Redistributed

In 1984, the education researcher Benjamin Bloom published a finding so striking it became known simply as "the 2 sigma problem." Students who received one-on-one tutoring performed two standard deviations better than students in conventional classrooms. That's the difference between an average student and one in the 98th percentile. The problem was cost: you can't give every student a private tutor. Bloom's paper was essentially a challenge — find a way to deliver tutoring-level instruction at classroom scale.

Forty years later, AI companies claim they've solved it. Khan Academy's Khanmigo, powered by GPT-4, uses a Socratic method to guide students through problems rather than handing them answers. Duolingo Max offers AI conversation partners for language practice. Dozens of startups are building AI tutoring systems for everything from calculus to chemistry. And millions of students aren't waiting for official tools — they're already using ChatGPT and Claude as de facto study partners, homework helpers, and essay editors. A 2024 survey by the Digital Education Council found that 86% of university students use AI tools regularly for coursework (Digital Education Council, 2024).

The pitch is seductive: personalised learning at scale, available to anyone with an internet connection. The reality is more complicated — and the redistribution questions are sharper than the edtech marketing suggests.

Where AI Tutoring Actually Works

Credit where it's due — and there is genuine credit to give. AI tutoring is genuinely useful for certain kinds of learning. Structured problem-solving — algebra, physics exercises, coding syntax — is where these tools shine. The student makes an attempt, the AI identifies the error, offers a hint, and guides them toward the solution. This feedback loop, delivered instantly and patiently at any hour, is something no classroom of thirty students can replicate.

Language learning benefits similarly. Duolingo's AI conversation feature lets learners practice speaking with a partner who never gets impatient and adjusts difficulty in real time. For factual recall — memorising vocabulary, historical dates, scientific nomenclature — AI-driven spaced repetition can optimise review schedules with a precision no human tutor would bother to calculate.

Khan Academy reports that students using Khanmigo show measurable gains in math proficiency, particularly among students who were previously struggling. The mechanism makes sense: struggling students benefit most from patient, individualised feedback, which is exactly what AI can provide cheaply and consistently. This is not a trivial achievement. Building systems that work patiently with struggling learners at scale is hard engineering, and the teams doing it deserve recognition for tackling one of education's most persistent problems.

Where It Falls Apart

But education is not primarily a problem of information delivery. If it were, textbooks would have solved it centuries ago.

Critical thinking — the ability to evaluate arguments, weigh evidence, and construct original analysis — develops through dialogue with minds that can genuinely disagree, that have their own perspectives, that can be surprised. An AI that generates Socratic questions from a rubric is performing the surface pattern of inquiry without its substance. It can ask "why do you think that?" but it cannot be genuinely perplexed by your answer.

Creative expression suffers similarly. Writing is not a skill you optimise; it's a way of thinking you develop through struggle, feedback from readers who have their own aesthetic commitments, and the hard work of figuring out what you actually mean. An AI that polishes your prose is not teaching you to write. It's teaching you to outsource writing.

Then there's what schools actually do beyond instruction: socialisation, emotional development, exposure to difference, the experience of being part of a community. Ursula Franklin described education as fundamentally a "growth model" — one that requires time, attention, and the unpredictable dynamics of human relationships — as opposed to a "production model" that treats learning as something to be optimised (Franklin, 1989). AI tutoring, however sophisticated, is a production model. It can optimise for measurable outcomes. It cannot replicate the unmeasurable ones that matter most.

The Integrity Crisis

The most immediate disruption AI has brought to education is not tutoring — it's cheating. Or, more precisely, it's the collapse of assessment methods that assumed students would do their own work.

Millions of students now use AI to draft essays, solve problem sets, and complete assignments. Universities are scrambling to respond. Some have banned AI tools entirely. Others have tried to embrace them, redesigning assignments to incorporate AI use. Many are stuck in an uncomfortable middle, issuing policies that are essentially unenforceable.

AI detection tools have emerged to fill the gap — and they're unreliable. A 2023 Stanford study found that GPT detectors disproportionately flag writing by non-native English speakers as AI-generated, creating a bias that punishes the students most in need of support (Liang et al., 2023). The false positive rate is high enough that no responsible institution should use these tools as primary evidence of cheating. Yet many do.

