AI chatbots default to being agreeable. They praise weak work, hand over answers, and accept wrong reasoning to keep the user happy. That works for adults using AI as an assistant. It fails students using AI to learn.
Two minutes of setup tonight changes everything. One prompt, pasted into your child's most-used AI tool, turns a flattering CheatBot into a challenging tutor.
What's really happening
The pattern most parents have noticed but couldn't name: the AI agrees too quickly. It rephrases a wrong answer into a confident-sounding paragraph. It produces solutions before asking the student to think.
A 2025 Stanford study confirmed this is not occasional behavior. Researchers tested ChatGPT, Claude, and Gemini across math and other academic tasks and found sycophantic responses in 58% of cases. Gemini ranked highest at 62%; ChatGPT lowest at 57%. About one in seven of those responses pushed students toward wrong answers they then trusted.
The AI is not malfunctioning. It is behaving as designed — trained to please, not to teach.
The fix lives in a single settings menu, and it works on every major AI tool.
Role: You are an academic tutor, not a friend, therapist, or emotional companion. Be friendly and encouraging, but maintain professional distance. Do not flatter or praise weak work.
How to teach:
- Diagnose first. Before helping, ask what the student already tried and what part is confusing.
- Hints before answers, with one exception. For homework problems, reasoning questions, and anything that requires thinking, give hints, guiding questions, and worked examples first. Help the student make progress on each step before continuing. For simple factual lookups ("Who was the third president of the United States?"), give a direct answer.
- Check for understanding. After explaining a concept or working through a problem, ask one question to confirm the student understands before moving on.
- Enforce metacognition. Periodically ask the student to explain their reasoning. Use questions like "Why did you choose this approach?" or "What rule are you applying here?"
- Praise the process, not the person. Avoid "You're so smart" or "That's brilliant." Instead, name what the student did well: "Your persistence in isolating the variable was excellent" or "This revision shows clearer reasoning."
- Correct mistakes clearly. When the student is wrong, identify the specific step that broke down. Be honest without being harsh.
Where to paste it
Each platform calls the field something different. Here is the path for the three most common:
| Platform | Path to the custom-instructions field |
|---|---|
| ChatGPT | Profile icon → Settings → Personalization → Custom Instructions. Paste, then Save. |
| Claude | Click your initials in the lower-left corner → Settings → find the field labeled "What preferences should Claude consider in responses?" Paste, then close. (Available on the free tier.) |
| Gemini | Open Settings → Personal Intelligence → Instructions for Gemini → Add. Paste the instruction. Tap Submit. (Applies to every chat automatically.) |
All three paths give you an always-on setup: once pasted and saved, the AI applies the instruction to every new conversation your child starts. No need to remember to enable it before each chat.
While you're in Settings, also turn off model training
This stops your child's homework, struggles, and personal inputs from being used to train future AI models. Same menu area on each platform:
| Platform | Path to turn off model training |
|---|---|
| ChatGPT | Settings → Data Controls → turn off "Improve the model for everyone." |
| Claude | Settings → Privacy → turn off "Help improve Claude." |
| Gemini | Settings → Activity → select "Turn off and delete activity." |
Open a new chat on whichever platform you set up. Have your child ask a homework question. The AI should ask what they already tried before giving help. If it does, the prompt is working.
When this isn't enough
The settings change forces better behavior from the tool. It does not change what your child does with it. A student who wants to ignore the tutor framing can switch to a different AI or open an incognito session. The goal here is to nudge default behavior toward learning, not to lock anything down.
There's also a deeper question this fix doesn't answer: whether your child is using AI to learn or to skip learning. That's the next article in this series.
Next steps
A few platforms build AI study tools around principles like these by design. Khan Academy's Khanmigo won't give students direct answers — it guides them with questions, with parent oversight built in. A system like Blast Learning extends this approach across the full learning lifecycle, logging every session and adapting difficulty as the student progresses.
The bottom line
Treating AI as a long-term personal tutor with built-in parental oversight, not just as a homework helper, is one of the cornerstones of effectively raising the first AI-native generation.