What Detection Couldn’t Teach

I. The question of the week

Somewhere between the winter of 2025 and this autumn, a large number of American schools stopped trying to catch their students. Not out of permissiveness, and not because the cheating had ended, but because the machines they had bought to police the cheating turned out to accuse the innocent at a rate no dean could defend. Turnitin’s detector, the flurry of startups promising to sniff out synthetic prose, the browser plug-ins that scored an essay’s “AI probability” like a weather forecast — these were, by the third quarter of the year, being quietly shelved. In their place rose a slogan that has since hardened into policy: stop detecting, start teaching. Do not ask whether the student used the machine; ask whether the student can think with it, around it, against it.

This is the tension the column takes up this week — between detection as a regime of enforcement and AI literacy as a regime of judgment, and the possibility that the second is being asked to do the work the first failed at, without anyone examining whether it can. The arc is not a simple conversion narrative, prohibition giving way to enlightenment. It is messier and more interesting: a discourse that spent early 2025 frightened of what AI was doing to student cognition, pivoted mid-year toward a confident language of empowerment and “human-centered” integration, and arrived in the fall still unable to say what, precisely, replaces the detector once the detector is gone.

What follows traces two lines that rarely touch. The first is what the education commentariat has been saying — the swing from banning to embracing, the vocabulary of prompt engineering and critical thinking. The second is what has actually been happening — the false-positive scandals, the adoption statistics, the emerging research on cognitive offloading. Where these lines meet, there is genuine insight. Where they miss, there is a vacuum being papered over with a phrase.

II. What we’ve been saying

The conversation began in fear, and the fear was specific: that the student would disappear behind the output. In January and February of 2025 the dominant register was defensive, and our own earlier briefings registered it — the February 23 critical analysis of the AI-literacy discourse noted authors preoccupied with “empty vessels that parrot the outputs of AI,” while the January 20 briefing found the field organized almost entirely around a deficit, “closing the AI skills gap,” as though literacy were a hole to be filled rather than a habit to be formed. The early rhetoric, in other words, imagined the student as a site of contamination. The natural response to contamination is quarantine, and the detector was the quarantine made technical.

But even in that first quarter the counter-position was already loud. The case against banning arrived fully formed and slightly impatient, insisting that fear was the wrong posture. “Students using AI: It’s not that scary and shouldn’t be banned” is a title that does its whole argument in its headline, and it was not alone. “Generative AI in Education: Is Banning an Option?” reported that fifty-two percent of U.S. adults already believed schools should teach students how to use the tools rather than wall them off — a majority, however thin, for integration over prohibition. The rhetorical center of gravity was shifting before the policies did.

What is striking, reading the first quarter whole, is how the anxiety and the enthusiasm ran in the same current rather than against each other. The same season that produced “Is AI Making You Dumber?” — a headline built to alarm, reporting that younger users “often accept” AI output without scrutiny — also produced “Opinion: Beyond Grades,” in which a school director in Dubai argued the machine could foster the very faculty the alarmists feared it would erode. The discourse was not divided into optimists and pessimists so much as it was internally divided, each essay carrying its own hedge. “Critical thinking skills erosion: the downside of artificial intelligence” warned that automating decisions “can diminish our capacity for analytical reasoning” — and then, like almost every piece in the genre, pivoted to the redemptive clause: which is exactly why we must teach differently.

By the second quarter the hedge had resolved into a program, and the program had a vocabulary. The word “ban” nearly vanished; the phrase that replaced it was “human-centered.” “The Future of Work Is Human-Centered AI. All We Need Is a Shift in Mindset” is representative not because it was wrong but because it was frictionless — the conversation, it declared, had been “stuck in a cycle of extremes,” and the way out was simply to change one’s mind. “How AI encourages critical thinking and creativity” completed the inversion the alarmists had feared: the tool that would make you dumber was now the tool that would sharpen you. Our own May 18 briefing caught the institutional face of this turn, the “AI Essentials Summer Series” and its kin, initiatives designed to equip educators rather than to catch students.

