In 1984, a Researcher Proved One-on-One Tutoring Could Double Learning. Forty Years Later, an AI Did It for $4 a Month.

Two standard deviations. In education research, that number has haunted the field for four decades. In 1984, Benjamin Bloom published a study in Educational Researcher showing that students who received one-on-one tutoring performed two standard deviations above students taught in conventional classrooms. An average student, given a personal tutor, performed better than 98% of classroom-taught peers.

Bloom called it the "2-sigma problem" because the result was easy to produce and impossible to scale. A private tutor costs $40 to $80 per hour in 2025 dollars. A hundred hours a year of meaningful tutoring runs $4,000 to $8,000 per student. Multiply that by 50 million K-12 students in the United States at a $6,000 midpoint and you arrive at $300 billion per year, roughly 38% of the nation's entire K-12 education budget. No district, no state, no country could afford it. So the 2-sigma effect remained what it has always been: proof that the system fails most children by design.

In June 2025, a team at Harvard published a randomized controlled trial in Scientific Reports. They split 194 students into two groups: one received a pedagogically designed AI tutor, the other participated in active classroom learning with an experienced instructor. Students in the AI group achieved effect sizes of 0.73 to 1.3 standard deviations above the classroom group. They did it in 49 minutes instead of 60. And when tested afterward without the AI present, the gains held.

$300 Billion vs. $2.2 Billion

Nobody has published this cross-calculation, and once you see it, the 2-sigma problem looks different.

Approach	Annual Cost per Student	Effect Size	Cost per 0.1σ
Private human tutor	$4,000–$8,000	~2.0σ	$200–400
Khanmigo (district pricing)	~$44/year	0.23–0.31σ	$14–19
AI tutor (Harvard design)	~$20–50/year (API cost)	0.73–1.3σ	$2–7

At $44 per student per year, putting Khanmigo in front of all 50 million US K-12 students costs $2.2 billion. Putting a human tutor in front of all of them costs $300 billion. That is a 136x cost reduction for a tool that, at its best, delivers 37 to 65% of the human tutor effect. At its current real-world scale, the number is more modest: a WestEd longitudinal study of over 2 million Khanmigo users found a 0.23σ improvement in math and 0.31σ for English Language Learners. That moves the 50th percentile student to the 59th. Not the 98th. But for $44.

Cost per 0.1 standard deviations of improvement tells the story more precisely. A human tutor delivers each tenth of a sigma for $200 to $400. Khanmigo at scale does it for $14 to $19. A Harvard-quality AI tutor, running on API costs alone, does it for $2 to $7. Even if AI tutoring never reaches Bloom's 2-sigma ceiling, the cost-effectiveness ratio is 10 to 20 times better per unit of learning gained.

Why 0.23σ Is Not 1.3σ

Harvard's result came from a controlled environment: 194 undergraduates at an elite university, studying physics, with a tutor purpose-built by researchers who understood both the subject matter and the cognitive science of learning. Khanmigo's 0.23σ came from 2 million students across thousands of schools, many of them underfunded, with varying levels of internet access and teacher support.

That gap is not a mystery. It has a name, and it was identified before any AI tutor shipped.

A Stanford CEPA study in 2025 found that student engagement with AI tutors drops 60% after three weeks when there is no teacher facilitating the process. Students open the tool, click around, and stop. Without a human in the room who knows their name, knows they skipped breakfast, knows they failed the last quiz, the AI is just another app on a laptop that already has YouTube and Roblox installed.

This is the first of three failures blocking deployment.

Three Failures

Failure 1: Engagement collapse. AI tutors work when teachers are in the loop. Stanford CEPA's data is unambiguous: among K-12 students using AI tools without structured teacher facilitation, engagement dropped 60% over a single semester. Khan Academy's own results confirm it. Students who used Khanmigo 30 or more minutes per week saw measurable gains. But getting students to that threshold required teacher involvement. Districts that deployed the tool without professional development for teachers saw minimal impact. The technology is not self-executing.

