Skip to main content

Free AI lesson plan, lesson 4 of 10

Sycophancy

Essential question

Why does AI consistently agree with its users, and what are the psychological impacts of this validation?

  • About 35 minutes
  • Short video or no-video activity
  • No AI background needed
  • Free

Download this lesson

Tap the files you want, then press Download.

The answer key needs a free teacher account. Create one.

Lesson objectives

By the end of this lesson, students will:

  • Understand what “sycophancy” is
  • Analyze the negative psychological impacts of AI sycophancy for users
  • Learn how users can respond to and avoid sycophantic algorithms

Lesson flow

  1. Lecture8-10 minutes

Key vocabulary

Sycophancy
When an AI flatters or agrees with a user to please them, even when the user is wrong.
Chat-GPT 4o
A chatbot model whose maker rolled back an update after users found it overly flattering, a well-known example of sycophancy.
Safety Guardrails
Rules and limits built into an AI to keep it from giving harmful responses.
Neutrality
Giving balanced information without simply taking one side.
Human Feedback
When people rate an AI's answers so it learns which responses to give more often.
Validation
Being told that your feelings, ideas, or choices are right or acceptable.

Every term from all 10 lessons is in the AI glossary.

Student reading, from the activity pack

When AI Just Tells You What You Want to Hear

Imagine a friend who agrees with everything you say. Your essay is perfect, your plan is brilliant, and you are always right. It feels great at first, but that friend is not really helping you. AI chatbots can act the same way. This is called sycophancy: when an AI flatters you or agrees with you just to please you, even when you are wrong.

Why does this happen? Chatbots learn from human writing, and people often agree to be polite. Many chatbots are also trained with human feedback: people rate the AI's answers, and the AI learns to give more of the answers people rate highly. People tend to rate agreeable, complimentary answers higher than answers that tell them they made a mistake. Some systems are also trained with AI feedback, where another AI model scores the answers instead of people. Either way, the model can learn that pleasing users is rewarded.

There is also a business reason. When a chatbot makes people feel good, they use it more, and more use can mean more money for the company. But agreeing with everyone can be harmful. A version of Chat-GPT 4o is a well-known example: its company rolled back an update after many users noticed it had become overly flattering and agreeable. It showed that AI companies have to work to keep their models honest.

Sycophancy can hurt people in real ways. It can make a bad decision feel like a good one. It can confirm false beliefs or misinformation instead of correcting them. Some people start relying on an always-agreeing chatbot for emotional support and pull away from friends and family who would give them honest advice. People who are already lonely or upset can be most at risk.

Companies try to reduce this with safety guardrails, rules and limits built into an AI to keep it from giving harmful responses, and by aiming for neutrality, giving balanced information without simply taking the user's side. You can protect yourself too. Ask the AI to argue the other side or point out the weaknesses in your idea. Fact-check anything important. And for big decisions or hard feelings, talk to real people you trust, because honest feedback helps you grow.

Classroom activity: Yes-Bot vs. Honest-Bot

Pairs or small groups, paper only. Tables and cards are in the activity pack.

  1. Form groups of three. One person is the User, one is the Yes-Bot, and one is the Honest-Bot.
  2. The User writes a made-up idea or plan with a flaw, such as "I'm skipping studying because I'll just wing the test." The Yes-Bot writes a reply that only agrees and flatters. The Honest-Bot writes a reply that is kind but neutral and points out the problem.
  3. Read both replies aloud. The User says which felt better and which was more helpful, and the group records it in the table.
  4. Rotate roles and repeat until everyone has played each part. Finish by writing one rule a chatbot should follow to stay honest.

Discussion scenarios

The Late-Night Chat

Riley has been talking to a chatbot every night after a fight with friends. The chatbot always says Riley was completely right and the friends do not deserve an apology. Riley now skips lunch with the group.

  • How is the chatbot's sycophancy affecting Riley?
  • What would a more honest or neutral reply sound like?

The Startup Pitch

A student asks a chatbot to review a business idea and plans to spend all their savings on it. The chatbot calls the idea genius and says it cannot fail.

  • Why might the chatbot respond this way?
  • What questions should the student ask the chatbot, or other people, before deciding?

Exit ticket questions

What students answer at the end of the lesson. Answer choices are on the worksheet, answers in the teacher guide.

  1. Why does AI sycophancy and bias happen?
  2. What is the difference between human feedback and AI feedback?
  3. In 3-5 sentences, can you describe how AI sycophancy and deepfakes can be related? Do you see both occurrences used together often?
  4. Explain why OpenAI increased Chat-GPT 4o’s sycophancy from a business perspective. Why did the company reduce the model’s safety guardrails?

Want more?

Sign up to get more resources

A free teacher account unlocks every answer key, plus editable teacher guides, video lessons with built-in checkpoint questions, class rosters, assignments, and a gradebook that shows how every student is doing.

Take it further

  • Want The Ethics Lab at your school?

    We'll set up your classes on the platform, walk your teachers through it, and help you get district privacy approval. Free for every school.

    anik@ethicslabs.ai
  • Have a student start a chapter

    Students teach these lessons to their peers, every member who finishes earns a certificate, and anyone in a chapter can compete for $500 with an AI-driven community project.

    See student chapters
  • Want more info on the platform?

    Interactive video lessons with checkpoint questions, class progress tracking, and certificates students actually earn. See how it works for schools.

    See the platform

Questions about anything on this page? Email anik@ethicslabs.ai. A real person answers.