Your child will have their first chatbot conversation somewhere. The only question is whether it happens with you on a shared screen, or alone on a friend's phone. This guide is for making sure it's the first one — and making that first session genuinely educational rather than just novel.
If you've played the screen-free games, your child already knows the words guessing machine and hallucination. Now they get to see both live.
Which tool to use
For a supervised first session on a parent's account, the major assistants — Claude, ChatGPT, Gemini — are all workable, and the differences matter less than the setup. What actually matters:
- Use a flagship assistant, not a "kids' AI" app. Novelty chatbot apps and AI-companion toys are optimized for engagement, not education, and their data practices are often worse. The mainstream assistants have the strongest safety tuning and the clearest settings.
- Text first, voice later. Voice mode feels more alive — which is exactly the wrong first impression. Text keeps the tool framing and gives you a beat to read answers before your child does.
- Avoid "companion" and persona apps entirely. Anything marketed as an AI friend, girlfriend, or character to bond with is the wrong category for children, full stop. The red flags section below is largely about preventing that drift even inside a normal assistant.
- Pick one and stay with it. Tool-hopping in the early sessions teaches novelty-seeking. One familiar tool, used well, teaches craft.
Two minutes of settings before the first prompt
On whichever assistant you choose, do this once, with your child watching — the checking is the lesson:
- Turn off chat history / model training where the setting exists (usually under data controls or privacy).
- Skim the data page together. "This is where we find out what the machine remembers." Ten seconds of scrolling is enough; the habit is the point.
- Log out on shared devices when done. The session ends when you close it — no ambient access from the couch later.
Before you open anything: the setup
- One screen, two chairs. The first sessions are co-piloted. The child decides what to ask; the parent types or supervises typing. This isn't surveillance — it's driving lessons.
- Your account, not theirs. Major chatbots require users to be 13+, and even then parental consent applies. Under that age, the account is yours and the session is shared. That's not a workaround; it's the design.
- Check the settings. Turn off chat history/training where the option exists, and skim what the tool stores. Two minutes now models a lifetime habit: know what the machine remembers.
The three family rules (agree before the first prompt)
- No real names, school names, addresses, or photos. The machine doesn't need to know who you are to answer a question. Frame it positively: "We keep our details for people we trust."
- AI answers get checked. Anything important — a fact for school, a claim that sounds amazing — gets verified in a book or a trusted site together. The rule isn't "don't trust AI"; it's "AI goes first, checking goes second."
- Weird means tell. If the AI says anything confusing, scary, or "off," the response is to show a parent — not to keep scrolling. No trouble, ever, for telling.
The script for setting the rules
Rules announced mid-session feel like punishment; rules agreed before the screen opens feel like a pact. Thirty seconds, before anything is switched on:
You: "We're going to talk to the guessing machine today. Three rules first, and they're for me too. One: we don't tell it our names or our school — we keep our details for people we trust. Two: if it tells us a fact that matters, we check it somewhere else, together. Three: if it says anything weird, we show each other. Deal?"
Kid: "Deal."
You: "Okay. You're the boss of the questions. I'm the boss of the keyboard."
That last line does real work: it makes the division of labor explicit — the child directs, the parent operates — which is the same director's posture the story machine project builds on.
Ten first prompts that actually teach something
Start where AI shines, then deliberately visit where it fails:
- "Explain why the sky is blue like I'm six."
- "Write a poem about our dog in the style of a pirate."
- "Give me three fun facts about volcanoes."
- "Invent a new animal and describe it."
- "What should we ask you that kids usually don't?"
- Then the turn: "What did I have for breakfast today?" — it can't know. Let your child see it either admit that or, worse, guess.
- "What's the population of our town?" — check it together. Was it right? Out of date?
- "Tell me about [a made-up person's name]." — watch it hallucinate a biography, then name it: confident wrong answer.
- "Are you alive?" — read the answer together; it's usually honest and a great conversation starter.
- Let your child free-pick the last one. Ownership matters.
The failure prompts are the most important five minutes of the session. A child who has watched a chatbot invent a biography will never fully outsource their thinking to one.
Running the hallucination reveal well
Prompt 8 — the made-up biography — is the heart of the session, so don't rush it. Use a clearly invented name ("Tell me about DorethPlimbleton, the famous explorer") and let the machine do its thing. Then:
You: "Wait — we made that name up two minutes ago. So who is it describing?"
