Education in the Age of LLMs: Stop Grading the Robot's Homework

Somewhere right now, a student is submitting an essay they did not write, to a teacher who does not have time to read it properly, graded against a rubric a machine could apply faster, for marks that are supposed to certify a skill the student may not have.

Even though everyone in that chain more or less knows, the chain keeps running anyway.

LLMs are already in every student's pocket, and since that question has already been settled, I do not think "how do we stop students from using AI?" is a useful thing to ask anymore. The better question is what education is for when the answers can write themselves.

The detector arms race

The first reaction was panic, which is fair: ban ChatGPT, buy an AI detector, and make students write in airplane mode under supervision.

It has not gone well. Detectors have been shown to flag non-native English writers as AI while missing text that was lightly edited, and students figured out how to "humanize" output faster than schools figured out the tools. The only skill that arms race really builds is evasion.

If this feels familiar, it should, because it is the practical file all over again: sixty students, identical observations, and full marks for neat handwriting. We were grading fiction long before AI showed up, and AI simply made the fiction cheaper.

Here's the thing: if a take-home assignment can be completed entirely by a machine, it was never measuring much in the first place. The LLM exposed an evaluation that was already broken.

Ask them about it

For a long time we have relied on a lazy shortcut, which assumes that if the essay exists, learning must have happened. That shortcut does not work anymore.

What replaces it is much older than homework. You check whether the person in front of you understands, and the tool for that is a conversation. You wrote this? Great, walk me through the second paragraph. Why this argument and not the obvious counter? Here is your code, and the requirement has changed slightly, so go. A student who understands it will adapt, while a student who pasted it will freeze, and five minutes of that tells you more than five pages of polished prose.

Process counts too. The chat with the AI, the questions a student asked, the answers they threw away and the bits they corrected are the new rough work, and I would much rather grade those margins than the final page.

"But we have sixty students per class!" That is fair, although it is funny how there was budget for detection software. The same models causing the panic can also help with the scale: an AI can run a back-and-forth with every student, probe their reasoning, and flag where the understanding goes thin, so that the teacher can review the flags and have the conversations that actually need a human.

Crutch or exoskeleton

This is the part that makes both camps a bit uncomfortable.

Banning AI prepares students for a world that has already gone, and letting AI do everything does not prepare them at all. The distinction I keep coming back to is between a crutch and an exoskeleton.

A crutch replaces a function, and if you lean on it when you do not need to, the muscle weakens. A student who hands their thinking to an LLM ends up fluent and hollow, able to produce anything and explain none of it. An exoskeleton amplifies a function: the muscle still works, it just lifts more. A student who drafts their own argument and then uses the model to attack it, find counterexamples and try five other framings is doing more thinking, not less.

The same tool gives opposite outcomes depending on who is doing the cognitive work.

We have been here with calculators. Eventually we settled on learning long division first and getting the calculator after, and maths class moved on to harder problems. An LLM, however, is a calculator for thinking itself, so what you need before leaning on it is reasoning rather than arithmetic: building an argument, sensing a wrong answer, and noticing when the model confidently cites a book that does not exist.

That is skepticism, which school was already bad at teaching back when misinformation came as WhatsApp forwards with spelling mistakes. Now it arrives polished, grammatical and personalized. The bar has gone up, and most curricula have not noticed.

The bottom rung

This is the part nobody wants to say at a parent-teacher meeting.

A lot of entry-level knowledge work, such as formatting the deck, writing the boilerplate and summarizing the document, was the bottom rung of the career ladder, and that rung is now a feature in a monthly subscription. I am not saying this from a distance. Since I spend my days building with these tools, I can see exactly which tasks they have already taken over.

So the split I would worry about is between people who can direct AI and people who can only do what AI does. If you can break a messy problem into pieces, hand off the mechanical parts, check what comes back and own the result, you start your career a few years ahead of where a fresher used to. If you can only produce passable text or standard code on instruction, you are competing with a machine on price.

Twelve years of memorize-and-reproduce trains exactly the second kind. Sharma ji ka beta scored 98% by reproducing the textbook perfectly, but the textbook is an API call now, so what exactly is the 98% in?

I wrote a while back about how we measure students by marks instead of understanding, and how that habit follows people into performance reviews. The stakes are higher now. Gaming the system used to get you a survivable career, and now it can certify you for work that is already automated.

What I'd want in the classroom

Honestly, nothing here needs new technology. It needs us to stop competing with the machine as producers of output and to start training people to direct it well:

• Breaking vague goals into clear, delegable pieces, which is the core skill of working with AI and also just the core skill of thinking.
• Verification as a real subject: checking claims, testing code, and asking "what would make this wrong?"
• Taste. When ten options cost nothing to generate, all the value moves to picking the right one.
• Some deliberately AI-free time. We kept mental maths after calculators because the muscle matters, and you do not take a forklift to the gym.

And evaluation that looks more like a conversation than a submission, where the question is whether you can explain it, defend it and extend it.

Full disclosure: an LLM could have written a version of this post, probably faster and with fewer rants. But if you sat me down and asked me to defend any paragraph here, I could, and for the next generation, that is pretty much the whole exam.

Can you defend what you submitted?

P.S. If you're a student and you used AI on tonight's assignment, that is fine. Just make sure you could explain it to your teacher without it.