How much a tutor costs, when you actually need one, and what an AI tutor can and cannot replace: a human tutor, an AI tutor and self-testing, compared honestly

A private tutor's hourly rate varies by city, subject, the tutor's qualifications and whether the lesson is in person or online, and a page that gives you one number is guessing. What you can decide without the number is whether you need one: your own error log says so when the same skill comes back as "did not know" after two rounds of self-testing. Here is how to read that, what a human tutor does that nothing else does, what an AI tutor does well and badly, where free tutoring is, and how to spend a tutor's hour so it is not wasted.

How much a tutor costs: what the rate depends on

Six things set the rate, and they pull in different directions.

• The city. A big city costs more than a small one, for tutoring as for rent.
• The subject and the level. A Grade 12 calculus or chemistry tutor charges more than a Grade 6 reading tutor, because fewer people can do it.
• The tutor's qualifications. A university student in the subject, a certified teacher, and a specialist in a learning difference are three different prices, in that order.
• In person or online. Online is usually cheaper and the pool of tutors is far larger.
• One-to-one or a small group. A group of three splits the hour three ways.
• A centre or an independent. A tutoring centre carries rent and a brand; an independent tutor carries neither.

We looked for a public, citable range, from a statistics agency or a school board or a university's own tutoring page, and did not find one we could stand behind. So this page prints no price. The marketplaces publish rate charts, and those are advertising. Ask three tutors in your city and you will have the number in an afternoon.

What you can compare without a price is the hours. One tutor hour a week for eight weeks is 8 hours. Twenty minutes of self-testing every day for the same eight weeks is 20 × 56 = 1,120 minutes, about 18.7 hours. The second is free and more than twice as long. It is also harder to keep up, which is the honest reason tutoring exists.

When you need a tutor: the error log decides

A tutor is for a specific problem, and the way to know you have one is a record of your own work. The error log is that record: every wrong answer in practice, with the skill it belongs to and the kind of error (did not know how; knew how and slipped; ran out of time). Four signs in the log say a tutor's hour is worth it.

1. The same skill logged "did not know" after two rounds. You tested yourself, redid the worked examples, tested again, and it is still wrong. Self-study has stalled on that skill, and a person who can see your line is the fastest way through.
2. A gap from a previous course that this course assumes. Calculus assumes twelve ideas from the functions course, listed on the precalculus page; a student missing three of them is not failing calculus, they are missing Grade 11, and a tutor can fill a known gap quickly.
3. An exam with a date, and a list of skills longer than the weeks left. The log says eight skills; the calendar says three weeks. A tutor compresses.
4. A diagnosed learning difference. Then the right tutor is a specialist, and the conversation starts with the school, which may already provide one.

And the sign that is not a sign: one bad mark with no log behind it. The first move after a bad test is the log, not the phone, because a tutor cannot work from "I don't get it". A week of honest self-testing turns "I don't get it" into "these three skills", and that is a sentence a tutor can use.

What a human tutor does that nothing else does

Benjamin Bloom (1984) claimed that students tutored one to one performed about two standard deviations above students taught in a class, the "two sigma" result. It is quoted everywhere, and it is larger than the later, broader evidence found. VanLehn (2011) collected the controlled comparisons and put human tutoring at about 0.79 standard deviations over ordinary instruction, with well-designed tutoring software close behind it and both well below two sigma. The most recent large review, by Nickow, Oreopoulos and Quan (2020), pooled the randomised studies of tutoring programs and found an average effect of about 0.37 standard deviations, which still puts tutoring among the most effective things anyone has measured in schools. One honest note: nearly all of those studies are of structured programs, several sessions a week with trained tutors and a set curriculum, not of a private hour on Tuesdays. The research says tutoring works; it does not say that any hour of it works.

What makes it work is specific. VanLehn's reading of what tutors actually do is that the gain comes from the tutor letting the student do each step, noticing the step that fails, and correcting that step. What an AI tutor should do is built on the same finding. A tutor who lectures for the hour is a textbook with a voice.

What the research leaves to the person is what no software has: seeing confusion in a face before the student says anything, keeping a student going for months when the marks have not moved yet, a relationship that makes the student do the homework, and the judgement to throw out the plan when it is not working.

What an AI tutor does well, and where it fails

The standard for an AI tutor is the same as for a person: it should ask what you tried, hint before it answers, check your answer against a verified key rather than its own guess, show you the source, say "I was not given that" when it was not, and remember your gaps. An AI tutor that does those things does some of them better than a person. It is there at any hour, every day. It does not get tired of the same question. It checks your maths at once, in any equivalent form, instead of a week later in red pen.

Where it fails is the other list. It cannot see you. A language model's own solution is a prediction, and on a long calculation it can be confidently wrong; a tutor built on a verified answer key does not have that problem, and one that generates its own answers does. It does not know your teacher's rubric unless you show it. And a tool that only answers is not a tutor at all; it is an answer bot, and practice in which the answer arrives before your attempt skips the retrieval that does the teaching. What AI is good for in studying draws that line in detail.

