What happens to tutoring when AI can teach

Tutoring works, and it works best at a frequency most families cannot afford. A 2024 meta-analysis of 89 randomised trials found the largest effects for programmes running at least three days a week; at typical rates that is several thousand dollars a year. When the cost of a session stops being an hour of someone's time, the binding constraint on the best-evidenced intervention in education is removed - and that is the change worth thinking about, more than any claim about whether AI is as good as a person.

The economics as they stand

Tutoring is one of the most reliable interventions in education, with a pooled effect of about 0.29 standard deviations across randomised studies. The same research says what predicts a bigger effect: trained tutors, earlier grades, during the school day, and at least three sessions a week.

Now price it. US federal data puts the median tutor wage at $20.84 an hour, and families typically pay well above that once an agency's margin is included. Three sessions a week for a school year runs into several thousand dollars.

So the market offers a thing that works, at a dose most people cannot buy. Families buy one session a week instead, which sits below the frequency the evidence most supports. That is the structural fact the whole industry sits on.

What is being commoditised, and what is not

Commoditised quickly: explaining a standard topic from a standard syllabus. Generating practice. Answering follow-up questions. Being available. Marking work with a rubric.

Not commoditised: working out why a student who should be able to do this cannot. Deciding that today needs encouragement rather than content. Knowing this exam, this teacher, this course. Spotting that the maths problem is a sleep problem. Being a person a young person will show up for.

The first list is most of what is sold in the mid-market. The second list is what the best tutors have always been paid for, and what is hardest to describe on a pricing page.

Three things to expect

The market splits. Cheap, frequent, automated teaching at one end. Expensive human expertise at the other, concentrated on diagnosis, strategy, specialist needs and oversight. The squeeze lands on the middle: the competent non-specialist explaining standard material at $50 an hour.

Frequency becomes normal. If three sessions a week costs less than one does now, the default pattern changes, and the pattern changes toward what the evidence supports. This is the most under-discussed consequence and probably the most important one.

Human time moves up the stack. The tutor's hour becomes more valuable precisely because the routine part is handled elsewhere. A tutor who sees a student monthly to set direction, diagnose what is stuck and keep them going, while automated teaching handles frequency, is a better use of expensive expertise than that same tutor explaining logarithms for the hundredth time.

What should temper all of this

The evidence does not exist yet. The tutoring literature is about human tutoring. Transferring it to AI tutoring is an argument by analogy, and a reasonable one, but it is not a finding.

Some specific reasons for caution:

• Motivation may not transfer. A large part of why tutoring works might be that someone is expecting you. Software does not supply that in the same way.
• The relationship may be load-bearing. Hard to measure, easy to dismiss, and the first thing a human tutor will tell you matters.
• Good explanation is not sufficient. It was never the scarce input. Diagnosis was.
• Unsupervised use can go wrong in ways supervised tutoring does not: a student can get the answer instead of learning.

Anyone confident about the size of this shift is ahead of the data, including the companies building it.

What happens to the people

There are roughly 175,000 people employed as tutors in the United States, plus a large informal market of independents and students. Not all of those roles face the same pressure.

Most exposed: explaining standard material from a standard syllabus to a motivated student. That is the part being automated, and it is the bulk of the mid-market.

Least exposed: specialist diagnosis, learning-difficulty support, exam strategy for a specific institution, motivation work with a disengaged teenager, and anything where a parent is really buying accountability and trust.

Newly valuable: oversight. A tutor who reads what a student has done with an automated tool, decides what is actually stuck, and sets the next month's direction is doing something neither the software nor the family can do alone. That is an hour worth more than the hour it replaces.

The honest version for anybody tutoring now: the hours are likely to change shape before they change in number, and the tutors who move toward diagnosis and strategy will be fine.

What has to be true for the optimistic case

Three things, none of them yet demonstrated:

1. Motivation transfers. Part of why tutoring works may be that a person is expecting you. If that is a large part, software gets a fraction of the effect at any frequency.
2. Unsupervised use is safe. A tool that can be asked for the answer will be, at 1am, by a tired student. Products that refuse are a design choice rather than a given.
3. The effect survives measurement. Nobody has run the trial. Until somebody does, the claim is an argument from mechanism.

If the first fails, AI tutoring becomes a good practice tool rather than a replacement for instruction. That would still be useful and much less interesting.

What a parent should do now

• Do not pay for explanation. It is the cheapest thing in the market and getting cheaper.
• Pay for diagnosis and for accountability. Those are the scarce goods.
• Buy frequency wherever it is cheap. The evidence rewards three sessions a week; get them however you can.
• Keep one outside human in the loop. Someone who is not you and not software, who sees the student periodically and will say if something is wrong.

Where TruLearn fits

Our bet is explicit: the mechanics the evidence favours - a plan made in advance, teaching one student, explain-back graded against a human-written rubric, scheduled review - can be built in software, and doing so makes frequency affordable.

The parts we think stay human are the ones in the second list above, and we say so rather than claiming the whole job. We also have no outcome data of our own, which our research page states. Our intended bar is an external one: results on exams we do not write.

More: AI tutor versus human tutor, and what makes tutoring work.