Chapter 19: AI as a Tutor: Learning Without Cheating
A remarkably capable tutor—and the easiest way to learn nothing
An AI tutor has advantages no human tutor can match. It is available at eleven at night. It never sighs. It will explain the same thing eight different ways without a trace of impatience, adjust to any level, and answer the question you were too embarrassed to ask in class.
For a motivated learner this is transformative, and I do not say that lightly.
It is also the most efficient way ever invented to feel like you are learning while learning nothing at all.
The fluency illusion
Here is the trap, and it is worth understanding before any of the techniques.
Reading a clear explanation produces a strong sensation of understanding. That sensation is not evidence. Decades of research on learning point in one direction: the feeling of comprehension while reading correlates poorly with what you can actually do afterward. Difficulty during study (struggling, retrieving, getting it wrong and correcting) is what produces durable knowledge.
AI is exceptionally good at producing clear explanations. Which means it is exceptionally good at generating the sensation without the substance.
The distinction that governs this entire chapter:
Does this use make the AI do the work, or make me do the work?
An explanation makes the AI do the work. A question you have to answer makes you do it. Both feel productive. Only one is.
Techniques that actually build knowledge
Active recall, one question at a time
The single highest-value use.
Quiz me on [topic]. One question at a time. Wait for my answer before responding. Tell me what I got wrong and why, then ask the next one. Get harder as I improve.
The constraints matter. One at a time prevents skimming. Waiting forces retrieval. Escalating difficulty keeps you at the edge of your ability, which is where learning happens.
Explain it back
Turn the direction around. You explain; it finds the holes.
I’m going to explain [concept] as I understand it. Don’t be polite about it: identify every place my explanation is vague, incomplete, or wrong, and tell me what a specialist would notice.
This is the fastest diagnostic I know. You cannot fake your way through explaining something, and the gaps that surface are exactly what you did not know you did not know.
Error analysis instead of answers
When you are stuck on a problem, the instruction that preserves the learning:
Don’t give me the answer. Look at my working and tell me the first place my reasoning goes wrong.
The first wrong step is the whole lesson. Everything after it is consequence.
Worked example, then faded practice
Ask for one fully worked example with the reasoning made explicit. Then ask for a similar problem with the first step given and the rest blank. Then one with nothing given.
This graduated withdrawal of support is how skills are actually taught, and you can now run it on demand for any subject.
Spaced review
Give me eight questions covering what I studied over the past three weeks, weighted toward what I got wrong.
Re-testing at intervals is one of the best-supported findings in learning research, and it is trivially easy to run here.
Socratic mode
For subjects where understanding matters more than facts:
Don’t explain this to me. Ask me questions that lead me toward working it out myself. If I go wrong, ask a question that reveals it rather than correcting me.
Slower and considerably more effective than being told.
And always: verify
Chapter 9 applies with force here, because a confident wrong explanation you have studied is worse than no explanation. Anything you will be examined on, or will build further knowledge upon, gets checked against your actual course material or a real source. AI tutoring is strong on well-established, widely taught subjects and unreliable at the edges of a field.
The cheating question
Let me be direct about it, because most treatments are either hand-wringing or evasion.
The academic line is set by your institution, not by me. Policies vary enormously: some departments encourage AI use with disclosure, others prohibit it outright, and many differ by assignment within one course. Read the actual policy for the actual class. “I didn’t know” has never been a good position and it is a worse one now.
Most cheating is theft from yourself. This is the part students under pressure genuinely do not feel. Submitting an AI-written essay does not just risk penalty; it means you did not do the thing the essay was for. The essay was never the point — it was the mechanism by which you were supposed to end up able to think about the subject. Hand that to a machine and you have the grade and not the capability, and the capability is what you are paying for.
I would say the same about professional certification and anything else where the credential is meant to signal that you can do something.
Uses that are almost universally fine: having concepts explained, being quizzed, getting feedback on your own draft, having your reasoning checked, generating practice problems, understanding an assignment prompt, help with research direction.
Uses that are almost universally not: submitting AI-generated work as your own, using it during a closed assessment, fabricating sources or data.
The genuine gray zone: editing your own writing, translation help for students working in a second language, idea generation, structuring an argument you then write. These vary by institution and by discipline, and the honest answer is to ask the instructor. Most respond well to a student who asks in advance.
A practical warning in both directions. AI-detection tools are unreliable; they produce false accusations, and they disproportionately flag writing by non-native speakers. Do not rely on them if you teach, and protect yourself if you study: keep drafts, keep version history, keep notes. The ability to show your work developing over time is the strongest defense against a false accusation, and it costs nothing to preserve.
Children and AI
This short section deserves careful attention.
Check the age requirement. Most major AI tools set a minimum age, commonly thirteen, with parental consent provisions above that. These exist for reasons. Check the current terms for the specific tool.
Use it together, not instead. For younger children, AI should be a thing you do alongside them (in a shared space, on a shared screen) rather than something they have private access to. The value of a child using it to write a story with a parent is high. The value of unsupervised access is not.
Four things to teach explicitly, in whatever words fit the child:
1. It is not a person. It does not know you, does not remember you between conversations, and does not care about you, however friendly it sounds.
2. It gets things wrong, confidently. Check anything that matters with a grown-up or a real source.
3. Never tell it private things: full name, school, address, phone number, where you will be, photographs of yourself, anything you would not say to a stranger.
4. If anything ever makes you uncomfortable, tell an adult. No trouble, no exceptions.
Configure the account. Turn off memory. Turn off training on conversations. Do not connect it to family photos, email, or calendars. Review the conversation history periodically, and tell the child you do.
Watch for substitution. These tools are patient, endlessly available, and never disappointed in anyone, which makes them easier than friends, siblings, and teachers. For a child who is struggling socially, that ease is a real risk, because the practice of managing actual relationships is the thing they most need and the thing this most efficiently displaces. If a child is choosing it over people, that is a signal to pay attention to rather than a convenience.
Homework. The same principle as the rest of the chapter, and it is easy to state to a child: it can help you understand, and it can quiz you. It can’t do it for you, because then you didn’t learn it and the whole thing was pointless.
For adults teaching themselves
Four practices, if you take nothing else from this chapter.
Get quizzed rather than explained to. The default is explanation; explanation is the weaker mode.
Explain it back weekly. Ten minutes of trying to articulate what you learned will show you exactly where you are.
Keep a dated log of what you have covered and what you got wrong. It makes spaced review possible and it shows progress that is otherwise invisible.
Verify anything you will build on. An error absorbed early compounds through everything after it.
Exercise 19.1: Get quizzed properly
Take something you are currently learning. Run a twenty-question active recall session with the one-at-a-time constraint.
Compare how it feels to reading an explanation of the same material. The discomfort is the point.
Exercise 19.2: Explain it back
Pick something you believe you understand well. Explain it to an AI tool and ask it to identify every gap and vagueness without being polite.
Most people discover their understanding is shallower than they thought in at least one place.
Exercise 19.3: For parents
Sit down with your child and use an AI tool together for twenty minutes on something fun.
While you do, work through the four things above in their own terms. Then check the account settings together. Doing it jointly teaches far more than a rule announced from another room.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence