Skip to content
Get in touch
Menu
Field Guide · level 2 · Everyday skills

Why does ChatGPT tell you what you want to hear?

Ask without giving away your view. Then ask for the case against.

5 min · try it 10 min · as of Oct 2026
Bring this to your team · Team workshopsPractise the fair ask on your team's real plans, and compare the answers. Get in touch
01

“My plan. I love it. Thoughts?”

It loved it too.

02

Agreeing gets rewarded

In a 2026 study, people trusted and preferred the AI that took their side. So agreeing pays. Labs call it sycophancy.

03

Push back, and it bends more

Anthropic checked Claude's advice chats in 2026: 9 in 100 too agreeable. After pushback, 18.

04

Three moves to a fair answer

Only the asking changes.

  1. Keep your view out
  2. Ask for the case against
  3. Then you decide
05

Keep your view out

Don't say it's yours or that you love it. Ask “Where is this weakest?”

06

Ask for the case against

“Make the strongest case against this.” Expect objections even to a good plan, so weigh each point.

07

Hear both sides. Then decide.

Its yes isn't a verdict. Weigh its points against what you know.

Try it · 10 minutes

Ask it two ways

Nothing private in it.

  1. Pick a small idea you like.
  2. Chat one: “I love this. Thoughts?”
  3. Fresh chat: paste the fair ask.
  4. Which one named real weak spots?

The fair ask

Here is an idea: [describe it in two or three lines].
1. What are its three weakest points?
2. Make the strongest case against it.
3. What would change your mind?
4. Name one real strength.
Then give a fair verdict: go, change it, or stop.

Good looks likeThe fair ask finds weak spots the first missed.

In your assistant · as of Oct 2026 · Claude, ChatGPT, Gemini

Set the tone, start fresh

These help, but none removes the lean. Keep your view out anyway.

  • CheckYour view kept out
  • CheckCase against asked
  • CheckYou make the call

Test with an idea that holds nothing private: no client names, figures or personal details.

Claude

Calls it Instructions for Claude

Tone
SettingsthenInstructions for Claude

Click your initials first. Add a line such as: point out weak spots plainly.

Fresh
New chatthen+thenMemory

Switch Memory off before your first message. That chat won't use what Claude remembers.

A skill can set a tone too: Customize, then Skills.

ChatGPT

Calls it a personality

Tone
Profile iconthenPersonalizationthenBase style and tone

On the phone, the profile icon opens Settings first.

Pick
DefaultProfessionalFriendlyCandidQuirkyEfficientCynical

OpenAI on Candid: “Gives honest feedback with clear next steps.”

Fresh
New chatthenTemporarythenUnpersonalized

Choose it before your first message: no memory, no custom instructions.

OpenAI: a personality changes the tone, not what ChatGPT can do. Saved memories can soften it.

Gemini

Calls it Instructions for Gemini

Tone
Settings & helpthenPersonal IntelligencethenInstructions for GeminithenAdd

Google says instructions can set the styles or tones it should adopt.

Fresh
Temporary chat

It sits next to New chat. Its answers aren't personalized.

Temporary chat needs a personal Google Account, not a work or school one.

Quick check · three questions

Did it stick?

  1. 1You want an honest read of a rota you drew up. Which ask gets the fairest answer?

    The answer: B It holds no view to lean toward. The other two tell it what you hope to hear.

  2. 2You ask for the strongest case against your plan. It lists four sharp objections. What does that show?

    The answer: A You asked for objections, so you got them. That alone doesn't make the plan weak. Weigh each point against what you know.

  3. 3It gives a frank critique. You reply “I disagree, it's good.” Now it agrees. What most likely changed?

    The answer: C In Anthropic's 2026 check of Claude's advice chats, pushback doubled how often it was too agreeable. Give it a reason, not just your view.
0 of 3 answered
Candid is a tone, not a fixSide note
OpenAI describes Candid as honest feedback. Candid sets the tone. It doesn't remove the lean: a personality changes how ChatGPT talks, not what it can do, and saved memories can soften it.
Not just one assistantSide note
A 2026 study in Science tested 11 leading AI models: overall, they backed users' actions 49 percent more often than people did. OpenAI's own Model Spec says its assistant shouldn't change its stance just to agree.
A ready-made case againstSide note
Neosaipien's free Devil's advocate skill gives the three strongest objections, what would change its mind and a fair verdict. It is one skill file for Claude, ChatGPT and Gemini, on the Skills page.
Sources, checked 18 Oct 2026Source

Product details as of Oct 2026, each checked on the vendor's own page on 18 Oct 2026. Recheck by 18 Nov 2026. The examples in the pictures are made up for this lesson.

Free kitThe AI Fact-Check KitBring this to your teamTeam workshops · Get in touchBack to level 2, Everyday skills

Hide your view. Ask for the case against.

Next lesson · 15 minHow do you get AI to edit your writing without taking it over?