The Crutch and the Tutor: Cognitive Debt Is a Design Choice, Not a Property of the Machine

Jul 7, 2026

Every time a new cognitive technology shows up, the same panic arrives with it, and the panic is always shaped the same way.

In Plato’s Phaedrus, the god-king Thamus refuses the gift of writing. His objection is specific and, honestly, sharp: writing will produce “the appearance of wisdom, not true wisdom.” People will stop building knowledge inside themselves and start relying on external marks. Memory will atrophy. Students will seem learned without being learned.

He was right about the mechanism. We did offload memory to the page. He was wrong about the verdict — writing also gave us cumulative, cross-generational knowledge that dwarfed the cost. But notice the structure of his mistake, because it’s the same mistake being made right now about AI: he correctly identified a real trade-off, then assumed the trade-off settled the question. It didn’t. The value depended entirely on what got built on top of the offloading.

So when the headlines ask “Is AI making us dumber?”, they’re asking Thamus’s question. And Thamus’s question is malformed. The research that actually holds up doesn’t answer “does the technology harm cognition.” It answers something far more useful — and far more political.

The evidence everyone cites is the evidence you should trust least

The study driving most of the doom is MIT Media Lab’s “Your Brain on ChatGPT”. EEG caps, essay-writing, three groups: ChatGPT, Google, and no tools. The ChatGPT group showed the weakest neural connectivity and, when later asked to write unaided, couldn’t recall their own sentences. Dramatic. Widely shared.

It’s also a preprint that was never peer-reviewed, with 54 participants (only 18 in the crucial hand-off session), one narrow task, one narrow demographic. And the lead author told TIME she rushed it out specifically because she feared a policymaker might “decide, ‘let’s do GPT kindergarten.’” That’s advocacy, not a finding. She has since spent considerable effort walking back the coverage — “we didn’t find any brain rot,” she notes, and they never measured IQ. Science pointed out the obvious: making damning claims from preliminary data, ahead of review, is not how you responsibly prevent bad policy.

The other frequently-cited paper, Gerlich (2025), surveyed 666 people and found a negative correlation between AI use and critical thinking. Bigger sample — but it’s self-reported, cross-sectional, and can’t establish which way the arrow points. “AI erodes thinking” and “people who already offload more gravitate to AI” fit that data equally well. A correlation cannot separate them. Anyone treating it as proof of causation is smuggling in an assumption the study never earned.

This is the media-literacy lesson hiding in plain sight: the scariest study and the most-shared study are often the weakest, precisely because alarming preliminary claims travel faster than careful ones. Follow the motive on the fear as skeptically as you’d follow it on the hype.

The study that actually settles something says something more interesting

The strongest evidence we have is Bastani et al. (2025), in PNAS — a randomized controlled trial with nearly 1,000 high school students in Turkey. Peer-reviewed. Causal, not correlational. No disclosed financial stake in the outcome. And instead of one AI condition, it ran two, which is what makes it worth your time.

One group got GPT Base: a raw ChatGPT-style interface, told to act as a tutor. Another got GPT Tutor: the same GPT-4 underneath, but with guardrails — it gave hints instead of answers, and was shaped by teacher input. A third group got no AI at all.

The results, in order:

During practice, the GPT Base group performed 127% better than the control.
On the exam, with the AI taken away, the GPT Base group scored 17% worse than the control group that never had AI at all.
The GPT Tutor group’s harm disappeared — it matched the control on the exam.

Read that sequence again. The raw chatbot made students look brilliant and learn nothing. When the scaffold was removed, they were worse off than if they’d never touched it. The mechanism, per the authors: students used GPT-4 as a crutch, copying answers — including the wrong ones — without understanding. The guardrailed version, doing the same underlying task with the same underlying model, produced none of that damage.

Here is the whole point: the difference between the harmful arm and the harmless arm was not the technology. It was the same technology. The difference was the design.

