The Homework Industrial Complex

The research on homework is one of the best-kept secrets in education. Harris Cooper of Duke University has conducted the most extensive meta-analyses of homework research, and his conclusions, reported first in 1989 and updated several times since, are consistently awkward: for elementary school students, homework has essentially no effect on academic achievement. For middle school students, modest effects. For high school students, a positive correlation exists, but the optimal amount of homework appears to be somewhere between one and two hours per night, with negative returns above that threshold.

The research has been available since 1989. Homework has gotten longer, more complex, and more grade-dependent in the decades since. The policy didn’t respond to the evidence because the evidence conflicted with every constituency’s interests simultaneously: administrators who use homework as a quality signal, teachers who use it as curriculum coverage, and parents who use it as evidence that their school is rigorous. Homework persisted through pure institutional inertia.

AI has now changed the evidence problem in a way that institutional inertia can’t survive. When a student submits AI-completed homework, the teacher has no information about what the student knows. The assignment was always a proxy for learning; the proxy now mostly measures whether the student thought the assignment worth doing themselves. Schools are starting to confront what the research has suggested for 35 years: most homework doesn’t work.

Let’s be precise about what “doesn’t work” means. It doesn’t mean practice is useless. Practice is essential, and distributed practice over time is one of the most robustly supported interventions in educational psychology. It means that assigning practice to be done outside the classroom, unsupervised, with no feedback during the process and delayed feedback afterward, is a poor way to structure practice. Students get stuck and stay stuck. They practice errors. They do the easy parts and skip the hard parts. They run out of time and rush. They complete it mechanically without engaging with the content. The form of homework as typically practiced optimizes for completion over learning.

The research on why homework shows weak effects is itself well-developed. Robert Marzano’s analysis of the evidence found that homework is most effective when it involves practicing a skill students have already learned (not initially acquiring a new one), when feedback is immediate, when the difficulty level is appropriate to the student’s current level, and when the time required is short enough to be completed without excessive fatigue. None of these conditions are met by most homework assignments, which ask students to attempt new applications of partially understood material, provide feedback days later, use the same assignment for all students regardless of their current level, and routinely take 30-45 minutes on a single subject.

The interesting thing about AI and homework is that AI, well-configured, could produce exactly the kind of practice environment that research says works. Immediate feedback. Appropriate difficulty calibration. Short, focused practice sessions. Identification of specific errors rather than just “wrong” marks. The AI homework assistant, in this framing, is not a cheating tool. It’s an implementation of what homework was supposed to be but never actually was.

This isn’t the way most students are using AI on homework, of course. Most students are using it to get answers, not to practice with feedback. But the gap between how AI is being used and how it could be used is itself revealing about the homework design. If your assignment is designed so that getting the answer is separable from doing the learning, then the assignment design is the problem. An assignment where the student has to defend their reasoning, where the process of arriving at the answer is what gets assessed, where the product is genuinely difficult to generate without the underlying understanding, this kind of assignment is difficult to fully delegate to AI even if AI is available, because defending reasoning requires the student to actually engage with it.

The homework industrial complex, the phrase is not original to me but deserves wider use, refers to the ecosystem of homework-adjacent products, services, and practices that have grown up around the assumption that large volumes of homework are a good idea. It includes test prep companies that provide additional homework, homework-tracking apps, tutoring services that help students with homework, homework-extension policies, and the parental labor market that has emerged around monitoring and supporting homework completion. The tutoring industry alone, at an estimated $115 billion globally in 2023, is substantially homework-completion assistance.

AI has begun cannibalizing this market in ways that are uncomfortable for the companies that built it. Chegg, the homework help company, saw its stock drop 48% in May 2023 after CEO Dan Rosensweig acknowledged that ChatGPT was taking students who previously would have paid Chegg. Course Hero, Brainly, and similar platforms reported similar patterns. The irony is that these platforms were already being criticized for enabling homework completion without learning. They were in the business of providing answers, not teaching, and students were using them exactly as they were designed, regardless of what the companies claimed their educational mission was.