Failure 2: Privacy violations. A US Department of Education report in 2025 found that 78% of EdTech AI tools used in schools do not comply with FERPA's updated digital provisions. Student learning data, including misconceptions, emotional states, and pacing patterns, is being collected with unclear retention policies and no federal standard for AI-specific data governance in K-12. Parents signing acceptable-use policies often have no idea what data is being recorded or where it goes. Schools are deploying AI tutors faster than they are writing data governance policies for them.

Failure 3: Access gaps. According to the Learning Policy Institute, 411,549 teaching positions in the US are currently unfilled or filled by non-certified teachers. These districts, predominantly rural and low-income, are the ones that would benefit most from AI tutoring support. They are also the ones with the worst internet connectivity, the oldest student devices, and the fewest IT staff to deploy and maintain AI tools. Districts with the best infrastructure will deploy AI tutoring first, meaning the technology could widen achievement gaps before closing them. Broadband access in rural school districts remains spotty despite federal investment. And NAEP 2024 data shows reading scores still declining and math barely recovering from COVID losses. The students falling furthest behind are the hardest to reach with any digital intervention.

What the Critics Get Right

The strongest case against AI tutoring comes from the OECD's Digital Education Outlook 2026, a 247-page report published in January. It found that students using general-purpose chatbots like raw ChatGPT improved their assignment output but showed gains that disappeared on exams taken without AI access. In other words: they learned to get answers from the machine, not to think for themselves.

This is entirely valid. And it is also not what the Harvard study measured. Kestin's team built a tutor that used Socratic questioning, scaffolding, and cognitive load management. It did not give answers. It asked questions. When students got stuck, it broke the problem into smaller pieces. When they guessed correctly without understanding, it pushed them to explain why. And the post-test was administered without the AI. Students had to perform on their own.

The OECD's own report draws this distinction explicitly: purpose-built tutoring tools with pedagogical architecture produce sustained learning gains. Generic chatbots produce shallow performance improvements that vanish under independent testing. The model's capability matters less than the pedagogical design wrapped around it.

But here is the honest concession. The Harvard study enrolled 194 students in a single physics course at one of the most selective universities in the world. These are students who got into Harvard. They are already exceptional learners. Whether the same AI tutor design would produce 0.73σ with a struggling third-grader in a Title I school in rural Mississippi is genuinely unknown. Khanmigo's 0.23σ at actual scale, with actual kids, in actual classrooms, is the number we can trust. It is real, it is replicable, and it is modest.

Limitations

Several caveats apply to the analysis above. Bloom's original 2-sigma result has been critiqued for small sample sizes and controlled conditions that may not generalize to all subjects or age groups. The Harvard RCT, while gold-standard in design, is a single study with N=194 at an elite institution; replication across demographics, subjects, and grade levels is needed before the 0.73–1.3σ range can be generalized. The $300 billion human tutoring figure assumes 100 hours per student at the midpoint of current market rates; actual costs vary significantly by region and tutoring format. The 136x ratio compares Khanmigo's district pricing to human tutoring, but Khanmigo's costs are partially subsidized; full long-run costs may differ. The 78% FERPA non-compliance figure is from a Department of Education survey of EdTech tools broadly, not AI tutors specifically. And Stanford CEPA's 60% engagement drop has not yet been published as a peer-reviewed paper with full methodology available for independent scrutiny.

Where This Leaves Us

Bloom identified the ceiling in 1984: two standard deviations, achievable through individual human attention, unaffordable at any meaningful scale. Forty-one years later, the Harvard data suggests AI can reach 37 to 65% of that ceiling. Khanmigo, at actual scale with actual students, reaches about 12%. Both numbers will improve as models get better and pedagogical designs get sharper.

But the 2-sigma problem was never just about tutoring effectiveness. It was about the gap between what we know works and what we can deliver. AI narrows the cost side of that gap by two orders of magnitude. It does not solve the human side: the teacher who notices a student is disengaged, the IT staff who keeps the Chromebooks running, the parent who asks what their child's data is being used for.

The technology costs $4 a month. Making it work costs something else entirely.