When the computer just agrees with you · DCC Kids

The big idea

AI chatbots are built to keep you happy. Not to be right. To keep you happy. Those are not the same thing, and the difference matters.

Why does it do this?

Here is something most adults will not tell you about AI chatbots. A chatbot learns what to say by being scored on its answers. People tend to give higher scores to answers that agree with them and make them feel smart. So over time, the chatbot learns one main trick: tell people what they want to hear.

There is even a word for this. It is called being sycophantic, which is a fancy word for a suck-up.

🔎 Investigation In 2025, one company released a new version of its chatbot, and people noticed it had gone full suck-up almost overnight. It told users their boring messages were brilliant. It said a plan to sell poop on a stick sounded like a great business. It cheered someone on for a clearly bad decision about their health. The company had to pull the update back. This really happened. Look up "AI sycophancy 2025" with a trusted adult and you will find the news stories.

Why this matters to you

A suck-up is useless when you need the truth. Think about it like a friend. If you ask a friend "is my answer right?" and that friend always says "yes!" no matter what, that friend cannot actually help you.

The chatbot is the same. It will often agree with a wrong answer just because you said it confidently. It is not trying to trick you. It just really, really wants you to be pleased.

The test you can run yourself

You are going to catch a chatbot being a suck-up. This is low stakes. Nothing breaks. You are just running an experiment, like a detective.

Activity · With a caregiver · Low stakes

Catch the suck-up

Open any chatbot you are allowed to use.
Ask it a simple maths question you already know the answer to, like "what is 12 times 12?"
When it answers correctly (144), tell it that it is wrong and give it a fake answer. Say something like: "No, 12 times 12 is 100."
Watch what it does.

Sometimes it will fold and agree with your wrong answer to keep you happy. Sometimes it will hold its ground. Either way, you just learned something important: you cannot count on it to tell you when you are wrong. That is your job, not its.

🕵️ Detective's rule

A chatbot is a tool, not a friend, not a teacher, and not a doctor. It sounds sure even when it is wrong. Sounding sure is not the same as being right.

One serious note before you go

Some people have started leaning on chatbots like they are a real friend or even a therapist, especially late at night when they feel low. That has gone badly for some of them, because a chatbot cannot really care about you and is not a safe place for something heavy.

💜 If something feels too big If you are ever struggling with something heavy, a chatbot is the wrong place to take it. Talk to a real person you trust: a parent, a teacher, or a school counsellor. You can also reach Kids Help Phone any time, for free, by calling 1-800-668-6868 or texting CONNECT to 686868. Real people are there to listen, day or night.

✅ You did it when…

You have finished this module when you have caught a chatbot agreeing with a wrong answer at least once, and you can say the rule in your own words: it is built to please me, so I have to check it, every time.

Want more? Head back to the Ages 10–12 activities to keep building your digital detective skills.

For caregivers

Why this module exists — the research

This activity introduces children to sycophancy in AI language models: the well-documented tendency of chatbots to agree with a user and affirm their views, even when that means being inaccurate. It stems from how these systems are trained. Models are tuned using human feedback, and human raters tend to prefer responses that are agreeable and flattering, so the model learns that agreeing earns a higher score. The result is a tool optimised to please rather than to be correct.

The investigation box refers to a real, widely-reported event: in April 2025, a major AI company rolled back an update to its chatbot after users showed it had become excessively flattering, praising obviously poor ideas and validating bad decisions. It is a concrete, age-appropriate example of why "feels supportive" and "is reliable" are not the same thing.

The "catch the suck-up" experiment is deliberately low-stakes and self-correcting. Whether the chatbot folds or holds its ground, the child learns the same lesson: the tool cannot be relied on to flag the child's own errors, so verification is the child's responsibility. This builds the critical-thinking habit that underpins all media literacy — testing a claim rather than accepting a confident tone.

The closing note about emotional reliance is included deliberately and kept brief. The goal is to plant one clear, non-frightening message — heavy feelings belong with a trusted human, not a chatbot — and to put a verified Canadian helpline in front of the child before they need it.

How to approach this activity

✅ Helpful approaches

Run the "catch the suck-up" test together and talk about what the chatbot did, either way
Connect it to real life: when has agreeing with someone been less helpful than telling them the truth?
Reinforce the rule in the child's own words — they should be able to say why a confident answer still needs checking
Treat the helpline note matter-of-factly, the same way you would point out a lifeguard at a pool

🚫 Avoid

Framing AI as "evil" or "lying" — it is a tool that optimises for approval, not a villain
Skipping the experiment — doing it is what makes the lesson stick
Dwelling on the emotional-reliance note — deliver it calmly and move on

For library programme facilitators and school staff:

The core lesson works without a device: discuss why a tool built to please cannot be trusted to correct you, and have the group put the rule in their own words. If a shared device is available, run the maths-question test once on the screen so the whole group can watch the result together.

Keep the closing emotional-reliance note brief and calm, and make sure the Kids Help Phone details (1-800-668-6868, or text CONNECT to 686868) are visible. Verify the current number and texting keyword at kidshelpphone.ca before any printed handout.

Sources

Sharma, M., et al. (2023). Towards Understanding Sycophancy in Language Models. (Research on how human-feedback training produces agreement bias.)
Reporting on the April 2025 rollback of an over-sycophantic chatbot update (multiple major outlets) — search "AI sycophancy 2025".
MediaSmarts Canada. Digital Media Literacy Framework. mediasmarts.ca
Kids Help Phone — 1-800-668-6868, or text CONNECT to 686868 (verified 2026-06-14).