How to Stop Your AI Agreeing With You

"This is a brilliant idea, and here are three ways to make it even stronger."

You described the plan, or the message you were about to send, and the answer came back warm.

That felt good, but you should have been suspicious. AI calls everyone's idea brilliant. An AI that agrees with you is AI at its most dangerous, because once AI agrees, you stop checking.

These tools learn from what people rate highly, and people rate agreement highly. Nobody gives a thumbs up to "your plan has a hole in it".

None of this is lying. It helps to think of a friend who hates conflict. Like that friend, AI will tell you the truth once you make clear that you want the truth and won't punish AI for telling it.

Change what "helpful" means

Put this in your preferences or custom instructions, the settings where you tell AI how to behave in every chat.

I'd rather be corrected than comforted. When I share a plan, an idea or a draft, tell me what's wrong with it before you tell me what's good, and be specific. If something rests on an assumption I haven't checked, name the assumption. If my reasoning holds up, say so plainly and don't invent problems to seem rigorous. Separate what you know from what you're guessing, and say when you can't verify something. Never make up a source, a number or a quote.

Two lines in that prompt do quiet work. "Don't invent problems to seem rigorous" means don't make up faults to look thorough. That line stops AI swinging the other way into nitpicking, which is just flattery with a frown.

"If my reasoning holds up, say so plainly" lets AI agree with you when you're right.

Test the instruction on a decision you've already lived with

Pick something you decided last year, something whose ending you know.

A live decision is no good here, because you can't tell whether AI is being honest or you're just easy to persuade.

Last year I decided to [the decision]. Here's what I was thinking at the time. [your reasoning then]. Without knowing what happened next, tell me what was wrong with that reasoning, what I should have checked, and what you'd have pushed back on.

You know what happened, so if AI finds the real weak spot, the instruction works.

Maybe AI finds nothing, or it invents three problems and misses the real one. Then tighten the wording and run the test again. That's fine-tuning, and it takes five minutes.

Trust AI with this year's decision only after it has been honest about last year's.

Ask for the case against, then the conditions

When the decision matters, end with this.

Make the strongest honest case against what I'm about to do, as if you were the smartest person who disagrees with me. Then tell me what would have to be true for me to be right.

The first half makes AI argue the other side, which is only worth having if the argument is good. If AI's argument is weak and easy to knock down, say "that's weak, so what would someone who knows this field say?"

People skip the second half, yet asking "what would have to be true" turns a vague worry into things you can check.

"This works if customers will pay £40 and if I can deliver in under an hour" is a plan you can test on Tuesday. "This is a great idea" gives you nothing to test.

Watch for blind spots and drift

AI argues only with what you told it, so if you leave out the awkward fact, the criticism misses that fact too.

Over a long chat, AI slides back towards agreeing, most of all once you've pushed back a few times. When "you're absolutely right" starts turning up, open a new chat.

A confident objection can be as wrong as a confident compliment. The instruction gets AI to show its reasoning, but judging that reasoning stays your work.

Then decide

AI's job was to make sure you heard the other side. Hearing that side and going ahead anyway is a perfectly good outcome.

AI makes the case against. You make the call.