Self-testing: the free option the research ranks first

Before either kind of tutor, the two methods with the strongest evidence in all of learning research cost nothing. Retrieval practice, testing yourself before you look, is the finding Roediger and Karpicke (2006) measured: students who took a test on material remembered far more a week later than students who reread it. Spacing, coming back to an idea on separate days, is the finding Cepeda and colleagues (2008) measured across more than a thousand people.

Both need two things a student alone may not have. A source of checked answers, so the self-test is honest. And the discipline to do twenty minutes every day for eight weeks. Those two gaps are exactly where a tutor, human or AI, earns its place: the AI tutor supplies the checked answers and the patience; the human supplies the months.

The honest comparison, in one table

Every cell in words. No prices, because none would be honest.

	Human tutor	AI tutor	Self-testing
Cost	By the hour; varies by city and subject	Free to modest; often included with other tools	Free
Available	At the booked hour	Any hour, every day	Any hour, if you sit down
Diagnoses the exact wrong line	Yes, by watching you work	Yes, if built to check your working	Only if you mark honestly
Notices confusion	Yes, in your face	No	No
Checks against a verified answer	Usually, from the text or the key	Yes, if built on a key; no, if it generates its own	Yes, with an answer key; no, without
Keeps you going over months	Yes, the relationship does	Partly, by reminders	Only your own habit
Adapts the plan	Yes	Partly, from your record	Only if you reread your log
Can be wrong without knowing it	Sometimes; a good one says so	Yes, if it generates answers; rarely, if it checks against a key	Yes, if you mark your own work loosely

The line under the table: the three are not rivals. Self-testing is the base, every day. The AI tutor sits beside it for the checks and the hints. The human tutor is called in where the log says the base has stalled.

Where free tutoring is

Before paying anyone, the free list, in the order to try it.

• Your school. Peer tutoring, run by senior students, and your teacher's own office hours or lunch help. The teacher is the one person who knows the rubric.
• The public library. Many run homework programs for school students in the evening.
• A university's drop-in centre. For first-year courses, most universities run a math or writing help centre where a graduate student sits with you. It is included in your fees.
• In Ontario, the province's free online math tutoring. Students in Grades 4 to 12 at publicly funded English-language schools can book one-to-one online sessions with certified Ontario teachers, through the provincial educational broadcaster and funded by the Ministry of Education. The grade range and the hours are on its sign-up page.
• Community programs, through a community centre or a place of worship.

What to say when you arrive, wherever it is: "Here is my error log. These three skills." That sentence turns a drop-in hour from a lecture into a lesson.

Worked example: one student, one decision

A Grade 11 student in the functions course has kept an error log for two weeks. Three skills are still "did not know" after two rounds of self-testing and a redo of the worked examples: factoring a trinomial with a leading coefficient other than 1, finding the domain of a rational function, and solving an exponential equation. The unit test is in three weeks.

The four signs, checked: the repeated did-not-know is present (sign 1); the factoring skill is a Grade 10 idea this course assumes (sign 2); the test has a date and the list is three skills long, which is not longer than the weeks (sign 3 is not present); no learning difference (sign 4 is not present). Two signs present is enough.

The decision: four tutor hours over the three weeks, one a week plus one the week of the test, on those three skills only, with the daily twenty minutes of self-testing continued. The hours: 20 minutes × 21 days = 420 minutes = 7 hours of self-testing, plus 4 hours with the tutor. The tutor's hours are about a third of the total, and they are spent only on the three skills that stalled.

What the student brings to hour one: the log, the three skills named, and two wrong attempts at each, so the tutor can see the line that goes wrong without spending the hour finding it. What the hour should look like: the student writing, the tutor hinting and naming the wrong line, and the last ten minutes a cold redo with the tutor silent.

How to spend a tutor's hour

• Bring the log, not the textbook. Name the two or three skills and show two wrong attempts at each.
• The student writes, the tutor talks. A tutor who lectures for the hour is a lecture you are paying for. Ask for a hint, try it, and let the tutor say which line went wrong.
• The last ten minutes: a cold redo. The hour's skill, from a fresh question, with the tutor silent. If it works, the hour worked. If it does not, you know what next week is.
• The homework is the next error log. Twenty minutes a day of self-testing on this week's skills, marked, logged, and brought to the next hour. How to study math has the five-step cycle those twenty minutes run on, and chapter 7 of the free playbook has the questions to ask of any tutor, human or AI.

The decision, on one screen

• The rate depends on six things: the city, the subject and level, the tutor's qualifications, in person or online, one-to-one or group, centre or independent. No single number is honest.
• Four signs you need one: the same skill did-not-know after two rounds; a gap from a previous course; an exam date closer than the list of skills; a diagnosed learning difference. One bad mark with no log is not yet a sign.
• The three-way table: self-testing is the free base; the AI tutor adds checked answers and patience at any hour; the human tutor adds the eyes, the months and the judgement.
• Free help first: the school, the library, the university's drop-in centre, Ontario's free online math tutoring for Grades 4 to 12.
• How to spend the hour: the log, two wrong attempts per skill, the student writing, a cold redo at the end, the next log as homework.