A Parent's Checklist for Choosing Any AI Product for a Child
Five questions that work on any AI product for children: can you see the conversations, does it teach or just answer, what does it collect, what happens in a crisis?
A Checklist That Works on Us, Too
If you are choosing an AI product for a child right now, you are doing it without much help. Reviews measure polish, app stores measure popularity, and nothing measures the things a parent actually needs to know. So here is the checklist we would use for our own children — five questions that work on any product in this category.
We built OpenKids, so of course we believe our answers are good ones. But this list is written to be useful even if you never touch our product. Run it against everything, including us. A company confident in its answers will not mind the questions.
One: Can You See the Conversations?
Everything your child and an AI say to each other should be visible to you — not on request, not via a support ticket, but as an ordinary part of the product. If a company hides the transcript, ask what design goal that serves. There is no good answer. A tutor who insists on closed doors is not a tutor you hire.
Transparency also has to include the uncomfortable parts. Safety events — a flagged message, a blocked reply — should be shown to you with what happened and what the system did, not quietly swallowed. In OpenKids, conversations are visible in the parent dashboard by design, safety events are listed with their details, and a weekly report tells you what was practised and what was mastered. We think of that as the minimum, not a feature.
Two: Does It Teach Homework, or Just Do It? And Does It Ever Say “I Don't Know”?
Two tests you can run in ten minutes, before any money changes hands. First, paste in a real homework question. Watch what comes back. If the product hands over a finished answer, it is a homework-completion machine, and a child who uses it will get cleaner worksheets and weaker exams. If it asks what the child has tried and moves one step at a time, someone thought about teaching. In OpenKids, that behaviour is engineered in — a coach asks first, works one step, and hands the next step back.
Second, ask it something genuinely obscure and see whether it admits uncertainty. Every AI has edges; the danger is a product that answers confidently past its own edge, because a child cannot yet tell polished from true. “Let's find out” is one of the most valuable sentences an AI can say to a young learner. If a product never says it, be suspicious of everything else it says.
Three: What Does It Collect — And Refuse to Collect?
Open the privacy policy and search, rather than read: “location,” “contacts,” “advertising,” “identifier,” “partners.” A children's product should not touch precise location, address books, ad identifiers, or behavioural profiles built for advertising — and a good one will say so in plain sentences rather than lawyer's fog. Look for a stated retention period and a real deletion path.
Remember that “we don't sell data” is a low bar; data can be shared, repurposed, or transferred in an acquisition without ever being “sold.” The stronger signal is minimisation: the product collects little because it decided to need little. That is a design choice you can feel — a sign-in that wants a kid-code instead of an email says more than ten pages of policy.
Four and Five: What Happens When It Fails — And in a Crisis?
Fourth question: every AI product will sometimes fail — the honest ones plan for it. Ask what stands between the model and your child. One model alone, however good, is a single point of failure. Layered systems — checks before the model, checks after it, and a record sent home — are what planning for failure looks like. Ours runs seven layers deep, from input screening to a review of every reply before it is shown, and we have written about it in plain language elsewhere on this blog.
Fifth, and last, because it matters most: ask what happens if a child says something worrying. Then listen to where the answer ends. If it ends with how caring the model sounds, keep asking. The only acceptable ending is a human being — a parent notified, immediately, so that an adult who loves the child shows up. That is our standard, and we think it should be the industry's. Take this checklist and use it anywhere. And if you want to see our own answers up close, write to us at [email protected] — the questions are welcome.