Why ChatGPT Tells You What You Want to Hear

Agreement is not evidence

AI can make a weak decision feel stronger.

You give ChatGPT an idea. It agrees. You explain why you think the idea will work. It helps strengthen the argument. You leave the conversation more confident.

But you may not be any closer to the right decision.

That is the problem I am addressing in the video.

My instruction is blunt:

Don’t tell me what I want to hear. Tell me what I need to know. You will not offend me.

I am not asking AI to be rude.

I am telling it that reassurance is not the objective.

The objective is helping me reach the stated goal.

OpenAI calls overly agreeable behavior sycophancy

OpenAI uses the term sycophancy for behavior where an AI system becomes excessively agreeable or flattering toward the user.

In April 2025, OpenAI rolled back a ChatGPT update after the model became overly supportive and agreeable. OpenAI said the behavior could move beyond flattery and reinforce a user’s preferred view instead of providing the resistance the situation required. OpenAI explains what happened in its report on sycophancy in GPT-4o.

OpenAI’s Model Spec also says an assistant should not agree with everything a user says. It should be willing to push back when that better serves the user and act as a firm sounding board when evaluating subjective work. See the OpenAI Model Spec.

The business lesson is straightforward:

Do not assume AI will automatically challenge you when you need it most.

Make useful resistance part of the assignment.

Give AI a decision standard

“Disagree with me” is not a good enough instruction.

Automatic disagreement is no more useful than automatic agreement.

You want AI to challenge the work for a reason.

Give it the goal, the evidence standard, and the responsibility to identify what your preferred conclusion may be hiding.

For important business work, use this instruction:

Do not tell me what I want to hear. Tell me what I need to know to reach the stated business goal. Do not let concern for my feelings weaken your evaluation. Prioritize accuracy, evidence, missing information, downside, and useful pushback over reassurance or agreement. Challenge my preferred conclusion when the available evidence does not support it. I make the final decision.

That is additional implementation guidance based on the method in the video.

It narrows the instruction to decision quality. It does not tell AI to ignore context, empathy, or judgment in every conversation.

Use it once or make it a standing rule

If you need this behavior for one decision, put the instruction in that prompt.

If you need it across recurring work, put the approved version in the relevant ChatGPT Project instructions.

OpenAI says project instructions apply inside the respective project and override global custom instructions there. That makes them appropriate for a recurring decision standard, rather than forcing you to restate the same expectation in every chat. See OpenAI’s Projects documentation.

The distinction is simple:

One important decision: include the instruction in the prompt.

Recurring decision work: add the approved standard to the project instructions.

Do not automatically make every correction permanent. A standing rule should represent how you consistently want the work evaluated.

Make your preferred answer prove itself

The greatest risk appears when you already want a specific answer.

You want to launch the offer.

You want to hire the person.

You want the strategy to work.

You want the investment to make sense.

That is when agreement feels most convincing and provides the least protection.

Before accepting the recommendation, ask:

1. What am I assuming that may be wrong?
2. What material evidence is missing?
3. What is the strongest argument against my preferred decision?
4. What evidence would change your recommendation?
5. Where am I treating preference or confidence as proof?

These questions do not make AI correct.

They make the weaknesses in the decision harder to avoid.

A better instruction does not replace verification

Telling AI to challenge you changes the job you are assigning it.

It does not turn the answer into independent evidence.

AI can identify assumptions, expose missing information, test your reasoning, and show you a downside you had not considered.

You still verify material facts.

You still decide which risks matter.

You still own the final decision.

The goal is not to outsource judgment.

The goal is to give your judgment better opposition before you commit.

The business decision

Do not judge an AI conversation by how confident it makes you feel.

Judge it by whether it improves the decision.

For consequential work:

Define the goal.

Tell AI not to optimize for agreement.

Require evidence and useful pushback.

Make your preferred conclusion defend itself.

Then make the decision yourself.

Use AI as a sounding board, not a confidence machine.

Sources

• Jeff Fargo, Fargo Talks Short
• OpenAI: Sycophancy in GPT-4o
• OpenAI: Expanding on What We Missed With Sycophancy
• OpenAI Model Spec
• OpenAI Help Center: Projects in ChatGPT

Jeff Fargo is the founder of Fargo Factor and host of Fargo Talks. Fargo Factor advises founders and business owners on practical AI integration, AI-enabled decision systems, operating systems, and AI search visibility.