The deeper problem is structural. For decades, education systems have relied on essays and take-home assignments as proxies for understanding. AI hasn't created a new problem — it has exposed the fragility of assessment methods that were already imperfect. The question isn't how to detect AI use. It's how to redesign assessment so that it measures what we actually care about: whether the student understands the material, can reason about it, and can apply it in novel contexts.

The Teacher Question

Will AI replace teachers? The edtech industry's preferred framing is "AI won't replace teachers, but teachers who use AI will replace those who don't." This is a convenient line — it sells AI tools while appearing to honour the profession. It also sidesteps the harder question.

The economic logic of AI in education points not toward replacement but toward deskilling. If AI handles instruction, assessment, and personalisation, the teacher's role narrows to classroom management and emotional support — important work, but work that is typically valued less and paid accordingly. As Daron Acemoglu and Simon Johnson argue in Power and Progress, the critical question with any new technology is whether it's designed to augment workers or to replace their judgment. AI tutoring systems that position the teacher as a supervisor of automated instruction are, in practice, reducing the skill content of teaching (Acemoglu & Johnson, 2023).

The alternative — AI that handles administrative burden (grading routine assignments, tracking progress, generating practice materials) while freeing teachers to focus on mentoring, discussion, and the complex work that requires human judgment — is technically possible. Some schools are already pursuing this model. Arizona State University's AI-assisted courseware, for example, lets instructors see where students struggle in real time, creating more informed human teaching rather than replacing it. But scaling this approach requires deliberate design choices that prioritise teacher agency over platform control. The incentive structure of edtech companies does not naturally produce these choices — though the best ones, like Khan Academy, are trying.

The Redistribution Question

Here is where the infrastructure framing bites hardest. AI tutoring could be the most equalising force in education since the public library. Or it could be the next mechanism for widening the gap between rich and poor students. The outcome depends entirely on access, governance, and design.

The cost barrier is real. Khanmigo costs $44 per year for students, $99 per year for families — modest by American standards, but a meaningful expense for low-income households, and prohibitive for much of the Global South. Premium AI tutoring tools from startups can run to hundreds of dollars per month. Wealthy families are already hiring "AI learning coaches" to help their children use these tools effectively — tutors for the tutoring software. The pattern is familiar from every previous educational technology: those with resources adopt first, adapt fastest, and pull further ahead.

The school funding gap compounds the problem. Schools in wealthy districts have IT infrastructure, professional development budgets, and technology coordinators who can integrate AI tools thoughtfully. Underfunded schools — disproportionately serving Black, Latino, and rural students — often lack reliable internet access, let alone the institutional capacity to deploy AI tutoring effectively. A 2024 RAND Corporation study found that high-poverty schools were significantly less likely to have implemented AI tools, and those that had were less likely to have provided teacher training on their use (RAND Corporation, 2024).

The Global South is largely invisible in this conversation. AI education tools are overwhelmingly designed for English-speaking students following Western curricula. A student in rural Bihar, a teenager in Lagos, a child in rural Guatemala — they are not the target market. Their languages, their educational contexts, their needs are afterthoughts at best. The promise of "personalised learning for every student on earth" rings hollow when the tools don't work in most of the world's languages and don't reflect most of the world's educational traditions.

Then there's the data. AI tutoring systems collect granular data on student performance — every mistake, every hesitation, every area of weakness. This data is extraordinarily valuable, both for improving the tools and for training the next generation of AI models. When a child uses an AI tutor, they're not just learning — they're generating training data. The terms of this exchange are rarely made explicit to students or parents. As Kate Crawford documents in Atlas of AI, the extraction of value from human activity — including the activity of children learning — is a foundational dynamic of the AI industry (Crawford, 2021).

Who Benefits, Who Pays

Who benefits: Students with access, support, and the digital literacy to use AI tools effectively. Edtech companies building and selling these tools. AI providers whose models power them. Wealthy families who can layer AI tutoring on top of already-advantaged educational environments.