Underneath the optimism, though, a harder claim was surfacing — that the absence of teaching was itself the scandal. “The AI Education Gap: Why Schools Need AI Policies Now” put the reproach plainly: educators had adopted the tools faster than they had figured out how to teach with them, and the vacuum was pedagogical, not technological. By the third quarter this had become the settled frame. “Teaching Students to Think with AI: Why Prompt Engineering Belongs Across the Curriculum” no longer bothered to argue against banning; it treated integration as accomplished fact and moved to method. The question had migrated, over three quarters, from should they use it to how do we grade the thinking — a migration that quietly assumes the second question is answerable, and answerable by the same institutions that could not make the first one work.

III. What’s been happening

The reality that ran beneath this rhetoric began with a failure of the machines. The detector was supposed to be the neutral arbiter, and it was neither neutral nor accurate. “AI Detection Tools’ False Positive Problem in Student Writing” documented what teachers were finding on their own: honest students, disproportionately those writing in a second language or in the flattened prose that schools themselves reward, were being flagged as synthetic. The tool that promised to restore trust manufactured suspicion instead. This is not an incidental bug. It is what these systems are. Janelle Shane, describing image classifiers in You Look Like a Thing and I Love You (2019), notes that when a company’s AI “is unsure about an image, it gets sent to humans to categorize” — the machine’s confidence is a performance layered over a hand-off it cannot make on its own. A detector that renders a verdict on a teenager’s paragraph has no such hand-off. It emits a number and calls it certainty.

And the number’s provenance is genuinely obscure, even to its makers. Shane again, on training a model to attend to the right part of an image: researchers found that “more than just a tiny bit of influence would make it perform much worse. Even more confoundingly, researchers don’t know exactly why” (You Look Like a Thing and I Love You, 2019). The detector schools were bolting onto their disciplinary process was a system whose own designers could not fully explain its behavior — deployed as an instrument of judgment against students who would face real consequences. Once that became visible, abandonment was less a philosophical turn than a liability calculation.

Meanwhile the actual adoption of these tools raced ahead of any policy meant to govern them. By mid-year, sixty percent of teachers reported using AI in their lessons, according to the Forbes account of the education gap — a majority of instructors using the technology daily inside institutions that had, for the most part, no coherent statement about how students might use it. The K-12 sector absorbed generative AI in the roughly three years since ChatGPT’s release under what one account frankly called a “belief” — the belief that students and teachers need the tools — rather than under evidence that they help. The enterprise world moved in parallel, with the CIO account of workplace AI worrying openly about “digital amnesia,” the erosion of retention when a system remembers so you need not.

That worry has begun to acquire empirical weight, and this is the part of the record the enthusiasts move past quickly. A study warning that overusing AI tools may weaken critical thinking and brain activity circulated widely in late summer, giving the earlier anxieties a measurable spine. The mechanism it describes is cognitive offloading — the same phenomenon Meredith Broussard gestures at in Artificial Unintelligence (2018) when she insists on “a difference between human thought and computation,” the human capacity to rotate a cone in the mind, to hold and manipulate a thing without a machine’s help. When the offloading becomes habitual, that capacity is what atrophies. The fear of the first quarter, in other words, was not hysteria. It was pointing at something real, which the confident second-quarter pivot did not so much refute as decline to discuss.

What schools actually did, then, was neither ban nor detect nor coherently teach. They adopted first and reasoned later. The most honest institutional research of the year admitted as much: AI in higher education is “not simply a matter of technological availability, but a sociotechnical process shaped by individual perceptions, interpersonal networks, and institutional contexts” — a careful way of saying that what a given classroom does with AI depends less on policy than on the particular teacher’s trust, the particular department’s culture, the particular provost’s nerve. The India Today account pressing for “systems thinking and pedagogy” over hype was describing what did not yet exist. The retreat from detection was real and largely defensible. The construction of the thing meant to replace it was, and remains, mostly aspirational — a set of workshops, a summer series, a promise that judgment can be taught by the same offices that could not make enforcement work.