AI does what these services did, but better and cheaper. The loss is real for those companies. It is not a loss for education, which was never well-served by an industry that monetized homework dependency.

Here’s the argument I want to make that nobody in education policy is making clearly: AI has created a forcing function that will, if schools respond correctly, produce dramatically better homework practice. The forcing function works like this. If AI can complete any homework assignment that merely requires retrieval and application of known facts, then assigning that kind of homework is pointless, not because students will cheat, but because the assignment doesn’t measure what you care about. Schools that want homework to be meaningful will have to design assignments that require genuine cognitive work: synthesis, original analysis, explanation of reasoning, real-world application that requires judgment. These are harder to design and harder to grade. They’re also better.

The barrier to this transition is not intellectual — the principles of good assignment design are well understood. It’s economic. Designing genuinely demanding assignments requires more teacher preparation time. Grading the products requires more teacher expertise. Neither is possible without investment in teacher time, which requires either smaller class sizes or reduced non-instructional demands on teachers. Both cost money that most school systems are not currently spending.

There’s a second forcing function that’s less obvious. AI does homework. AI also tutors. If a student is stuck on an assignment and can use AI to get unstuck, not to get the answer but to understand the concept well enough to produce the answer, then the homework environment can approximate the ideal practice environment in ways it previously couldn’t. A student at 9pm with a confusing algebra problem previously had three options: call a friend who might or might not understand it, give up, or muddle through to a wrong answer and submit it. Now there’s a fourth option: ask an AI for an explanation, work through the concept, and then complete the problem with genuine understanding.

This fourth option is the one nobody talks about when they discuss AI and homework, because it doesn’t fit into either the “AI enables cheating” or the “AI disrupts education” narrative. It’s just better homework support, available to students who otherwise wouldn’t have it. Whether students use AI for this purpose or the answer-retrieval purpose depends heavily on how assignments are designed and how students have been taught to think about learning. But the option exists in a way it didn’t before, and it’s genuinely valuable for the students who use it well.

The homework industrial complex will contract. Whether it contracts in a direction that produces better learning or just fewer assignments with the same mediocre outcomes depends on whether school systems use the pressure as a prompt to redesign practice or just as a justification for doing less.

Finland provides the most useful counterfactual here. Finnish schools assign among the least homework of any OECD country, approximately 30 minutes per night across secondary school, versus 90-100 minutes in the United States. Finnish students consistently outperform American students on international assessments, including PISA. The Finnish education research community’s explanation for this is consistent: Finnish classroom instruction is better calibrated, so students arrive at class prepared and engaged, and the in-class time produces genuine learning that doesn’t need to be extended into homework. The homework that is assigned reinforces class work rather than attempting new material.

This is not a technology story. It’s a story about how the same amount of learning time can be organized more or less efficiently. What’s relevant to the AI conversation is that the evidence has been available for decades that the American model of heavy homework is not necessary for high educational outcomes, and American schools have not changed their homework practices because the constituencies that benefit from them, college prep tutoring companies, parents who associate homework with rigor, teachers who have built their curricula around homework completion, are politically organized in a way that the evidence base is not.

AI is the first force powerful enough to destabilize those constituencies. When Chegg’s revenue drops 40% and when district after district finds that homework submission rates have decoupled from learning outcomes, the political case for the homework-as-rigor model becomes very hard to make. This is an opportunity, if schools take it. The pressure AI creates toward abandoning ineffective homework could be used to build more effective in-class practice, more rigorous assessments, and more intentional use of out-of-class time for projects that genuinely require student initiative rather than completion of prescribed tasks.

The alternative, replacing homework with AI-generated homework, maintaining the form while further hollowing out the substance, is more likely if schools respond reactively rather than strategically. The choice between these outcomes is genuinely open. It will be made by individual principals, district curriculum directors, and state education boards over the next five years, mostly without anyone explicitly framing it as a choice. That’s how most important educational decisions get made: not as deliberate policy choices but as the accumulated drift of a thousand institutional responses to immediate pressures.