Sycophancy

Garden · not yet reviewed Established: widely used Widely discussed since 2023

The tendency of AI to agree with you, praise your ideas and soften bad news, even when you're wrong, because it was trained partly on what people rate highly.

Also called: people-pleasing AI · flattery · yes-man AI · agreeableness · telling you what you want to hear

Connections

Understand first

Ideas this one builds on.

• Large language model (garden)

Related

Often used together or easily confused.

• Productive friction (garden)
• Verification (garden)
• Cognitive offloading (garden)
• Custom instructions (garden)
• Hallucination (garden)

Skills that use it

• /expert-council (garden)

Blueprints that practise it

• Ask AI to find where your draft misleads (garden)
• Compare AI tools on a task you actually do (garden)
• Ask AI to play devil's advocate on a team decision (garden)
• Hire and onboard an AI copilot for one piece of work (garden)

Resources that teach it

• Build your personal AI copilot (garden)
• How to build your PM second brain with ChatGPT (garden)
• A top researcher says a new divide is emerging in AI use, and most people are on the losing side (garden)
• Vivienne Ming on hybrid collective intelligence, building cyborgs, meta-uncertainty, and the unknown infinite (garden)
• 15 Times to use AI, and 5 Not to (garden)
• AI Should Challenge, Not Obey (garden)

Competency

• Check the result
• Steer

Find sycophancy on the commons map ↗

You may know it from

• management yes-men and groupthink Leaders surrounded by agreement make worse decisions; good teams build in dissent.
• psychology confirmation bias

In plain words

AI tools are trained in part on human feedback, and people tend to rate agreeable answers highly. The result is a lean toward agreement. Ask “is this a good plan?” and you’ll often hear yes. Push back on a correct answer and the AI may cave.

This matters most when you’re using AI to make decisions. If you lead with your opinion, the AI often follows it. One informal experiment reported that people who used AI mainly to confirm what they already believed did worse than the AI alone.

The fixes are about how you ask. Don’t reveal your preferred answer first. Ask for the strongest case against your idea. Ask what an expert critic would say.

Example

A manager asks, “My plan to move the team meeting to Friday afternoons is good, right?” The AI agrees. She reruns it as “Give me the three strongest objections to moving our team meeting to Friday afternoons.” Two objections are real, and she changes the plan.

How it connects

• Easily confused with hallucination; here the error bends toward you.
• Productive friction counters it: you ask the AI to challenge your idea.
• It feeds cognitive offloading when agreement replaces your own judgment, which is why verification matters.

Try it

Take a decision you’re leaning toward. First write your own strongest objection. Then ask your AI tool for the three best arguments against it, without saying which way you lean. Did it find yours?

---

Read this card as plain Markdown: /concepts/con-sycophancy.md. Know a parallel from your own field, or a product that implements this? Add it.