The study that exposes academia's 'false positive': AI raises grades 18% and sinks exam scores 20%

By Zendoric · September 6, 2026.

There is one figure that should unsettle any university head: the higher the grade on an assignment, the worse the subsequent exam performance. This is not an isolated classroom paradox. It is the result of tracking 26,811 Chinese secondary school students for 30 months, in work by David Strömberg (Stockholm University) with Victor Lei and Yanhui Wu (Hong Kong). The use of artificial intelligence (AI) raised assignment grades by 18% and cut the time spent on them by 30%. Six months later, results on closed-book exams —no notes, no tools— had fallen by 20%.

The grade has become a false positive: it certifies command of an assistant, not understanding. And the problem does not stay in the classroom, because universities and education systems keep issuing diplomas and credentials based on that contaminated metric.

There is an underlying explanation, not just a correlation. Robert and Elizabeth Bjork have spent decades describing "desirable difficulties": the effort of retrieving a fact from memory, getting it wrong and trying again is precisely what fixes learning. AI removes that friction, which is exactly the necessary condition for learning. Hamsa Bastani and her team confirmed it in PNAS with nearly a thousand Turkish students: with the AI interface open, subsequent exam scores fell 17% against those who never used it. But with a version that only offered the teacher's hints —the Socratic method, asking instead of answering— practice improved by 127% and the negative impact vanished. Zara Contractor and Germán Reyes (Middlebury) reached a similar conclusion with university students: the advantage held a week later when AI was used to have it explain concepts, and was lost when it was used to generate the final text. The relevant question is not how much AI a student uses, but what they do with it.

Banning it is not a real option. 92% of Latin American university students already use some AI tool regularly, according to the 2026 survey by the Digital Education Council and Tec de Monterrey. In Spain, 59% of young people use it to study, the highest figure in the European Union according to Eurostat. Detecting its use is, at this point, a battle lost before it starts.

What is feasible is redesigning assessment and the instructions that govern the tool. The Chinese study yields an operational rule: time spent on the assignment should not fall; if AI saves half an hour, that half hour must be used to go further, not to finish sooner. Bastani yields another: the difference between asking and answering does not depend on the model but on the instructions that govern it, and today that pedagogical decision is made by the tool's vendor, not by the university. And in assessment itself, the logical shift is toward the in-person: closed book, oral defense, new problems in front of the teacher. Princeton has already done it, in barely a year: its faculty voted to proctor all exams, something banned at the institution since 1893.

One piece is missing that almost no set of rules contemplates: a "debt rule". Work done with AI would be admissible if the student can reconstruct their reasoning unaided in three minutes; if they cannot, the work is not theirs. It is an expensive check —it takes time, it demands teacher training— but it is the only one that distinguishes understanding from delivered output.

That is where inequality appears, and there is no way to soften it here: verification costs money. Princeton can pay for proctoring and small groups; Mexico invests $3,650 per student against the OECD average of $13,210. PISA 2022 already showed that, in Latin America, three out of four 15-year-olds fail to reach the basic level in mathematics, against 31% in the OECD. AI does not create that gap, but it masks it with flawless assignments and could widen it tomorrow: the inequality measured today in learning will shift to the credibility of the diploma itself. In Spain, the CYD Foundation records that 46% of university students say their faculty does not promote the use of AI —six points more than a year ago— and that 61% have received no training on it at all. The vacuum of institutional judgment is being filled by students themselves, alone.

The most uncomfortable data point, however, is physiological. Nataliya Kosmyna and her team at MIT Media Lab hooked 54 Boston university students up to an electroencephalogram —a device that measures the brain's electrical activity— while they wrote essays: some with ChatGPT, some with a search engine, some with no help at all. Those who wrote unaided showed the broadest brain connectivity networks; the search engine group, intermediate ones; the AI group, the weakest. 83% of the latter were unable to quote a single sentence from their own essay. The authors called it "cognitive debt": a liability someone will have to pay later, whether the graduate facing their first job, the patient facing their doctor or the citizen facing the bridge somebody calculated.

Our reading is that this crisis is not an argument against AI in education, but against institutional laziness in integrating it. The data itself proves it: with Bastani's Socratic version not only did the negative impact disappear, but practice improved by 127%. The problem is not the tool; it is that today the vendor designs it and no one evaluates it. This fits something we had been observing in the education sector: the one who wins with AI is not the teacher who merely transmits content and grades deliverables, but the one who orchestrates it —designing the necessary friction, verifying reasoning, acting as an augmented tutor. That role becomes more valuable, not less, and it is exactly the kind of judgment- and relationship-based work that survives automation in every sector we have analyzed so far.

There is also a parallel worth noting. Just as image verification is shifting from the pixel to provenance —who captured and signed the file, not whether the image "looks" real—, learning verification will have to shift from the deliverable to the process: not what was handed in, but whether whoever handed it in can reconstruct it live. In both cases trust migrates from the final artifact to the chain that produced it. It is an expensive and still nascent institutional redesign, but it is the only answer that does not depend on winning a detection arms race that, as we have already seen with deepfakes, is lost before it starts.

In the short term, the diagnosis is harsh and must be said plainly: the poorest systems are the ones with the least room to pay for verification, so the educational gap PISA already documents could turn into a credibility gap for their diplomas. But over the longer term, the same study that exposes the problem points to the solution: well-designed AI, with deliberate friction and verification by reconstruction, can multiply real learning instead of replacing it —just as the Turkish Socratic version did. That is the abundance possible in education: not less cognitive effort, but better-directed effort, with the teacher freed to do what a model cannot do for them, which is to know when a student has truly understood and when they have merely delegated the thinking.