”Does AI harm cognition” is the wrong question. “Designed by whom, tuned for what” is the right one.

Once you see that the harm lives in the design rather than the model, the entire framing shifts. Cognitive debt isn’t something AI does to you. It’s something a particular configuration invites — and a different configuration prevents. The variable is whether the tool hands you finished cognition or scaffolds you into producing your own.

And the moment “it’s a design choice” is on the table, the next question is unavoidable: who makes the choice, and whose interest is it tuned to?

This is where the comfortable version of the conversation ends and the honest one begins. A crutch and a tutor are not just two designs — they are two business models. A tool optimized for engagement, session length, and daily active use has a structural incentive to be the crutch, because dependence is the product. A tool that made you so capable you needed it less would be, from that vantage, a failure. The answer-vending default isn’t a bug or an oversight. It’s the design that maximizes the metric the tool was built to maximize.

That’s not a reason for despair, and it’s definitely not a reason to romanticize going back to the textbook. The Bastani tutor arm proves the good version is buildable. The hint-not-answer, keep-the-human-in-the-loop design isn’t hypothetical — it exists and it works. The question is purely one of incentives: whose cognitive interest is the tool aligned with?

And the only durable way to align a tool with the user’s interest is for the user to be able to shape the tool. Steerable systems. Open weights and open prompts you can actually inspect and modify. Models you can run yourself, configured to scaffold rather than to vend. This is the unglamorous, concrete core of “digital sovereignty”: not a slogan about freedom, but the specific technical capacity to make your tools serve your understanding instead of someone else’s engagement graph. If cognitive debt is a design choice, then controlling the design is how you refuse the debt.

The question this frame doesn’t answer

Notice what the design argument settles and what it leaves wide open. It tells you a tool can be built to scaffold instead of to vend, and that whoever builds it decides which. That’s the collective, structural half — the half about power and control. It’s the half worth fighting over.

But it doesn’t tell you what happens inside you when you sit down with even a well-built tool, because there’s a second variable underneath the first, and the studies barely touch it: not just which design you’re handed, but which layer of your own thinking you hand over.

Every intellectual task has a mechanical layer — retrieval, derivation, boilerplate, the rote first pass. Thamus’s page has been carrying that layer for two thousand years, and offloading it is mostly fine; it’s arguably the whole point of having tools. But every task also has an evaluative layer — the judgment that lets you tell whether the output you just got is any good. Whether the formula is real or invented. Whether the fluent, confident paragraph is also a wrong one. These two layers are not the same faculty, and — this is the part that matters — they do not fail together. You can offload the mechanical layer and keep the evaluative one, and come out genuinely ahead. Or you can offload the evaluative layer itself, and lose the one capacity you’d need to notice you’d been handed garbage.

That distinction is where the most seductive pro-AI argument — discovery beats acquisition, who cares if I can’t derive it as long as I can find it — turns out to be true, with a catch sharp enough to deserve its own essay. Hold onto the two layers. We’ll need them.

A note on how early this all is

One honest caveat, because the field deserves it: this research is thin and young. The RCT above is one study, one subject, one age group, one culture. The scare studies are weaker still. Nobody has the longitudinal data yet to say what a decade of this does to a developing mind, and anyone claiming otherwise — in either direction — is selling something.

But “we don’t fully know yet” is not the same as “we know nothing.” We know enough to reject the malformed question. The evidence doesn’t support AI makes you dumber. It supports something more precise and more actionable: outsourcing the thinking means you don’t learn the thing you outsourced — and whether a given tool encourages that outsourcing or prevents it is a design decision, made by someone, in someone’s interest.

Which means the most important question isn’t about your brain. It’s about who’s holding the pen that designs the tool your brain now runs on — and whether you have any say in it.

That’s the argument for controlling the design. Part two turns the same lens back on us — on which layer of our own thinking we’re actually handing over — and asks a harder question: whether the research measuring all of this is even looking at the right thing.