IV. Where they meet, where they miss

Where the two lines meet, they meet on a true insight: detection was never literacy, and treating a student’s use of AI as a crime to be forensically proven was a category error that deserved to collapse. The commentariat that argued against banning was right, and the schools that abandoned their detectors were, on the evidence of the false-positive problem, acting responsibly. A regime that accuses the honest to catch the dishonest has failed on its own terms. To that extent the arc bends toward something better.

But the miss is large, and it sits exactly where the rhetoric is most confident. “Teach AI thinking” has become an incantation precisely because it is unfalsifiable — nobody is against it, and almost nobody has specified it. The prompt-engineering-across-the-curriculum proposal is the closest thing to a method, and even it collapses a deep question into a technique: as if judgment about when not to reach for the machine could be taught as a subspecialty of learning to reach for it well. The harder truth the enthusiasts skirt is that the cognitive erosion the brain-activity study describes is not a misuse of the tool. It is a normal use of the tool. You do not offload your thinking by prompting badly; you offload it by prompting well and accepting the result. Fluency and dependency are the same gesture seen from two angles, and no curriculum yet distinguishes them.

There is also a deeper mystification worth naming, because the whole “think with AI” project rests on it. Kate Crawford, in The Atlas of AI (2021), returns to the founding moment at Dartmouth where John McCarthy “boldly argued that the differences between human and machine tasks were illusory.” That claim — that the human and the machine are doing the same kind of thing, only at different speeds — is the unexamined premise beneath every promise that a student will learn to “think with” a language model as a partner in cognition. Broussard’s cone says otherwise. The MIT Press volume AI Ethics frames the same fault line as a set of “tensions” that “continue to divide the minds and hearts in this discussion,” and warns of “the risk of new forms of paternalism” when we hand judgment to systems that manage us in the name of helping us. A detector that surveils students and a personalized tutor that quietly shapes what they attend to are not opposites. They are two faces of the same delegation of authority to a machine, and the second is harder to see because it wears the language of empowerment.

This is where the column plants itself. The retreat from detection was the right retreat for the wrong reason — abandoned for its inaccuracy rather than for its logic of surveillance — and so the logic survives, migrating from the crude detector into the friendlier tutor. Schools traded an instrument that policed students for instruments that manage them, and called the trade literacy. The reader deserves the plainer account: nobody has yet built the thing that “teaching AI thinking” names, and the phrase is doing the work of pretending they have.

V. The longer view

The honest position at the end of this arc is neither the alarmist’s nor the enthusiast’s. The alarmist was right that something in the student’s cognition is at stake and wrong to think a filter could protect it. The enthusiast was right that prohibition was a dead end and wrong to imagine the alternative had already been invented. What actually happened is that a failed technology of enforcement was retired and a rhetoric of judgment was installed in its place before the practice of judgment existed — a substitution of vocabulary for method, performed by institutions that adopted the tools, as the K-12 record shows, on belief rather than evidence.

The work that remains is unglamorous and specific. It means teaching, against the grain of every tool’s design, the discrimination between the task you should offload and the task whose difficulty is the point — the essay whose struggle is the learning. It means treating the personalized tutor with the same skepticism the false-positive detector finally earned, because both ask a student to trust a system that cannot explain itself. Detection asked the wrong question — did you use it? — and got a wrong answer often enough to discredit itself. The question that should have replaced it is not how do you use it well but what, in you, must remain unassisted for the using to mean anything. Until a school can answer that, it has not replaced the detector. It has only stopped looking.

What Detection Couldn’t Teach

I. The question of the week

II. What we’ve been saying

III. What’s been happening

IV. Where they meet, where they miss

V. The longer view

References