The simplest explanation for why the internet didn’t transform overall learning outcomes is the one Justin Reich offers in Failure to Disrupt: even with ubiquitous knowledge, learning still requires motivation and effort. Today, many “AI will improve education” claims are just internet-era claims, reformulated.

Here’s what’s different this time: AI makes learning optional for productivity.

The motivated individual learner is the one success story of the MOOC era, and AI is breaking it. In a previous essay, I wrote about how AI is making it harder for schools to assess learning. This one is about how AI is hurting self-directed learners too.

In the pre-AI days, if you didn’t learn the tool, you couldn’t build the thing. Wanting to build the thing was the motivation that made you learn the tool. But today, AI can build the thing for you. The learning part is now optional. AI cuts the motivational link between learning and productivity in a way that the internet never did.

Paul Graham makes this observation about writing in his essay Writes and Write-Nots. He predicts the world will split into those who choose to write and those who never do, because AI makes writing skills optional. But it applies to more than just writing. It applies to any self-driven creative pursuit. You used to hit a wall trying to make something, and you’d have to learn your way past it. Now, AI can unblock you, so you have to decide where to pause and actually learn.

No previous generation of self-directed learners has had this much choice. Before, if you wanted to build software, you had to study math and programming, and you probably didn’t have time to also be a social scientist. If you wanted to build new models of human behavior, you had to read about the previous attempts, and learn statistics and experimental design, and you probably didn’t have time to build software. The prerequisites are now mere suggestions.

You can build the app this afternoon, then make the behavioral model, and be done before dinnertime. The output is free now. But learning still takes time and effort. AI accelerated productivity but didn’t accelerate learning. So you have to decide which skills are still worth learning, and which you can fast-forward past.

The research on the productivity-learning split is clear: unconstrained AI use produces students who score better on the AI-assisted assignment (“accelerated productivity”), but then score worse on an AI-free exam (“didn’t accelerate learning”). A 2024 Wharton study by Bastani et al. reached this conclusion, and so did a 2026 Stockholm University study by Strömberg et al., which produced this chart.

These studies are from schools, but the effect applies beyond the classroom. If you skip the effort, you skip the learning.

But why even bother learning something the AI knows? Because you need to understand a concept to be able to think about it yourself, and have the next idea.

It’s fashionable among the AI-pilled to say that you just need the higher-level concepts now that AI can do everything. If they traveled a couple of generations back in time, these same people would be saying that you don’t need to learn multiplication because a calculator can do it. Of course, the calculator is faster. The point is to train your brain. You can outsource the multiplication after you understand it. To understand the concepts that sit above multiplication, like area in geometry, the multiplication needs to be automatic.

Learn the things you want to be able to have ideas about. And then feel free to collect your AI productivity dividend on the rest of it. The people actually building the models at OpenAI and Anthropic? I bet they know how to multiply numbers without a calculator.

We had just walked everyone through an activity to come up with ideas for businesses they would like to start. Then I told them to add more of their own ideas to their online journal. The journal was part of the custom software we built for the summer camp, and it even included an “AI with guardrails” designed to help students who were stuck, but not do the work for them. As they typed in their journals, I walked around the room. Coming around to one seventh-grader’s desk, I saw he had switched tabs from our custom-AI-with-guardrails to plain ChatGPT and had told it to write the whole business plan for one of his ideas. Then, realizing the output was too fluent, he added a second prompt: “No this sounds too smart. You have to make it sound like a kid wrote it.”

Another day, I was working with one of the younger kids. This kid wanted a cool logo for her product, but her drawing skills didn’t live up to her imagination. She had even rejected a few AI attempts. She just wanted HER logo, but drawn better. She gave the AI a hand-drawn version and it gave her back an image with no coloring outside the lines. Her creative vision, executed with steadier hands. She was beaming.

Both kids used AI, but with different of intent. One of them worked hard to manage my impression of the output, the other worked hard to meet her own standards. This wasn’t a difference between the kids, it was a difference in how meaningful the work was to them.

