Does using AI to learn actually make you learn less?

AI Study: Homework Scores Up 18%, Exam Results Down 20%

Often, yes: a new study by Stromberg, Lei and Wu tracking 27,000 students over 30 months found generative AI raised homework scores 18 percent while closed-book exam scores fell 20 percent within six months. Paul Bloom's Psych explains why — memory is built through effortful retrieval, so let AI check your work after you attempt it, never instead of you.

I have been waiting for someone to measure what AI does to the part of learning you cannot see, and this week a research team finally did it at scale. David Stromberg of Stockholm University, with Victor Lei and Yanhui Wu of the University of Hong Kong, tracked roughly 27,000 students aged 12 to 18 across nine subjects for 30 months, combining homework data, monthly closed-book exams and high-stakes entrance exams. Al Jazeera covered the working paper on September 2, 2026, and the headline numbers deserve to be read twice: after adopting generative AI, homework scores rose 18 percent and time per assignment fell about 30 percent, from 64 minutes to 45. Then the same students sat monthly exams without AI, and their scores fell 20 percent within six months. Entrance-exam performance declined too, with the full effect emerging more slowly.

Output went up. Learning went down. Both at once, in the same people, from the same tool.

The OECD's Digital Education Outlook 2026, published in January, reached the same conclusion from the broader research base: where generative AI is neither built to teach nor used under any structure, outsourcing tasks to the model improves performance with no real learning gains. Researchers working with Dragan Gasevic have a name for the erosion of productive cognitive effort behind this: metacognitive laziness. And an Ipsos survey for the Vodafone Foundation of 7,000 European students shows how the tool actually gets used unsupervised — 56 percent for getting information, 31 percent for complete solutions to tasks, and only 20 percent for anything resembling a personalised learning plan.

If you run a company, this is not a story about teenagers. Professionals have closed-book exams too — the board Q&A, the negotiation, the incident call at 2 a.m., the sales conversation where you cannot pause to consult a chatbot. Every one of those moments tests what actually got encoded in your head, not what your AI produced on your behalf. Paul Bloom's Psych is the best short explanation of why the gap opens: memory is not a recording, it is a reconstruction built through effortful encoding and retrieval, and we routinely confabulate — we mistake the fluency of watching polished output appear for understanding we never built. The homework felt learned. The exam proved it wasn't.

Here is the protocol I now run for anything I genuinely need to know cold, as opposed to work I am happy to delegate forever. I call it the closed-book protocol:

Attempt cold first. Fifteen minutes on the problem with no AI — a draft, a decision memo, a diagnosis. The struggle is the encoding; done means you produced something wrong-but-yours.
Use AI as examiner, not author. Paste your attempt and ask it to find errors, gaps and counterarguments. You are done when it critiques your work instead of replacing it.
Retrieve after 24 hours. Next day, write what you remember before opening the chat log. Done when the recall attempt exists on paper.
Ship one artifact without the transcript. Explain the topic to a colleague or make the call with the chat closed. If you cannot, loop back to step one.

Now the limits, because they matter. The study population is students aged 12 to 18 in China, not executives; effect sizes for experienced professionals with strong priors may differ, and the OECD is careful to note that AI built to teach — structured tutoring that withholds answers — is a different animal from AI that completes your work. This paper measures unstructured adoption, which is also how most professionals I coach use it. And I will admit the other side plainly: I still let AI write plenty of things I never intend to remember. That is the actual skill now — deciding which knowledge must live in your head for the closed-book moments, and refusing to outsource that part, while cheerfully delegating the rest.

Sources: Al Jazeera — Faster homework, poor exam results (Sep 2, 2026); Stromberg, Lei & Wu — SSRN working paper; OECD Digital Education Outlook 2026.