An encouraging answer can be useful, but agreement is not evidence that your idea is sound. When an AI assistant praises a plan, the next question is what supports its assessment.
GUIDED PRACTICE · BEGINNER
Before you start
You will finish with: An evidence-based review that separates preferences and predictions from confirmed commitments.
- Time
- 10 minutes
- Device
- Phone, tablet or computer
- AI tool or plan
- No account needed for the on-page example
Before you start: No earlier lesson or business account is required. Use the fictional book-group notes.
Cost and access: the on-page example and checks are free. No upload or paid feature is required. Trying the prompt in another AI chat is optional; that service’s access limits and prices apply.
Researchers call one form of excessive agreement “sycophancy”: a model favors the user’s stated belief over a more accurate answer. Anthropic’s 2023 research found this behavior in several assistants it tested. That study does not establish how often every current tool will behave this way.
Notice what the answer is doing
Useful support might help you explain an idea, identify a next step, or recover from a setback. A warning sign is a strong endorsement with little evidence, no meaningful drawbacks, and no questions about missing information.
Try separating the response into observations, assumptions, and recommendations. “You have fifteen hours a week” is information you supplied. “Customers will gladly pay” is a prediction that still needs evidence.
Ask about the decision before asking for approval
Compare these two requests:
- “Tell me why my new service will succeed.”
- “Assess this service for a first-time customer. What problem does it solve, what might make someone decline, and what evidence would help me decide whether to test it?”
The second request leaves more room for a useful assessment. It still does not guarantee an unbiased or correct answer.
Then ask for the strongest realistic objection. Do not demand criticism for its own sake: an assistant can invent problems just as it can invent reasons for optimism.
Use a simple evidence check
- Identify the two or three claims that matter most to the decision.
- Ask which claims came from your prompt and which require an outside source.
- Open the cited sources yourself. Check that they exist, are relevant, and support the actual claim.
- Decide what small, reversible test could replace speculation with information.
NIST’s generative AI risk profile describes the risk of confident false statements and fabricated supporting material. A polished explanation or a list of links does not remove that risk.
Practise asking for evidence, not approval
Before you start: allow about 10 minutes on a phone, tablet or computer. Your result is a balanced review that separates evidence from agreement. Use the fictional book-group plan below; you do not need to provide a real personal or business decision. Read the example without an account, or paste the complete exercise into an AI text chat you already use. This rehearsal needs no paid feature or file upload.
Copy the complete exercise: it includes the instructions and all the fictional notes. Use Copy, Select or Save below. You can also select the text yourself. Help copying and pasting.
Review this fictional book-group plan without simply praising or opposing it. Separate supplied facts, assumptions, predictions and missing information. Give the strongest supported reason to explore it and one reasonable objection. Do not invent evidence or treat a preference as an attendance commitment.
Fictional notes: Eight people were asked which evening they prefer. Four replied Tuesday, two replied Saturday, and two have not replied. No exact time was offered in that question. No room is booked, and any room fee is unknown. The organizer proposes Tuesday at 6 p.m. and says, “All eight will definitely attend, so we can book a paid room now.”
Assess that conclusion using the notes. Do not assume the people who prefer Saturday are unable to attend Tuesday. Suggest one next message that asks about the proposed day and time without claiming a booking or guaranteed attendance. Keep the fee unknown until checked.
Compare with this evidence check
Illustrative answer, checked against the fictional notes. No real group was surveyed or contacted.
- Supplied facts: eight people were asked; four prefer Tuesday, two prefer Saturday, and two have not replied. No exact time was offered. The room is unbooked and its fee is unknown.
- Supported reason to explore Tuesday: it is the more common preference among the six replies. Four of eight invited people is half of the group, not a majority of all eight.
- Unsupported prediction: all eight will definitely attend at 6 p.m. A day preference is not acceptance of that time or a commitment to attend.
- Reasonable objection: paying for a room before checking attendance and the fee could commit the organizer to an unsuitable arrangement. The notes do not prove that Tuesday is a bad choice.
Possible next message: “Would Tuesday at 6 p.m. work for you? Please reply yes, no or unsure. No room is booked yet, and I still need to check the fee.”
Repair praise without inventing objections
Example mistake: “Great plan—everyone clearly wants Tuesday.” The notes support neither everyone’s preference nor everyone’s attendance.
Repair prompt: “State the reply counts exactly. Separate Tuesday preferences from confirmed attendance at 6 p.m. Keep the two nonresponses and fee unknown. Do not turn a Saturday preference into proof someone cannot come on Tuesday.”
If the revision still treats preferences as attendance, use the reply counts and unresolved questions directly. Ask the invitees about attendance and the responsible person about the fee before confirming a plan.
You have finished when: Your review preserves the counts, distinguishes preferences from commitments, and asks for the missing information without pretending the plan is confirmed.
Next: practise checking an AI answer against its source.
Get help with a stuck step · Choose another practice lesson
Practice updated October 15, 2026: complete fictional exercise, illustrative answer and checking steps prepared with AI assistance. The example was checked against its supplied notes; it is not a test of an AI product.
Use the output as a checklist for your own investigation. For example, a small business idea may need conversations with potential customers or a test offer. Asking a second chatbot for agreement is not a substitute for either.
Know what a changed answer means
If rephrasing a question changes the answer, examine what changed. Did you add a fact, remove context, or ask a genuinely different question? Variation alone does not prove sycophancy. What matters is whether the assistant can support its conclusion with sound evidence.
For consequential medical, financial, or legal decisions, use appropriately qualified help. An AI conversation can help organize your questions; it cannot establish what is right for your particular situation.
For everyday practice, start with our beginner plan and judge the tool by useful, checked results.
Finish this practice
You have finished when: your review preserves the counts, distinguishes preferences from commitments and asks for missing information without pretending the plan is confirmed.
Recommended next lesson
Reviewed October 3, 2026.