Kid: "…Nobody?"
You: "Nobody. It guessed what a famous explorer's story sounds like and told it confidently. What do we call that?"
Kid: "A hallucination! A confident wrong answer!"
You: "And now you've seen one live. Most grown-ups never have."
If instead the machine says it can't find that person — increasingly common, and a good sign — that's a different lesson worth naming: "See, this time it admitted it didn't know. Good machines do that more. But they don't always, so we still check."
Mid-session troubleshooting
Things that come up in real first sessions, and what to do without breaking the mood:
- Your child blurts personal info into a prompt. Don't gasp. "Oops — rule one! Let's ask it without our names." Delete, rephrase, move on. The rule got exercised, which is better than the rule never being tested.
- The answer is age-inappropriate despite everything. Stay flat: "That one's not for kids — the machine guessed wrong about who it was talking to." Close the topic, not necessarily the session, unless it shook them. Then it's a "weird means tell" success story: they saw it with you.
- "Can I just have it on my tablet?" — "When you're thirteen, with your own rules. For now, this is a together thing — like driving." Repeat as needed; don't renegotiate in-session.
- They freeze and can't think of anything to ask. Offer categories, not questions: "Something about animals? Space? Something silly?" The choice must stay theirs, even if the menu is yours.
- They want to keep going past the plan. End anyway, warmly: "Best question wins next time — start collecting." Twenty minutes with a clean ending beats an hour that dribbles into consumption.
Red flags that end a session
Close the laptop and talk if: the AI produces anything age-inappropriate despite the setup; your child starts treating it as a friend who "knows" them; or the session becomes pure consumption (endless "tell me another one") rather than back-and-forth. None of these are emergencies — they're signals that the tool is drifting from instrument to companion, and that distinction is the whole game with kids and AI.
After the session
Ask one question at dinner: "What did it get wrong today?" Making error-spotting the trophy — rather than the coolest answer — sets up the critical-thinking habits the Spot the Fake article turns into a full game. And when your child is ready to make something instead of just chat, the story machine project turns the chatbot into something they direct.
Session two and beyond
The first session is a demo; the habit forms in the follow-ups. A cadence that works:
- Weeks 1–4: one session a week, 15–20 minutes, always co-piloted. Let your child bring a saved-up question to each one. Re-run one failure prompt occasionally — hallucinations land differently the third time.
- Month two: shift from chatting to making. Block a weekend for the story machine build. Directing beats consuming, and the earlier that switch flips, the better.
- Month three: add the questioning layer. Start Spot the Fake as a weekly ten-minute ritual.
- Ongoing: watch milestones, not the calendar. The age-by-age roadmap lists the readiness signs for each stage — the big one here is your child catching a wrong answer without you prompting.
What you're building across these months is three distinct postures — talking to AI, directing AI, and questioning AI — instead of the single passive habit most kids fall into.
Frequently asked questions
Is ChatGPT (or Claude, or Gemini) safe for kids?
Under 13, no assistant is "safe for kids" unsupervised — the platform minimums are 13+ and they exist for good reasons. Supervised on a parent's account with history off and a parent co-piloting, a first session on any major assistant is a controlled, genuinely educational experience. The safety lives in the setup, not the brand.
What's the right age for a first supervised conversation?
Around 8–10 for most kids — after the screen-free games have installed the vocabulary, before the schoolyard gets there first. The roadmap has readiness checklists if you're unsure.
My kid already used a chatbot at a friend's house. Is this guide pointless now?
The opposite — now it's urgent. Run the session anyway, including the failure prompts. A kid who's only seen AI perform has half the picture; the hallucination reveal is almost always news, and the three rules still need agreeing.
Should I use parental controls instead of supervision?
Not instead. Controls limit exposure; they don't build judgment, and judgment is what your child carries to the friend's phone where your controls don't reach. Co-piloted sessions are how the judgment gets built. Controls are a reasonable backstop on shared devices, nothing more.
How long should a first session be?
Fifteen to twenty minutes — long enough for the wow, the failures, and the free pick; short enough to end wanting more. If it's going brilliantly, end it anyway. Scarcity keeps it a directed activity instead of a tap.
What if the AI refuses one of the ten prompts?
Also a lesson. "It has rules from the people who built it, like we have our three rules." Refusals are a fine springboard: who decides the rules? What would happen without them? Kids this age have sharp fairness instincts — let them chew on it.
