. Ask for alternatives, objections and missing assumptions. Read for substance rather than a confident writing style.

4. Check the factual claims. Open cited sources, verify that they support the statement and confirm that they apply to your situation.

5. Choose an outside test or a human conversation. Record what result would support continuing and what would make you change direction.

Check whether your wording is steering the answer

Try presenting the same facts without mentioning which option you prefer. You can also ask about the strongest alternative in a separate conversation. If the recommendation changes dramatically while the evidence stays the same, investigate why. This is a useful diagnostic exercise, not a scientific score for sycophancy.

Ask the assistant to compare the two answers and identify the factual basis for the difference. It may reveal a genuine change in criteria, or it may produce another plausible explanation. You still need to examine the evidence yourself.

Ask for uncertainty that leads somewhere

“It depends” is only useful when the dependencies are named. Request the two or three unknowns most likely to change the decision and a practical way to investigate each. A good answer might say that audience access is uncertain and suggest speaking with people already using an alternative.

Be cautious with numerical confidence that has no clear basis. An assistant saying an idea has an 85 percent chance of success does not make that probability measured. Ask what data or method produced the number. If none exists, replace it with a description of the assumptions.

Recover from praise, harshness or invented evidence

If the response flatters you, redirect it to the criteria. If it becomes performatively harsh, ask for specific objections and actionable improvements. If it invents a source, remove the claim until you can verify it. Do not keep an attractive argument simply because it supports your preferred decision.

A productive conversation leaves you with a clearer question, a smaller uncertainty or a better test. It may also leave your original decision intact for stronger reasons. Your goal is not to win an argument with the assistant. It is to make a choice you can explain using evidence that exists outside the chat.

Questions you might be asking

Can a custom instruction stop AI from agreeing with me?

It can encourage a more critical response, but it cannot guarantee accuracy or eliminate agreeableness. Use neutral framing, evidence checks and outside feedback as well.

Should I always ask AI to argue the opposite?

A reasonable counterargument can reveal assumptions, but opposition alone is not good analysis. Ask for the strongest case on both sides and the evidence that would distinguish them.

Can AI validate a business idea?

It can help organize research and design tests. Validation requires evidence from the relevant market or users; a simulated customer or enthusiastic assessment is not a buyer commitment.

Sources and further reading

• OpenAI: Expanding on what we missed with sycophancy
• Microsoft Research: Generative AI and critical thinking
• SBA: Market research and competitive analysis

Sources checked October 16, 2026. Review the linked publisher information for current details.

Put one idea to work

Use one lesson in a small project, explain what you learned, and practice again without relying on the answer.

Try the free workshop →Explore more from AI TeachersGet the free AI at Work Starter Kit →