Picture a stack of forty written reflections. They’re well structured. The paragraphs flow. Each one cites the assigned reading, closes with a tidy insight about what the student learned, and reads like it was produced by someone who genuinely thought about the material. And every single one of them could have been generated in under a minute by a tool the student has open in another tab.
That’s the moment a lot of instructors hit recently, and it’s the moment described in a post that’s been making rounds on Hacker News: “How I changed teaching after AI managed to do all my homework assignments.” The author’s response was blunt. Written reflections, which had been part of essentially every assignment, got replaced. In their place: a scheduled 15-minute in-person session with a TA after each assignment.
I build bots for a living. I spend my days making systems that produce plausible text on demand. So I read that post less as an education story and more as a systems story, and it maps onto something I keep running into in my own work.
The artifact was never the point
Here’s the structural problem. A written reflection is a proxy. Nobody ever wanted the paragraphs. They wanted evidence that thinking happened, and paragraphs were the cheapest available signal for that thinking. For decades the proxy held up, because producing decent prose about a topic required having engaged with the topic.
Generative models broke the correlation. Not the prose quality, the correlation. The signal and the thing it measured came apart, and every assessment built on that link stopped measuring what it claimed to measure.
If you’ve ever shipped a metric-driven system, this should feel familiar. You pick a measurable stand-in for the thing you actually care about. It works. Then something changes that makes the stand-in cheap to produce directly, and your dashboard goes green while the underlying reality gets worse. Teachers aren’t facing a cheating problem. They’re facing a broken instrument.
Why the 15-minute conversation works
The fix in that post isn’t nostalgic, even if it looks backward at first glance. Oral exams, in-person interactions, projects, discussions — these aren’t retro for retro’s sake. They’re expensive to fake. A student sitting across from a TA answering follow-up questions in real time is being measured on the actual target, not on a proxy.
The author is honest about the cost, and names it as abandoning evidence-based pedagogy practices in favor of oral and written exams plus rebuilt assignments. That’s a real tradeoff, and the fact that it was made anyway tells you how badly the old instruments were failing.
The deeper reframe is one I find genuinely useful: homework becomes a learning process rather than a product to be submitted. Once the artifact stops being the deliverable, the question changes from “did you produce this” to “can you operate on this.” Those are different skills, and only one of them survives contact with a language model.
What the research space is pointing at
Work published in Frontiers on homework in the AI era suggests some concrete directions, and two stand out to me as a builder:
- Error analysis tasks — hand students AI-generated solutions and have them find and correct the mistakes.
- Self-generated problems — have students design their own question modeled on class material, then solve it.
Both of these are what I’d call review work rather than production work, and that’s exactly the shift I’ve watched happen on my own side of the fence. I don’t write most boilerplate anymore. I read it, judge it, catch the part where the model confidently used an API that doesn’t exist, and decide what stays. My value moved from typing to evaluating.
An error-analysis assignment trains precisely that. It’s not a workaround for cheating, it’s job training for the way technical work already operates. The second one, designing your own problem, is arguably harder to fake than anything on a traditional problem set, because it requires knowing what a good question looks like in the first place.
The uncomfortable part
None of this scales cheaply. Fifteen minutes per student per assignment is real labor, and a lot of institutions will look at that number and decide detection software is easier. I think that’s the wrong call. Detection is an arms race against systems specifically optimized to produce output indistinguishable from human writing, and I know which side of that race has better funding.
The instructors rebuilding their assessments from scratch are doing the harder and more durable thing. They’re rebuilding the instrument instead of patching the old one. If you build systems for a living, you already know how that story usually ends for the patchers.
🕒 Published: