Two-Sigma Problem

An educational-psychology finding by Benjamin Bloom (University of Chicago, 1984): students given one-on-one tutoring with mastery learning outperformed students in conventional classroom instruction by roughly two standard deviations — the average tutored student scored above the 98th percentile of the classroom control group. Bloom named the paper "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring."

The word problem is the load-bearing part. Bloom's own framing: tutoring at scale works, and no education system could ever afford to deliver it. The paper is a challenge to find classroom methods that close the gap without the cost.

The finding in plain form

Control group: conventional classroom (1 teacher, ~30 students, standard curriculum).
Treatment 1: mastery learning (formative assessment, corrective feedback, retest until 80% competency before moving on) → ~1σ improvement.
Treatment 2: one-on-one tutoring using mastery-learning methods → ~2σ improvement.

The 2σ figure means the average tutored student outperformed 98% of the classroom students. That's an enormous effect size; if it held generally at that magnitude, it would restructure education globally.

The replication story (important honesty)

Bloom's exact 2σ result has proven hard to fully replicate. Later studies typically find:

Tutoring effect is real — every meta-analysis since finds tutoring meaningfully outperforms classroom instruction.
Magnitude is often smaller — often 0.4σ to 1σ in careful contemporary studies, depending on age, subject, tutor training, dosage, and how tightly the mastery-learning protocol is enforced.
Bloom's methodology had limitations — small samples, specific curricula, high-motivation contexts.

The honest one-line: tutoring beats classroom instruction by a lot; whether "a lot" equals Bloom's 2σ specifically is contested; whether the effect is real is not.

Why it matters for AI

The Two-Sigma Problem is the theoretical anchor for the argument "AI can democratize world-class tutoring." If Bloom's original 2σ (or even a discounted 1σ) result is real, and if AI can plausibly deliver a version of one-on-one interaction at scale, the implication is a step-change in learning outcomes at a fraction of the cost.

The claim is common now in ed-tech (Khan Academy's Khanmigo, OpenAI's education partnerships, Anthropic's Claude for Education, plus every consumer-facing "AI tutor" startup), but it has an obvious weak spot: no published measurement yet establishes that AI tutoring produces the 2σ effect — or even reliably matches a human tutor. The mechanism has plausibility; the empirical grade is still unresolved.

Where this cuts:

Pro-AI-tutor — even a fraction of Bloom's 2σ, delivered universally, would be transformative. And the marginal cost of AI-tutor delivery is close to zero.
Skeptical — Bloom's effect included a human who could read affect, catch confusion the student couldn't articulate, and hold them accountable. AI tutors currently miss most of that layer; a confident-wrong tutor is worse than no tutor for a novice who can't spot the error.

In Sandeep Swadia's ALTER Framework

Sandeep names the Two-Sigma Problem as the theoretical justification for the Tutor role in his five-role AI-education stack — see How To Become Dangerously Self-Educated With AI (theMITmonk) and ALTER Framework. His specific claim: "AI completely breaks the economics of human intelligence." His practical form: voice mode (Gemini voice, ChatGPT voice), with the two prompts "teach me" and "test me" framing an interaction that's supposed to feel like an office hour, not a chat window.

Sandeep does not name the replication caveat — his citation of Bloom's number is casual and delivered as settled canon. That's a minor honesty gap on the source page; noted here so the vault carries the nuanced version.

Adjacent concepts in this vault

ALTER Framework — Sandeep's five-role AI-education stack; Two-Sigma is the Tutor role's justification
TRAP Framework — Sandeep's retention practice; Two-Sigma is about delivery-modality effect size, TRAP is about practice-discipline effect size — different levers on the same outcome
Fluency Illusion — a good tutor's job is to break the fluency illusion; the mechanism through which Bloom's 2σ effect likely operates in part
Desirable Difficulties — the tutor forces the desirable difficulty the student would skip on their own; another mechanism candidate
Higher-Order Thinking — Justin Sung's concept; a Bloom-quality tutor is one of the fastest paths to developing it because the tutor forces relational reasoning rather than fact recall

Sources

How To Become Dangerously Self-Educated With AI (theMITmonk) — canonical vault mention; Sandeep Swadia's framing as the anchor for AI-as-tutor
Bloom, B.S. (1984). "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring." Educational Researcher, 13(6), 4–16. — the original paper; not in this vault as a source page (not clipped)