How to Tell if an AI Tutor Is Safe and Accurate

Before you let a child near an AI tutor, run it through five checks: can it tell you where its answers come from, does it admit when it doesn't know, is it clear about what it does with your child's data, will it step back and hand over to a human when a question needs one, and is there a real, verified human standing behind it at all. A tool that clears all five is worth using. One that fails any of them — the last two most of all — is a convenience with a hidden cost. Everything below is how to run each check yourself, in a few minutes, without taking anyone's marketing on trust.

The reason this checklist has become necessary is that "AI tutor" no longer describes one kind of thing. It covers a bare chatbot dressed up inside a revision app, a subject assistant grounded in a real curriculum, and a full platform where an AI sits alongside human tutors you can vet. On a phone screen these look almost identical. What separates them only shows up when a child asks something the tool gets wrong, types something worrying into it, or needs more help than a generated paragraph can give.

The landscape changed, so the old signals stopped working

Two things are true about AI in education in 2026, and they pull in opposite directions. The technology genuinely improved — modern language models explain a tricky topic more clearly, more patiently, and more in line with the curriculum than anything available a couple of years ago. That is a real gain for a child stuck on quadratic equations at nine in the evening. At the same time, confident, fluent, plausible-sounding output is now the default of almost every AI product, whether or not it was built with a child's syllabus, data, or safety in mind.

That combination breaks the instinct most parents rely on. A well-written, self-assured answer used to be a reasonable proxy for a competent tool. It no longer is, because fluency is free and universal. So the question shifts from "does this sound right" to "how was this built, and what happens when it is wrong." The five checks below are simply that question, broken into parts you can actually test.

"Safe" and "accurate" are not the same question

Parents tend to fold safety and accuracy into a single worry, but they fail independently. An AI tutor can be entirely safe — no unsuitable content, no careless handling of personal data — and still be confidently wrong about the specification, the mark scheme, or the method a particular exam board rewards. The reverse also happens: a tool can be broadly right on the facts and still be unsafe in how it treats a child's data, a distressed message, or a moment that plainly needs an adult rather than a chatbot.

Because they break in different ways, one "is it safe?" question can't cover both. You have to ask two things: is the information correct, and is the way it reaches a child appropriate. The checks that follow deliberately test each side in turn.

Check 1 — Where do its answers actually come from?

A general-purpose chatbot draws on whatever its underlying model absorbed during training: a huge, unlabelled sweep of the internet, with no promise that any of it matches the exam board your child is sitting. It will answer with the same steady confidence whether the source was a peer-reviewed textbook or a forgotten forum post. Confidence and correctness are not the same property.

A tool built for learning should be able to show its footing. Does it draw from a curated, curriculum-aligned knowledge base, or is it improvising from general training data? You can test this directly: ask it a curriculum-specific question — a set text, a required practical, a formula on a particular syllabus — and then ask where that came from. A tool made for education can usually point to its grounding. A bare chatbot usually can't, because it doesn't actually know where its own answer originated.

Check 2 — What does it do when it doesn't know?

This is the single most revealing test, and it costs you nothing to run. A weak AI tutor guesses and dresses the guess as fact — the failure researchers call hallucination. A carefully built one is designed to hold back: to say "I'm not certain" rather than invent, and to track what a learner has genuinely grasped instead of assuming.

To check it, ask something obscure, or feed it a subtle error and see whether it corrects you or nods along. Ask about a topic slightly outside the syllabus and watch whether it flags the boundary. A tool that never once says "I don't know" is telling you something structural about how it was made: it was optimised to always produce an answer, not to be honest about the limits of one.

Check 3 — What happens to your child's data?

The UK has a clear floor here. The Information Commissioner's Office's Age Appropriate Design Code — the Children's Code — requires any service likely to be used by under-18s to act in the child's best interests, minimise the data it collects, limit profiling, and be plain about what it gathers and why. That is the legal baseline, not a nice-to-have. A safe AI tutor should be able to tell you, in ordinary language, what it stores, why it stores it, and who can see it — without you having to excavate a privacy policy written for adults.

The Department for Education's guidance on generative AI in education points the same way for schools and edtech: protect pupil data, keep an accountable human responsible for decisions that affect a child's learning, and treat AI as a support tool rather than an unsupervised decision-maker. Read together, the regulator and the department set a simple test. If a tool can't give you a straight answer to "who sees what my child types into this," it hasn't cleared the bar, whatever its marketing says.

Check 4 — Does it know when to hand over to a human?

An AI tutor that never escalates is a tool with no ceiling. Real learning eventually reaches something a chatbot shouldn't handle alone: a child in genuine distress, a safeguarding concern, a question that calls for a qualified adult's judgement rather than a generated paragraph. The design question is whether the tool is built to recognise that edge and stop, or whether it will keep producing text no matter what it has just been asked.

This is exactly where "AI tutor" and "AI tutoring platform" separate. A standalone chatbot has nowhere to pass a child on to — the handover simply doesn't exist as an option. A platform that pairs AI help with a set of real, checkable human tutors does have somewhere to go, which is the whole reason for building the two together rather than shipping one and pretending it replaces the other.

Check 5 — Is there a real, verified human standing behind it?

Most AI-safety checklists stop at check four, because they examine the AI on its own. But the honest answer to "is this AI tutor safe" is rarely just about the software — it's about what backs it up. A chatbot with no accountable human behind it asks you to trust a program that faces no consequences for being wrong. A platform where the AI works alongside verified human tutors asks you to trust a person whose credibility you can actually inspect.

On Tutorwise, that inspection has a name: Credibility as a Service, or CaaS. A human tutor's standing here isn't a self-written bio or a star rating that enough reviews can inflate — it's a computed score built from real, checkable signals across six categories: delivery (how sessions actually go), credentials (qualifications that have been checked, not claimed), network (professional standing), trust (identity verification and an enhanced DBS check), digital presence, and demonstrated impact. A verified DBS check and a verified identity sit inside that trust category and lift the number; their absence lowers it. No tutor is given any score at all until they have verified their identity or completed onboarding — a hard floor before a single figure appears. The mechanism is set out in full in how CaaS works and what a credibility score actually measures.

The practical upshot for a parent is the opposite of "trust us." You don't get a blanket promise that everyone has been vetted. You get to see, tutor by tutor, who has verified their identity and who holds a verified DBS check, and to watch that evidence move a score you can compare against other tutors. That is a far stronger position than a claim, and it's why reviews alone aren't enough to judge a tutor: a five-star average tells you sessions felt fine, not whether the person is who they say they are or has been through the specific, legally defined process a DBS check involves.

What this looks like with Sage, in practice

Tutorwise's own AI tutor, Sage, is built on this principle rather than having it bolted on afterwards. Sage answers from a curriculum-focused knowledge base, is designed to signal uncertainty instead of bluffing, and keeps track of what a learner has actually mastered over time. When a question needs more than it should safely give — real accountability, a qualified judgement call, or simply a level of teaching a chatbot shouldn't attempt — it points the learner towards a verified tutor on the platform whose CaaS score is visible before any session is booked.

Here is how a parent can run the whole checklist in one sitting. Say your Year 10 child wants to use Sage for GCSE maths revision. Ask it a syllabus-specific question — something tier-dependent, or a non-calculator method — and check the explanation against the actual specification (Check 1 and 2). Feed it a deliberately wrong step and see whether it corrects you (Check 2). Look at what the account holds about your child and who can see it (Check 3). Then, when the child needs more than a chatbot should offer, follow where Sage hands them: to a human tutor, and look at whether that person's credibility is shown as an earned score rather than asserted in a bio (Check 4 and 5). That last step is the one a standalone AI app structurally cannot offer, because there is no verified human on the other end of it.

This is also the line between AI genuinely helping a child and a product overselling what its AI can do. Sage is useful precisely because it is candid about its limits and backed by people whose standing you can verify — not because it claims to replace a qualified human. It doesn't, and it isn't designed to. Tutors on the platform also carry safeguarding duties that are set out in law, not merely in a platform policy.

Red flags — the checklist failing in real time

A few signs an AI tutor wasn't built with any of this in mind: it answers everything with total confidence regardless of subject or difficulty; it can't say where its information comes from; its privacy policy is generic and never mentions children specifically; there is no route from the AI to a real, checkable human when a question needs one; and any "verification" or "trust" language turns out to be marketing copy rather than something you can independently confirm, like an identity check or a DBS check. Treat a tool that trips two or more of these as a study aid to supervise closely, not something to leave a child alone with.

Overclaiming deserves a warning of its own. Any tool that markets itself as a complete replacement for a qualified human — "never need a tutor again," implying it can shoulder safeguarding, or promising a level of accuracy no language model can honestly guarantee — is telling you it is built for the sale, not for your child. The safest tools tend to be the ones that are upfront about what they can't do, because that honesty is itself evidence the rest of the design was handled with care.

FAQ

Is Sage free to use? Yes. Sage is Tutorwise's AI tutor, and you can start using it without paying, bringing in a verified human tutor whenever you want or need one.

How is Sage different from ChatGPT or a general AI chatbot? A general chatbot answers from broad training data with no accountable human behind it. Sage answers from a curriculum-focused knowledge base, is designed to flag uncertainty rather than guess, and sits inside a marketplace of tutors whose credibility is a computed, checkable score rather than a claim. For revision with general-purpose tools, see how to use ChatGPT for GCSE revision safely.

Can an AI tutor replace a human tutor? No — and a well-designed one won't pretend to. AI is strong at instant, patient, curriculum-aligned explanation; it can't replace a qualified human's judgement, accountability, or the safeguarding a real relationship needs. The full comparison is here.

What should I actually check before trusting an AI tutor with my child? Where its answers come from, whether it admits uncertainty, what it does with your child's data, whether it can hand off to a real human when needed, and whether that human's credibility is independently verifiable rather than self-declared.

Is AI tutoring safe for a child with additional needs? It depends heavily on the platform and the specific need — a curriculum-grounded AI tutor backed by verified human specialists is a very different proposition to a generic chatbot. This covers AI tutoring for SEN students specifically.

---

Traderwise, Trainerwise and Adspots are separate Tutorwise verticals; this article covers tutoring only.

More in this series:

• Can an AI Tutor Replace a Human Tutor? An Honest Answer
• How to Use ChatGPT for GCSE Revision Safely and Well
• AI Tutoring for SEN Students: Useful Support or Risk?