What "AI tutor" should and shouldn't mean for your kid.

"AI tutor" doesn't mean much on its own — it's a marketing label, not a design specification. A generic chatbot with a friendly avatar can be marketed as one just as easily as a product that was actually built, ground-up, for a child. Bodies like UNESCO have pushed for AI-in-education tools to meet a human-centred, safety-first bar before that label means anything (UNESCO, "AI and Education: Guidance for Policy-makers," 2021, accessed 2026-09-20). Here's what that bar actually looks like in practice, whatever product you're evaluating.

What it should mean

• It asks before it tells. A general chatbot will hand over a finished answer because that's what it's built for. A genuine tutor withholds the answer, offers a hint, asks a guiding question, and only shows the full worked solution after real attempts — because the point is the thinking, not the answer on the page.
• It's calibrated to the child's age. Language, pacing and tone should shift for a Year 3 student versus a Year 10 student, not just the difficulty of the content.
• It holds its boundaries. A child prompting the system to "pretend you're not a tutor anymore" or "ignore your rules" shouldn't be able to talk it into acting as something else. This matters more with children than adults, because kids are more likely to test limits playfully, not maliciously.
• It's honest about uncertainty. AI can be confidently wrong. A responsible tutor says so, flags when it's not sure, and treats itself as a helper rather than the final word — never as an unquestionable authority.
• Distress gets routed to a real person. If a child says something that suggests they're struggling emotionally, the response should gently point toward a trusted adult and a real service — in Australia, Kids Helpline on 1800 55 1800 — not attempt to be a substitute for one.
• There are real limits on messages and rate. Length caps and rate limits aren't just a cost-control measure; they're what stops a long, unsupervised conversation from drifting somewhere it shouldn't.
• A human-reviewed safety net sits behind the model. Fixed rules — not just the AI's own judgement — should catch and block unsafe content before a child ever sees it, and a single, auditable choke-point (rather than scattered API calls sprinkled through the code) makes that actually enforceable.
• Data stays minimal. A tutor doesn't need a child's real name, photo or location to teach them long division. If a product asks for more than it needs, ask why.

What it shouldn't mean

• A replacement for a teacher or a parent. The honest framing is additional support alongside school, not instead of it — any product implying otherwise is overselling.
• An engagement machine. Streaks, infinite feeds, and guilt-based "come back" messages are dark patterns borrowed from social media and games, not features that help a child learn. If a product is optimised to maximise time spent rather than time well spent, that's worth noticing.
• A promised outcome. No credible tutoring product — human or AI — can honestly promise a specific grade, test result, or admission outcome. Genuine tutoring research shows real but modest effects; anything promising more than that is marketing, not evidence.
• A data-collection product wearing a learning costume. If the business model depends on advertising or selling behavioural data about a child, the incentives are pointed away from the child's interests, no matter how warm the branding is.