Bloom's 2-Sigma and the Tutor in the Room

In 1984, Benjamin Bloom published a two-page paper that has haunted education ever since. His research group compared three ways of teaching the same material: a normal classroom, a classroom using mastery learning, and one-on-one tutoring paired with mastery learning. The tutored students scored about two standard deviations higher than the classroom baseline. Put plainly: the average tutored student performed better than 98% of the students in the conventional class. Move a kid from the middle of the room to a good tutor, and they leap from the 50th percentile toward the top of the distribution.

Then Bloom did the honest thing. He called it a problem. Two sigma is a spectacular result and a useless one, because you cannot give every child a private human tutor. The economics never close. So he framed the challenge as a search: find group methods that approach the two-sigma effect without the one-to-one cost. For forty years, that search mostly came up short. The tutor stayed the gold standard nobody could afford.

What the 2-sigma result actually says

It's worth being precise, because the number gets thrown around loosely. Bloom's finding was not "tutoring is nice." It was a specific claim about effect size under specific conditions — and the conditions matter more than the headline.

Two things were doing the work, not one. The first was one-on-one attention: a tutor watches a single learner, notices the exact moment confusion starts, and adjusts. The second was mastery learning: you don't advance to the next unit until you've genuinely got the current one, with feedback and a second attempt built in. Bloom's big number came from stacking both. Classroom mastery learning alone bought roughly one sigma; the tutoring relationship on top bought the second.

That decomposition is the whole reason this essay exists. If the gain were pure human warmth, software would have nothing to offer. But half of it is a loop: diagnose, correct, re-test, don't move on until the learner is ready. A loop is exactly the kind of thing you can encode.

Why it never scaled — and why that just changed

The reason two sigma stayed a curiosity is arithmetic. A tutor costs a salary. A class of thirty amortizes one teacher across thirty kids; one-to-one does not amortize at all. No public system on earth can fund a personal tutor per student for six hours a day, so the method that works best is the one we can least afford. That is the trap Bloom named and could not escape.

Here is the part that should make you sit up. The binding constraint was never pedagogy. It was the cost of a patient, competent adult who will sit with one learner and adapt in real time. For the first time, that cost is not a salary. A per-learner tutor that runs on a language model is close to free at the margin — available at 11pm, on the fourteenth repetition of the same question, without fatigue or a sigh.

The two-sigma problem was never a teaching problem. It was a supply problem — the supply of individual attention. AI does not prove it can teach like Bloom's tutors. It proves, for the first time, that a tutor per learner is affordable. Those are different claims, and the honesty of this whole field lives in the gap between them.

That is the real news, and it is smaller and bigger than the hype. Smaller, because affordability is not attainment. Bigger, because affordability was the wall. Forty years of research assumed the tutor was a fixed, scarce resource. That assumption is now false, and almost none of our institutions have metabolized it.

A mediocre AI tutor is not a 2-sigma tutor

Now the caveat that most of the marketing skips. Presence in the chat window is not pedagogy. A model that cheerfully answers every question is not a tutor — it's a smarter search box, and it may actively lower learning by handing over answers the student should have struggled toward. That is the anti-pattern the first essay in this series is about: an AI tutor that refuses the answer is doing more real teaching than one that supplies it.

Be honest about the gap. A good human tutor does things today's models do unreliably: reads a slumped posture, remembers that you cried over fractions last month, decides when to push and when to back off, calls your bluff when you say "I get it" and you don't. The two-sigma tutors in Bloom's study were good tutors running a disciplined mastery loop. Nothing guarantees a language model reproduces that. Drop a raw chatbot in front of a struggling fourteen-year-old and you may get confident wrong answers, sycophancy, and a learner who feels helped while learning nothing.

So the responsible framing is a hypothesis, not a result: AI has removed the cost barrier that made two-sigma impossible; it has not been shown to reproduce the two-sigma effect. Anyone claiming the latter is selling. The opportunity is real and unproven at the same time, and the work of this field is turning affordability into attainment through better pedagogy — the mastery loop, the refusal to over-help, the assessment loop that writes the test first — not through more availability.

Where AI genuinely moves the needle

Strip out the overclaims and there is still a substantial, defensible list of things a per-learner AI tutor does that a shared classroom teacher structurally cannot:

• Availability. The tutor is there at the moment of confusion, not on Tuesday at 4pm. Most learning breaks happen alone, at night, mid-problem. That is exactly when the classroom tutor is unavailable and the AI one isn't.
• Infinite patience. The fourteenth explanation lands with the same tone as the first. No human tutor sustains that, and the difference matters most for the students who need the fourteenth explanation.
• Per-learner pacing. This is the mastery mechanism, encoded. Don't advance until it's solid; loop back without embarrassment; move fast where the learner is strong. A class of thirty cannot pace to thirty rates. One tutor per learner can.
• Subject routing. A single interface can hand you to a maths specialist, an English specialist, or a wellbeing check depending on what you actually need — the pattern behind subject-routing wellbeing tutors like The Mentor. The learner sees one door; behind it, the right room opens.

None of those four is "teaches as well as Bloom's tutor." Each is "supplies a specific ingredient that scarcity used to ration." Stack them under a real mastery loop and you have a plausible run at the problem. That is what the education-focused skills on the marketplace are trying to build — structured tutors like the KS3–GCSE AI tutors and Professor Pi for KS3 maths — not chat wrappers, but tutors with a subject, a level, and a pedagogy baked in. Whether any of them clears one sigma, let alone two, is an empirical question we should insist on measuring rather than assert.

The arithmetic changed; the burden of proof didn't

Bloom handed us a target and a wall. The target: two standard deviations, the difference between a middling student and a top one, available in principle to every child. The wall: a tutor per learner is unaffordable, so the target is out of reach. For forty years the wall held, and the field spent its energy trying to sneak up on two sigma with clever group methods.

The wall is now gone — not because we got better at teaching, but because the scarce ingredient got cheap. That is a genuine inflection, and it deserves neither dismissal nor breathless certainty. A cheap tutor is a precondition for two sigma, not a delivery of it. The honest position is to hold both truths at once: the economic impossibility Bloom named is over, and the pedagogical proof he'd have demanded has not yet been produced. Build for the first, measure against the second — and browse the AI tutoring skills as attempts, not answers. The tutor finally got into the room; now we have to prove it can teach.

---

Part 7 of the AI Tutoring series. Previously: Explain It Like I'm Twelve. Next: Metacognition Is the Real Curriculum. Browse the AI tutoring skills or more builder insights.