The homework apocalypse already happened: this is an assessment crisis, not a learning crisis (and the universal tutor isn't here yet)

THE THESIS. AI hasn't broken education. It has broken the contract that held it together: send work home and assume the student did it. That contract is gone. 84% of high school students use AI for homework, according to a 2025 College Board study cited by Fortune, and only 29% to 33% of schools have a formal AI policy. Our thesis has three parts. One: detection is a lost battle, and chasing it causes more harm than the cheating it aims to stop. Two: the scientific evidence says AI tutoring works, but only with specific pedagogical design and humans in the loop. Three: the bottleneck for the 'universal personal tutor' is neither technology nor price; it is student motivation. That is where it will be decided whether AI narrows educational inequality or widens it.

DETECTION IS DEAD. The numbers are brutal. An analysis of 14 AI-detection tools, reported by Fortune, found false positives (accusing an innocent student) of up to 50% and false negatives (missing actual AI use) of up to 100%. Worse: 61.3% of texts written by non-native English speakers were wrongly flagged as AI. In other words, detectors mostly punish students who write in simple sentences, not students who cheat. Our reading is blunt: any school building its academic integrity on a detector is building on sand — with a built-in bias against its most vulnerable students. The serious answer isn't technological; it's design: in-person, oral, and process-based assessment (defend what you wrote, show your drafts, solve it in class). It costs more teacher time. There is no shortcut.

WHAT THE EVIDENCE SAYS. Here we need to separate signal from noise, because studies point both ways. In favor: a randomized controlled trial (RCT, the gold standard — one group with AI, one without, assigned at random) published in Nature Scientific Reports found that a well-designed AI tutor in a Harvard physics course made students learn more than twice as much as the best active-learning classroom, in less time. In Nigeria, a World Bank program with 800 students and six weeks of teacher-guided AI tutoring improved scores by 0.3 standard deviations (a measure of effect size; 0.3 is a notable effect in education), which its authors equate to 1.5 to 2 years of business-as-usual schooling. Caveat: several critics note that one outcome measure was too closely aligned with the tool itself, and that the 'years of learning' framing is generous. We attribute the claim; we don't buy it whole. Against: a recent arXiv study, 'Faster Completion, Less Learning', documents that when AI simply hands out answers, students finish math exercises faster and learn less. The pattern emerging from all of this — confirmed by Brookings' review of four RCTs — is consistent: AI that forces you to think (Socratic questions, hints, elaborated feedback) teaches; AI that thinks for you, atrophies.

THE 15% PROBLEM. And here comes the most uncomfortable number of the year, because it dismantles the easy story of a 'universal tutor already available'. Khanmigo, Khan Academy's AI tutor and the sector's most serious project, went from 40,000 to 700,000 students with access in the 2024-25 school year. But in April 2026, Sal Khan admitted that only 15% of students with access actually used it, and that for most it was, in his words, 'a non-event'. Khan Academy is redesigning it to offer proactive help starting this summer. The lesson is huge: Bloom's tutoring problem (the old evidence that one-on-one tutoring dramatically boosts learning) was never just about cost. It was also about wanting to. A human tutor seeks you out, knows you, doesn't let you slip away; a chatbot waits patiently for you to show up. And the student who shows up on their own is usually the one who already had motivation, family support, and study habits. Left uncorrected, the free AI tutor could end up widening the very gap it promised to close: equal access, unequal use.

OUR READING. The public debate is framed wrong. We argue 'AI yes or AI no' when the evidence says the decisive variable is the mode of use: the same technology produces the best documented result in university physics and the hollowing-out of learning in math homework. This connects with our running thesis on jobs and education: the teacher who orchestrates AI wins; the teacher who merely transmits content loses. The short term will be rough, and we won't sugarcoat it: a real academic-integrity crisis, overloaded teachers redesigning assessment with no support (remember: two-thirds of schools have no AI policy), and a serious risk that the advantage concentrates among the students who need it least. None of this gets fixed with detectors or ostrich-style bans; it gets fixed with teacher training, redesigned assessment, and AI tutors that seek out the student instead of waiting for them.

THE HORIZON. That said, we hold on to our long-term optimism, because the data supports it. For the first time in history, there is a technology capable of delivering individualized, patient, unlimited tutoring at near-zero marginal cost. The serious RCTs show that, well designed, it works: more learning in less time. The 15% usage problem is a product and human-support problem, not a law of physics, and the whole industry is already working on it. Our forecast: within the next decade, the combination of a proactive AI tutor plus an orchestrating teacher will become the standard in systems that take it seriously — and the effect will be largest precisely where resources are scarcest today, as the Nigeria trial suggests despite its methodological loose ends. School will not disappear; it will stop being the place where information is transmitted and become the place where learning is verified, motivated, and where students learn to think with (and without) the machine. That world, painful transition included, is better than the one we had.