AI tutor or human tutor: an honest comparison

A good human tutor is still better than any AI tutor at motivation, judgement and relationship, and human tutoring has a randomised evidence base that AI tutoring does not. AI wins on availability, frequency and cost, and frequency happens to be one of the strongest predictors of tutoring working at all. For most families, though, the honest comparison is not AI against a human tutor. It is AI against nothing.

Start with what is actually established

Human tutoring is one of the best-evidenced interventions in education. A 2024 meta-analysis of 89 randomised experiments found a pooled effect of 0.288 standard deviations, with the largest effects for trained tutors, three or more sessions a week, and delivery during the school day.

AI tutoring has no equivalent body of evidence. There are promising individual studies and a great deal of marketing. Any company, including ours, that cites the human tutoring literature as though it transfers is arguing by analogy. The honest position is that the design can follow the evidence while the outcome remains unproven.

Where each one is genuinely better

A human tutor is better at:

• Reading the person. Distinguishing "does not understand" from "has had a terrible week" is the single hardest thing to automate, and it changes what the right next move is.
• Knowing when to push and when to stop. Judgement built from having taught many students.
• Relationship as motivation. A teenager turns up for a person they like. That is doing real work, and it does not transfer to software.
• Local knowledge. This school, this teacher, this exam board, this professor's habits.
• Noticing the non-academic. The maths problem that is actually a sleep problem, an eyesight problem or a home problem.

An AI tutor is better at:

• Being there. Immediately, at any hour, without arranging anything.
• Frequency at a price. The thing the evidence most favours is the thing an hourly rate makes unaffordable.
• Perfect memory. Every previous session, every specific error, with no notes to keep.
• Inexhaustible patience. The fourth explanation of the same idea is delivered exactly as well as the first.
• Consistency. No bad days, no rushing because the next student is waiting.

The comparison most families are actually making

In most households the choice is not between an AI tutor and a human tutor. At roughly $60 an hour, three sessions a week is around $7,000 a year, which is simply not available to most families. So the real options are: an AI tutor, occasional human tutoring at a frequency the evidence does not support well, or nothing.

Framed that way the question changes. It is no longer "is this as good as the best option" but "is this better than the option I actually have".

Where the gap is closing, and where it is not

Closing fast: explaining a concept clearly, generating practice, adapting wording to a student's level, answering follow-up questions, being available in any subject.

Closing slowly: judging when a student needs encouragement rather than instruction, deciding to abandon the plan because today is not the day, building the kind of relationship that gets a reluctant student to the desk.

Possibly not closing: being a person a young person trusts. It is worth being sceptical of anyone confident about this in either direction.

How to decide

The row worth dwelling on is the second. A student who has decided they are bad at maths does not have a content problem, and more explanation will not fix it. That is a job for a person.

What a session actually feels like, side by side

Worth being concrete, because the abstract comparison hides where the difference lives.

With a good human tutor. They arrive having thought about last week. They notice within two minutes that you are flat today and adjust. They remember that you always rush the setup step, so they make you say it out loud before you write. When you get something wrong they usually know, from having seen it before, which of three misunderstandings you have. At the end they tell your parent one specific thing.

With a good AI tutor. It arrives with a plan built from the exact record of what you got wrong. It is identical whether it is Tuesday morning or Sunday night. It will explain the same idea a fifth way without a flicker of impatience. It cannot tell that you are flat, and will teach the plan at a student who is not in the room mentally. It remembers everything and understands nothing about your week.

The pattern: the human is better at the judgement calls, the AI is better at the consistency. A student who turns up and works gets more from the AI than most families can afford in human hours. A student who is avoiding the subject needs the person.

The hybrid that probably beats both

If you can manage it, the configuration with the best claim on the evidence is not either column.

• AI for frequency. Three or four short sessions a week, doing the teaching, the practice and the scheduled review. This is the dose the research rewards and the dose hourly pricing prevents.
• A human monthly, or fortnightly. Not to teach the content, but to set direction, diagnose what is genuinely stuck, handle exam strategy, and be a person who is expecting them.
• A parent weekly. One conversation where the student explains what they covered. No subject knowledge needed, and it is the check that catches a student who is going through the motions.

That splits the work by what each party is actually good at, and it costs a fraction of weekly private tutoring.

Questions that separate a real AI tutor from a chat wrapper

Ask these of any product, including ours.

1. What does it do before I say anything? If the answer is "wait", it is a chatbot.
2. How does it know I understood? If the answer is "it asks you", that is not a check.
3. What happens when I ask it for the answer? A tutor declines. A tool complies.
4. What does it remember from six weeks ago, specifically? Not "your progress" - which idea.
5. What is it mapped to? A standard, a specification, a rubric a person wrote, or nothing.
6. What does it refuse to claim? A product with no stated limits is selling rather than teaching.

What to demand of an AI tutor

If you go this route, the bar is higher than "it answers questions":

• Does it decide what to teach, or wait to be asked? Waiting is the defining weakness of a chatbot.
• Does it check understanding, or assume it? Ask whether it ever makes the student explain something back.
• Will it refuse to do the work? A tool that completes the problem set is actively harmful.
• Does it remember specifically? Not "you are doing well" but which idea you had backwards in March.
• Can you tell what it covered? Opaque progress is no progress you can act on.

Where TruLearn fits

Those five questions are the product specification we wrote for ourselves. TruLearn arrives with a plan rather than waiting; it asks the student to explain each concept back and grades that against a rubric a person wrote; it is built to let a student struggle rather than rescuing them; and it stores the wrong idea in the student's own words and brings it back days later.

It teaches AP Biology and introductory college biology today, has no parent dashboard, and - as this article insists of everyone else - has no outcome data of its own. Our research page says what we are building on and where that evidence stops.

More: is ChatGPT a good tutor, and what makes tutoring work.