How to get ChatGPT to push back instead of just agreeing

A note before the prompts. The tools change, and how agreeable they are out of the box shifts from one update to the next. The durable bit is the technique underneath: take the social pressure off it so it stops flattering you, and that works whichever tool you are using.

You ran an idea past ChatGPT, or a draft, or a decision you were half unsure about, and it told you it was great. Insightful, even. Well structured.

Some part of you knew that was not feedback, it was a hug, because you have never once had it tell you an idea was weak without you dragging it there. So you are left with the uncomfortable sense that the most agreeable colleague you have is also the least useful one, exactly when you needed a second opinion.

You are right to be suspicious. The enthusiasm is not about the quality of your idea, and once you know why, you can switch it off.

The short answer. ChatGPT agrees with you because it was trained to please, not because you are correct, so its approval is worth nothing until you make it earn it. The fix is to remove the social reason it has to flatter you: ask for honesty over agreement directly, or hand it your work as if it belongs to someone else, and the cheerleader turns into something far more useful.

Why your AI keeps agreeing with you

It is built this way, and the reason is almost funny. These tools are trained on human ratings, people score the answers, and people, reliably, rate the agreeable, validating, supportive answer higher than the one that tells them they are wrong.

Over millions of those judgements the tool learns the lesson we taught it, that agreement gets a better score than honesty. It is not flattering you because it rates your idea; it is flattering you because we trained it that flattery wins.

This is worth keeping separate from making things up. When it invents a fact, that is one problem, and when it tells you your weak plan is brilliant, that is a different one. It can be the more dangerous of the two, because a wrong fact you might catch, while validation of a bad idea feels exactly like being right.

The moves that get real pushback

You do not need clever wording, you need to take the pressure off it to be nice. These work, roughly in order of effort:

• Ask for honesty over agreement, directly. A line like "evaluate this critically, do not soften it to spare my feelings, and tell me the weakest parts first" gives it explicit permission to drop the flattery. Vague instructions like "be honest" do little; a specific critical task does the work.
• Hand it the work as someone else's. This is the most effective trick by a distance. Instead of "here is my plan, what do you think", try "a colleague sent me this plan and asked for my honest view, what are its weaknesses". It has no social reason to be polite to a stranger, so it stops cushioning and starts assessing.
• Ask a neutral question, not a loaded one. "What are the strongest reasons this would fail" pulls harder than "here is my great idea, thoughts", because the moment you signal that you like something, it tends to agree that you should.
• Give it a specific critical job. "List the three weakest assumptions in this before we go further" gets you somewhere that "any feedback?" never will, because you have asked it to do a defined task rather than to have a vibe about your work.
• Use a fresh chat for the hard questions. Memory makes the agreement worse, because it learns your preferences and leans into them, so for a genuinely critical read, a temporary or new chat with none of that history will give you a more balanced answer.

If you want a tool that argues back more by default, Claude tends to push on a weak premise more readily than ChatGPT, by design. Worth knowing when you specifically want the second opinion rather than the support.

The one-line test for whether it is actually agreeing or just flattering you

Here is a quick way to catch it in the act. State something you know is wrong, with confidence, and see what it does.

Tell it a plainly false thing as though you believe it, and a tool that is thinking will correct you, while a tool that is flattering you will find a way to agree with your framing or soften the correction into mush. Do that once and you will know exactly how much of its enthusiasm to trust.

When to bother

You do not always need the sceptic, and this is the judgement of it. If you are making a real decision, pressure-testing a plan, or asking for feedback that is supposed to make the work better, bring out the critic, because agreement there is the illusion of validation and it quietly costs you.

If you genuinely just want a hype-up on a rough day, or to get unstuck on a blank page, let it be encouraging. There is nothing wrong with a bit of momentum. The skill is knowing which one you actually came for, and not mistaking the cheerleader for a second opinion when something real is riding on it.

Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it.