Breaking down the cognitive debt of AI 

What if the essay is getting better while the mind writing it is getting weaker? What if the grade is rising, the polish is improving, the citations are cleaner and underneath all of it, something essential is quietly atrophying? And what happens to a civilization that spent centuries building education systems to strengthen human thinking, the moment it stops noticing whether thinking is still happening at all? 

That is the uncomfortable question sitting at the center of a wave of evidence emerging from UNESCO and the OECD this year — evidence that deserves to be treated not as an academic curiosity, but as one of the most consequential findings of the AI era. More than 25 education ministers and national representatives gathered in Paris during UNESCO’s Digital Learning Week and adopted a joint statement insisting that education remain a human right and a common good as AI adoption accelerates, laying out eight priorities spanning governance, accountability, and the protection of critical thinking itself (UNESCO). That a room full of national education leaders felt compelled to explicitly defend critical thinking as a priority — as though it might otherwise be quietly designed out of the system — tells you how serious the underlying data actually is. 

The numbers that should stop everyone 

The core evidence comes from a UNESCO survey of higher-education educators, and it contains two figures that deserve to sit side by side, because their relationship is the entire story. Ninety-three percent of surveyed academics believe students are using generative AI to complete assignments (UNESCO). That number, on its own, is almost banal at this point — of course students are using AI, the tools are ubiquitous, free, and fast. Nobody serious expected otherwise. 

The second number is the one that should stop every university administrator in their tracks: 63% of the same educators perceive a deterioration in students’ cognitive capabilities (UNESCO). Not a change. Not a shift in study habits. A deterioration — a decline in the underlying mental capacity to reason, synthesize, and argue. 

And here’s what makes the finding even more striking: this isn’t a community divided between alarmists and enthusiasts. Fewer than 1% of respondents supported banning AI outright. Only 2% advocated unrestricted, unmanaged use (UNESCO). The overwhelming majority of educators occupy the uneasy middle ground — genuinely unsure how to reconcile a tool they know is useful with a decline they’re watching happen in real time. That is not the profile of a moral panic. That is the profile of a slow-motion crisis being observed by the people closest to it, who don’t yet have the institutional tools to respond. 

The OECD’s parallel evidence, drawn from PISA data, adds a crucial second dimension: higher-performing students tend to report less AI use, not more, and the specific capabilities most valuable in an AI-saturated environment — distinguishing fact from opinion, interpreting ambiguity, critically evaluating information — are showing measurable weakness. Put those two data sets together and a troubling pattern emerges: the students most capable of thriving without AI assistance are precisely the ones using it least, while the skills AI can’t replace are exactly the skills showing the most decline. 

The metric universities have been optimizing is wrong 

Here is the structural error underneath all of this, and it’s worth stating as plainly as possible: better output does not equal better cognition. 

For the entire history of formal education, these two things were functionally inseparable. A superior essay was reliable evidence of superior synthesis, recall, argumentation, and judgment, because there was no other way to produce it. The artifact was a proxy for the thinking, and the proxy was trustworthy precisely because faking it required doing the actual cognitive work anyway. 

Generative AI severs that link completely. A student can now submit a more polished, better-structured, more sophisticated piece of writing than they were capable of producing independently a year ago — while exercising less synthesis, less recall, less argumentation, and less judgment than a student who struggled through a mediocre essay entirely on their own. The artifact has become decoupled from the process that used to guarantee its meaning. 

This is precisely why UNESCO’s ministerial statement dedicates one of its eight core priorities specifically to protecting critical thinking — calling for AI use that actively prompts reasoning rather than substituting for it, and for building AI competencies among both teachers and students so the tool augments cognition instead of quietly replacing it. That’s a remarkably specific policy response, and it exists because the alternative — universities continuing to grade the artifact rather than the thinking process behind it — creates a genuine institutional trap: measured academic performance can rise indefinitely, even as the underlying capability the grade was supposed to certify quietly collapses. 

Borrowing against a balance we can’t see 

The phrase “cognitive debt” is worth taking literally rather than metaphorically. Financial debt lets you consume today what you haven’t yet earned, on the promise that you’ll pay it back later, with interest. Cognitive debt works the same way: a student can produce today the intellectual output they haven’t yet developed the capacity to produce independently, borrowing that capability from a machine — and the interest accrues invisibly, compounding as a widening gap between apparent performance and actual capability. 

The danger of any debt is that it feels like abundance right up until the moment it doesn’t. A student who has outsourced years of synthesis and argumentation to AI tools will look, on paper — in grades, in transcripts, in polished portfolios — indistinguishable from a student who developed those capacities independently. The debt becomes visible only under conditions the artifact was never designed to test: an unscripted interview, a live problem no template can solve, a professional situation where the AI tool isn’t available or doesn’t have the answer. That is the moment cognitive debt comes due, and by then it’s a workforce entering the economy with a capability gap nobody detected in time to address. 

This is not a hypothetical concern reserved for future graduating classes. It’s already showing up in the OECD’s PISA-linked findings: the specific competencies that matter most precisely because AI exists — separating fact from opinion, sitting with ambiguity instead of demanding a clean answer, evaluating information critically rather than accepting it — are the competencies showing the clearest signs of erosion. These aren’t arbitrary skills. They are, almost by definition, the exact human capabilities that remain valuable because they are the ones AI is least reliable at replicating independently. Watching them weaken at the same moment they become most economically and civically important is close to the worst possible version of this story. 

What a real response actually requires 

To their credit, the education ministers who convened in Paris didn’t respond with prohibition, and the data explains why that would have been the wrong instinct — fewer than 1% of educators wanted an outright ban, and for good reason: banning a tool that 93% of students are already using, quietly, unsupervised, and ungoverned, doesn’t eliminate cognitive debt. It just removes the possibility of managing it. The UNESCO statement’s eight priorities — spanning governance, auditability of AI systems, teacher agency, learner data rights, age-appropriate safeguards, equity by design, full cost accounting before adoption, and pooled governance capacity across nations — read less like a technology policy and more like an attempt to rebuild educational assessment for an environment where the artifact can no longer be trusted as a proxy for the mind that produced it. 

That rebuild is the actual work ahead, and it’s far harder than writing a policy statement. If the essay can no longer certify the thinking, institutions need new ways to observe thinking directly — oral defenses, in-person problem-solving, iterative drafts that show reasoning develop over time, assessments designed around ambiguity and synthesis rather than polished final output. Two-thirds of higher education institutions already have, or are developing, some form of AI guidance but guidance about permitted use is a different, much easier problem than redesigning what gets measured and how. The former manages behavior. The latter protects the actual purpose of education. 

The question that should keep every educator up at night 

So here is where this leaves us, and where the hardest questions have to live. If 63% of the educators closest to students are already watching cognitive capability erode in real time, how much of that erosion is happening silently, undetected, inside institutions that haven’t yet built the tools to even measure it, let alone the other 37% who may simply not be looking closely enough?  

If the students who most need to build independent reasoning capacity are, according to the OECD’s own data, the ones leaning hardest on AI rather than developing it themselves, are we quietly building a generation stratified not by access to intelligence — since AI access is nearly universal — but by who resisted the temptation to borrow it? 

And the hardest question of all: when this cognitive debt eventually comes due — in a workplace, a crisis, a decision with no machine available to make it for you — will we discover the deficit early enough to correct it, or will we only find out how much was borrowed the moment someone desperately needed to repay it, and couldn’t?