Who pays: Students without access, who fall further behind as the baseline shifts. Teachers whose professional judgment is subordinated to platform logic. Educational equity, as a new axis of advantage opens up along familiar lines of income, geography, and language. And students themselves, whose learning data fuels an extraction cycle they didn't consent to and don't benefit from.

What Equitable AI Education Would Look Like

The gap between what AI education could be and what it's becoming is not inevitable. It's a design choice. An equitable approach would require:

Free, high-quality AI tutoring as public infrastructure. If AI tutoring works — and in structured domains, the evidence says it does — then it should be available to every student, not just those who can pay. There is precedent: the Rural Electrification Administration (REA) brought electricity to 90% of American farms within two decades, not because the market found it profitable but because political will made it happen. Khan Academy's mission of free education for anyone, anywhere, is the right frame, but it requires public funding, not dependence on philanthropic goodwill and corporate partnerships.

Teacher-controlled deployment. Teachers, not platforms, should decide how AI tools are used in their classrooms. This means tools designed to augment teacher judgment, not to circumvent it. It means professional development that gives teachers genuine expertise, not just training on a specific product.

Robust data privacy protections. Student learning data should not be available for commercial AI training without explicit, informed consent. Children's educational records deserve at least the same protections as their medical records. FERPA, the US federal law governing educational records, was written decades before AI and is badly in need of updating.

Multilingual, culturally relevant tools. AI education tools should work in the languages students actually speak and reflect the curricula they actually follow. This requires investment in language models for underserved languages — an area where open-source efforts like AI2's OLMo and community-driven projects like Masakhane and Mozilla Common Voice could play a crucial role, if funded. The rise of small language models makes this more feasible than it was even two years ago: a well-tuned 7B model running locally can serve as an effective tutor without requiring expensive cloud infrastructure.

Assessment reform. Rather than an arms race between AI-assisted cheating and AI detection, education systems need assessment methods that AI can't trivially game: oral examinations, project-based learning, portfolios, demonstrations of skill. These methods are more expensive and harder to scale. They're also better measures of actual learning.

The Deeper Question

There is a question beneath all the policy specifics that rarely gets asked plainly: what is education for?

If the purpose of education is to optimise test scores and credential students for the labour market, then AI is a powerful tool. It can drill math facts, polish essays, and prep students for standardised tests with superhuman patience and precision. Measured by the metrics that currently dominate education policy, AI tutoring will look like an unambiguous success.

But if the purpose of education is to develop humans — to cultivate judgment, creativity, ethical reasoning, the capacity to live in a community and participate in democratic self-governance — then AI's contribution is, at best, partial. The skills that matter most are precisely the ones that emerge from human relationships, from struggle, from the unpredictable dynamics of a classroom where thirty different minds encounter the same idea and make thirty different things of it.

Bloom's 2 sigma problem was never really about finding a cheaper way to deliver instruction. It was about what happens when a student is truly seen — their specific confusion addressed, their particular strengths recognised, their individual development taken seriously. AI can simulate some of this. It cannot do it. The question is whether we'll use AI to create the conditions for more genuine human attention in education, or whether we'll use it as a justification for providing less.

The technology is ready. The infrastructure decisions — who pays, who profits, who gets access, whose data gets extracted — are being made now. They will determine whether AI in education narrows the gap or widens it. Right now, the trajectory favours widening.

Sources

Bloom, Benjamin S. "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring." Educational Researcher 13, no. 6 (1984): 4–16. JSTOR
Liang, Weixin et al. "GPT Detectors Are Biased Against Non-Native English Writers." Patterns 4, no. 7 (2023). Cell
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
Acemoglu, Daron & Johnson, Simon. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs, 2023. Publisher
Franklin, Ursula. The Real World of Technology. Anansi, 1989. Publisher
Digital Education Council. "2024 Global AI Student Survey." 2024. DEC
RAND Corporation. "The State of AI in K–12 Education." 2024. RAND

General

AI and Education: Tutoring, Testing, and Who Gets Left Out