When Homework Became a Conversation Instead of a Paper

What if the homework problem in schools right now isn’t that students are cheating, but that homework itself was never the point?

That question has been rattling around in my head since I read an educator’s account of what happened when AI handled every assignment he gave out. Not some of them. All of them. His response wasn’t to build a better detector or write a sterner syllabus. He rebuilt the assignments from the ground up, and in doing so, he stumbled onto something that anyone trying to understand AI agents should pay attention to.

The assignment that stopped proving anything

Here’s the setup. Written reflections used to be part of pretty much every assignment he handed out. A student reads, thinks, writes down what they thought. Reasonable. Evidence-backed. The kind of task that shows up in teaching guides for good reason.

Then a language model could produce a perfectly acceptable written reflection in about four seconds. The artifact still looked right. It just no longer carried any information about the person who submitted it.

So he swapped it out. Now, after every assignment, each student schedules a 15-minute in-person session with a teaching assistant. Same learning goal. Completely different format. You cannot outsource sitting in a chair and explaining your own thinking out loud to a human being who can ask a follow-up question.

He’s been candid that this cost him something. Moving to oral and written exams meant walking away from practices that research supports, and redoing most of his assignments from scratch. That’s not a triumphant story. That’s a teacher making a trade under pressure.

Why this matters beyond the classroom

I write about AI agents for people who don’t build software, and this shift maps onto something I keep trying to explain.

An AI agent is very good at producing the thing that looks like the answer. It is not good at being the person who understood the question. For years, a lot of our institutions used the first as a proxy for the second, because there was no cheaper way to check. An essay stood in for comprehension. A completed worksheet stood in for practice. A report stood in for analysis.

Those proxies worked because producing the artifact was hard. Once producing the artifact becomes trivial, the proxy collapses, and we’re left staring at what we actually wanted all along.

Teachers are just the first group to hit this wall in public. Hiring managers reading AI-polished cover letters are hitting it. Managers reading AI-summarized status updates are hitting it. Anyone whose job involves evaluating other people’s written output is going to hit it.

Homework as a process, not a product

The reframing that came out of all this is the part I find genuinely useful. Homework stops being a product you hand in and becomes a process you go through. The submitted file is no longer the evidence. The conversation is.

Teachers making this shift are leaning on a few moves:

• In-person interactions and oral exams, where a student has to hold their own reasoning in real time
• Projects and discussions, which surface engagement in ways a finished document can’t
• Error analysis tasks, where students are handed AI-generated solutions and asked to find and fix what’s wrong
• Self-generated problems, where students design a question modeled on class material and then solve it

That third one is my favorite, because it flips the tool into the curriculum. Critiquing a machine’s answer requires more understanding than producing an answer from scratch. You have to know what correct looks like before you can spot where the model wandered off. An AI agent can write you a solution. It cannot reliably tell you which parts of its own solution are nonsense, which is exactly the skill everyone using these tools needs and almost nobody is being taught.

The cost nobody should skip over

I want to be honest about the tradeoff, because the optimistic version of this story does everyone a disservice.

Fifteen minutes per student per assignment does not scale for free. It requires teaching assistants, scheduling, rooms, and hours. Oral exams are harder to grade consistently and harder on students who freeze up in conversation. Redoing an entire course’s assignments is unpaid, invisible labor that lands on individual instructors who didn’t ask for any of this.

So when you hear that AI is pushing education toward more human contact, hold both halves of that. The direction may be good. The bill is real, and right now it’s being paid by teachers.

For the rest of us, there’s a question worth borrowing. Look at the documents you produce at work and ask which ones exist to demonstrate that you understood something. Those are the ones an agent can now fake. Figuring out what should replace them is the same problem this teacher solved with a chair, a TA, and 15 minutes.

🕒 Published: September 26, 2026