What Every Parent Needs to Know About AI Tutoring (Before Their Child Uses One)

There is a particular kind of helplessness that comes from watching your child use technology you do not understand. AI tutoring tools are arriving fast in UK secondary schools and homes, and the marketing around them tends to promise everything while explaining nothing. This article draws on Part II of AI Tutors for Key Stage 3 — the book's plain-English AI-literacy primer — and its aim is simple: give you enough understanding to ask the right questions, not to turn you into a software engineer.

You need to understand five ideas. Nothing more.

1. The Language Model: A Very Good Guesser

The engine inside almost every AI tutoring product right now is called a large language model, or LLM. The name sounds imposing, so here is the plain version: an LLM reads an enormous quantity of text and learns, in fine statistical detail, which words tend to follow which other words. When you ask it a question, it generates a response by predicting, one word at a time, what a plausible and coherent reply would look like.

That one sentence contains both the power and the limit of the technology.

The power is that "a plausible and coherent reply" can be extraordinarily good. Given enough training text, the model has absorbed the patterns of clear explanation across millions of documents — textbooks, worked examples, teacher notes, academic papers.

The limit is that the model is not looking anything up. It is not checking its answer against a reliable source in the way a teacher might pause and verify a fact before speaking. It is predicting. Most of the time this produces something accurate. Occasionally it produces something fluent and wrong. Parents who understand this know to watch for it — and know to ask providers how they handle it.

2. The Harness: Where the Rules Live

A raw language model has no particular values, no patience, and no awareness that it is talking to a fourteen-year-old who should be learning, not just receiving answers. On its own, it is a powerful but undisciplined know-it-all.

The harness is the program that wraps around the model and changes everything. Think of it as a strict job description handed to the model before every conversation. The harness sets the rules: answer questions with questions before giving solutions, never reveal the answer to a homework problem directly, always speak at an appropriate level for this student's age and subject, and — critically — respond in a specific way if the student appears to be in distress.

When you evaluate an AI tutoring product, the harness is what you are really evaluating. Ask: what are the rules? Who wrote them and how are they tested? What happens if the model tries to ignore them? A provider who cannot answer those questions clearly has either not thought carefully about the harness, or is not willing to tell you.

This is also why the approach we took at aitutors.me matters — the choice of underlying model affects what the harness can reliably enforce.

3. Agents: Thinking in Loops

Early AI chatbots answered in a single shot. You asked, they responded, the conversation moved on. Modern AI tutoring tools increasingly use a pattern called agents, which works differently.

An agent does not answer immediately. Instead it moves through a loop: think about what is needed, take an action (look something up, retrieve a past exercise, consult a teaching plan), observe what came back, then think again before responding. This loop can repeat several times before the student sees any output.

The practical effect is that the tutor feels more deliberate. It can check what the student has already covered, consult a skill set for the relevant topic, and decide on a Socratic prompt rather than a direct answer — all before typing a single word. If you have noticed that a good AI tutor feels less like autocomplete and more like a thinking conversation, this is why.

The Socratic method in particular requires this kind of looping: knowing when to ask rather than tell is not a single-step decision.

4. MCP: The Socket That Connects Tools

You may see the term MCP in technical descriptions of AI tutoring platforms. It stands for Model Context Protocol, and the analogy is a standard electrical socket.

Different devices have different plugs, but a standardised socket means any compliant device can connect. MCP does the same thing for AI applications: it gives a general-purpose AI assistant like Claude a standardised way to plug into external tools — a tutoring curriculum, a student progress tracker, a set of teaching exercises.

For a parent, the important implication is portability. MCP means the tutoring tools are not permanently baked into one application. They can be updated, replaced, or extended without rebuilding everything from scratch. It also means the AI you already use can, in principle, become your child's tutor — rather than requiring a separate, closed system.

5. Skills and Memories: Becoming a Specialist

A general-purpose language model knows something about everything, which makes it useful and unreliable in equal measure. Skills are a way of loading specific, structured know-how on top of the general model for a particular purpose.

Think of it as the difference between a generalist who has read every textbook and a specialist who has spent years teaching Year 9 maths in UK secondary schools. The specialist has internalised not just the content but the common misconceptions, the typical stumbling points, the language that works for that age group, and the curriculum structure they need to follow. Skills package that kind of expertise into something the model can apply consistently.

Memory is related: the ability for a tutor to carry knowledge of this particular student from one session to the next. What topics have been covered? Where did the student struggle last week? A tutor with no memory starts from zero every time; a tutor with memory can build on what came before.

Both skills and memory are design choices. They do not come for free with the underlying model. Ask any provider how their product handles both.

6. Safeguarding: The One Non-Negotiable

Safeguarding deserves its own mention, not because it is technically complex — it is not — but because it is the one area where a harness failure has real consequences rather than merely an inconvenient wrong answer.

A safeguarding hook is a rule wired into the harness that detects language associated with distress, self-harm, or crisis, and redirects the conversation to a trusted adult or a helpline such as Childline (0800 1111). This rule must be explicit, tested, and impossible for the model to override.

No language model comes with this built in. It is built by the people who write the harness, and it needs to be verified by them regularly. If a provider cannot tell you specifically how their safeguarding hooks work and how they test them, that is the most important gap in their answer.

The Questions That Actually Matter

Once you understand these five ideas — the language model, the harness, agents, skills, and memory — you have exactly the vocabulary you need. The questions reduce to a short list:

• What rules does the harness enforce, and who verifies them?
• How does the product handle the model being wrong?
• What does the tutor do if my child appears to be in distress?
• What does the product remember, and who has access to that data?

Any provider who takes these questions seriously will have clear answers ready. Any provider who cannot answer them clearly has not done the work.

AI Tutors for Key Stage 3 by Duke Harewood (Innovatorly Ltd, 2026) is available in English and Simplified Chinese. It is written for parents, not technologists, and Part II — from which this article draws — covers these foundations in full without requiring any prior technical knowledge. Find out more at aitutors.me/books, or explore the companion tutoring service at aitutors.me.