Lots of folks, including me in an earlier piece, have been skeptical of AI use while learning, because it makes cognitive offloading attractive. That’s true. But here is the other side of the coin: seeing work just beyond your capabilities can be inspiring, too. I know, because I’ve felt it. I remember copy-pasting code I didn’t understand in some of my hobby projects – that was the 2010s-equivalent of vibe coding with AI. It allowed me to get the project working. Of course, I had to go back later on to learn how that worked, often when I hit the next wall. But feeling like you’re making progress is important too. AI used with ownership can be the difference between excitedly finishing a project, and giving up and deciding that “I’m not a _____ person.”

I’m grateful for what I learned from these kids over the summer. Kids really do care about the work when they feel ownership of it.

I’ve talked to teachers and school leaders from elementary school to graduate school. I see three main ways schools are handling this wave of AI use:

1. Punish it: Detect AI use on homework; treat it as cheating.

2. Prevent it: Move to live assessments – oral or offline – that are hard to cheat on.

3. Embrace it: Assess the synthesis, not the output.

While all three are common, the first one is a dead-end.

The first approach: Punish AI use

This is widespread today – AI detection tools are everywhere. But teachers should avoid this. Even with good intentions, it usually sparks an adversarial game where students then also use detectors and humanizers to hide their AI use. And so far, students have adapted faster than the tools trying to catch them. The focus shifts from learning to avoiding detection.

School network bans and tutors-with-guardrails don’t work either, at least not standalone. They can’t protect take-home assignments when uncensored ChatGPT is one tab away.

We should just stop grading take-home work that AI can do. If you are testing foundational facts, Prevent is more effective and less adversarial. If it’s about rewarding creativity and sustained effort, Embrace is better.

The University of Chicago Law School just announced a thoughtful AI policy that even non-law schools should look at. The point is to protect critical thinking, not to ban AI. Punishment and detection are nowhere to be found in that document.

The second approach: Prevent AI use

Some knowledge is foundational: basic literacy and numeracy, basic history, and basic science, for example. You don’t absorb this in your everyday experience – it needs to be taught. Students should learn these things, even though computers can do them perfectly. Cognitive scientist Daniel Willingham tells us why. You can only think with the knowledge actually in your head, not the knowledge sitting in a book or chatbot. So we all still need to learn foundational concepts – though what counts as foundational will keep evolving.

Some struggle is useful for learning them. Using AI to generate answers before you understand removes this useful struggle. Assessments need to preserve what Terence Tao calls natural friction to train your mind.

For foundational knowledge that requires struggle, we should make the assessment itself AI-proof. Paper exams are making a resurgence – good. So are oral exams. It’s hard for a student to fake understanding at the speed of a live conversation. The only downside is that it doesn’t scale. Maybe scaling this kind of assessment is where AI should actually be helping.

AI could shift from cheating-enabler/detector to knowledge-assessor – and move the teacher from adversary to coach.

The third approach: Embrace AI use

Once you have enough of a foundation, you can learn to work with AI. This is where the economic value lies – engineers, lawyers and doctors are all using AI today, embedded in their professional tools.

A long essay or computer science project – the kind that requires multiple sittings – is still valuable as a take-home, but only if working with AI is the point. The grading moves from the artifact to an oral defense of it. If you outsource your argument to an AI, you’re caught in the first minute of a live back-and-forth. If you came up with a first draft and went back-and-forth with the AI to strengthen it, that will survive a ten-minute conversation with your teacher. Students show their work by pointing to their own ideas, and by sharing the AI session itself.

The teacher’s job here is to encourage effective AI use. Get this right, and it lets students take on vastly bigger problems, and get a feel for work that is just beyond their abilities. Success looks like a middle schooler in a summer camp drafting their first novel, or a high schooler launching an app in their community. I’ve seen both of these stories. That kind of excitement compounds, and gives a taste of the world they are entering.

So stop punishing the cheating. Protect the learning foundation with assessments that can’t be gamed. Then on top of that foundation, work with AI to build big things in the real world.

In the spirit of Embrace, watch my collaboration with Claude as I wrote this essay: homeworkbench.org/essay (this is my real drafts and actual conversation; the webpage itself was 100% AI-generated)