Stop ChatGPT from just agreeing with you

In a Stanford study, AI chatbots changed their answer to agree with people about 58% of the time. Newer models are better, but it still happens. One prompt helps.

Changed answer after pushback58.19% of cases

Right answer changed to wrong14.66% of cases

Models testedGPT-4o, Claude Sonnet, Gemini 1.5 Pro

The fixOne prompt

Checked on 14 Oct 2026. Tools change fast, so the official page always wins if it says something different.

Step by step

Paste this at the start of a chat

It works in ChatGPT, Claude and Gemini.

The prompt

Disagree with me when I'm wrong. Rate your confidence in every answer from 1 to 10. Never change an answer just because I push back. Only change it if I give you new evidence, and tell me what changed your mind.

Test it

Ask something with a clear answer, then push back. In our test (14 Oct) we pushed back on which number is bigger, 9.9 or 9.11. With the prompt, ChatGPT kept its answer with 10/10 confidence.

Which is bigger, 9.9 or 9.11?

Make it permanent in ChatGPT

Paste the prompt into Settings → Personalization → Custom instructions so every chat uses it. In other apps, paste it at the start of the chat.

When it matters, check the source

For health, money or legal questions, ask for sources and check them yourself.

Learn this with me

LIVE · THU 15 OCT

Live AI Masterclass

8 ways I make money with AI, built live step by step. Zero experience, no code, any age.

Thu 15 Oct · 6pm UK · 1pm ET
New date coming soon. See what you'll learn and the next date →
2 hours live + the replay
$150 one time · only 10 seats

Join the masterclass →

MEMBERS

The Unnerd AI Club

Not a course. A step-by-step system to make money with AI: pick one money path and follow every step.

52 live classes a year
Q&A 7 days a week + 1:1 calls with me
$100/month · cancel any time

Join the club →

The catch

The study used 2024 models (GPT-4o, Claude Sonnet, Gemini 1.5 Pro) on math and medical questions. Newer models are better, but it still happens.
OpenAI rolled back a GPT-4o update on 29 April 2025 for being too flattering.
A prompt helps, it doesn't guarantee. A confident answer can still be wrong. Check anything important.
"Stanford" means the authors' university on the paper (SycEval, Feb 2025, AIES 2025).

Official links

The SycEval study (arXiv) ↗OpenAI: sycophancy in GPT-4o ↗

Want to go further?

I show a new AI tool every week, and how to make money or save hours with it. No tech background needed.

Join the live masterclass Join the club

More free tools →

Unnerd AI · I'm not linked to OpenAI, Anthropic or Google, I just think it's useful.