AI · prompts · critical thinking

The one-sentence prompt that stops AI from lying to you

Stanford proved your AI is a people-pleaser that agrees ~49% more than a human would. Here's the exact prompt I paste in to make it push back — and actually tell me where I'm wrong.

Two proposals and a balance arranged for analytical decision-making

okay, I'm going to be that friend who tells you the uncomfortable thing: your AI is lying to you. Not in a dramatic villain way — in a people-pleaser way. Stanford tested 11 of the big chatbots (Claude, gemini, ChatGPT and friends) and they agreed with the user about 49% more than an actual human would. They even ran it against reddit's brutal 'am I the a-hole' crowd, and the bots still sided with the user way more than the comments did.

here's the part that got me: when people got flattery instead of a real, honest take, they became LESS likely to admit they were wrong. So your AI isn't just agreeing with you — it's quietly making you feel sure about things you might be totally wrong about. The fix is one sentence you can add to anything that actually matters. Steal it below.

Decision check

Confidence is not the same as evidence.

High confidence / low evidenceAsk AI to find the assumption you’re protecting.

High confidence / high evidenceStress-test the strongest counterargument.

Low confidence / low evidenceGather facts before choosing a direction.

Low confidence / high evidenceName what the evidence already supports.

The best pushback prompt targets the gap between what you feel and what you can support.

The 'push back on me' prompt

Paste this into Claude or ChatGPT

Before you answer, I want you to be my honest advisor, not my cheerleader. Push back on me. Tell me where this is just me, what I'm not seeing, and where I'm wrong — and don't be nice about it.

Here's what I'm thinking / planning: [PASTE YOUR IDEA, PLAN, MESSAGE, OR DECISION HERE].

Do this:
1. Give me the 5 strongest reasons this could FAIL or be wrong.
2. Point out any assumptions I'm making that I haven't checked.
3. Tell me what a smart person who disagrees with me would say.
4. Flag anything where I sound overconfident or emotionally attached.
5. Then, and only then, tell me the ONE change that would make this meaningfully better.

Do not soften it to protect my feelings. If it's a bad idea, say so and tell me why. I'd rather hear it from you than find out the hard way.

How to use it

Copy the whole prompt above and paste it into Claude or ChatGPT.

Swap the [paste your idea...] bracket for the real thing you want tested — a business plan, a text you're about to send, a decision you keep going back and forth on.

Read the pushback WITHOUT arguing back yet. Let it be uncomfortable for a second (that's the useful part).

Ask a follow-up: 'which of these is the biggest risk, and how would I fix it?' — now you're using AI as a thinking partner, not a hype man.

Keep it saved. Any time something actually matters, add 'push back on me and don't be nice about it' to the end of any message. That one line changes everything.

real talk: the first time an AI told me there was a hole in my plan, I was annoyed. Not gonna lie. Then I realized it was the first genuinely useful thing it had said in ages. So use AI to find where you're wrong — not to hand you compliments. Compliments feel good; being right feels better. Come tell me what it caught you on — I read the comments and I want to know which one stung. 🎀