The Great Unwiring

Right now, Generation AI is sitting exams. Not with their noses in textbooks, as my generation did, but with ChatGPT on their phones whilst Claude writes synopses based on information found via Google AI. Are these the skills they are training for the future? Or are we quietly sawing off the branch on which education systems and human enlightenment have grown for the past hundreds of years?

AI researcher Andrej Karpathy has observed: "You can outsource your thinking but not your understanding." I don't know if he is right. But before AI arrived, we never thought to ask the question. Today we see the answer playing out in classrooms everwhere.

Rebecca Winthrop leads the Centre for Universal Education at the Brookings Institution. In January 2026, she and her colleagues published the report A New Direction for Students in an AI World, the result of a year of global work they call a "pre-mortem": an attempt to learn from the mistakes of social media before it is too late.

As part of this work, the report introduces the concept of "the great unwiring" - a parallel to Jonathan Haidt's "great rewiring" concept from social media - describing specifically how AI erodes the cognitive capacities that schooling was supposed to build.

Reading the report and listening to interviews with Winthrop, there is little doubt that the risks of AI in education currently outweigh the benefits. I have gathered the five most alarming findings from the report here, and close with examples of possible paths forward.

1. AI is taboo

It is perhaps unsurprising that AI in education is shrouded in the forbidden. A key finding from Winthrop's research confirms that young people know they should not be using AI, and she compares it in an interview to the sensitive conversations adults have with teenagers about drugs and sex:

As a father of two children in secondary school, I cannot help but recognise the elephant in the room. My daughter graduated last year as part of the first generation to see AI overtake and outmanoeuvre them in most areas of knowledge. Today the technology is everywhere and impossible to avoid. Everyone pretends to do one thing in the open and does something else behind closed doors. Few talk openly about it with parents or teachers, because it touches on feelings of dishonesty and the fear of not being intelligent enough.

2. AI is teaching pupils to cheat

The report also documents a double layer of concealment. Pupils use AI to write assignments, then run them through "AI humanisers", systems that strip out all traces of machine-generated text so the work can be submitted without detection. Teachers use AI to produce presentations and worksheets, but say nothing. Both parties know the other is doing it, yet nobody speaks openly about it.

It is almost impossible to do anything about. The plagiarism-detection software available to teachers is unreliable and produces false positives for neurodivergent pupils, second-language speakers, and those with learning difficulties, whilst missing re-humanised AI output. Pupils know this, and they exploit it.

It is worth reflecting on what skills and values we are cultivating right now. Those who fare best in the current system are those who master the art of quiet deception, working out how to game it.

3. AI is creating a fear of becoming less capable

The greatest worry among my daughter's friends last year was that they had almost forgotten how to write an introduction and a conclusion when they sat their written exams, because several of them had spent the year outsourcing that part to ChatGPT.

Cognitive offloading, or "stunting" as Winthrop calls it, is the term used to describe delegating one's thinking to AI. Many pupils allow themselves to use AI when they already know the material, telling themselves it is fine for routine work. But it is a slippery slope, and students in the report express concern about whether they are losing the ability to start work on their own. The capacity to initiate, to reach for a blank document without an AI to ask, is something many pupils feel they are losing.

The evidence in the report for AI's negative impact on cognitive abilities such as reading and writing is considerable. Teachers describe how it undermines engagement, curiosity, independence and self-confidence. They speak of how AI introduces a strange double feeling: a student proud of a high mark, yet ashamed that the work is not entirely their own.

Most striking is that the concern does not come primarily from adults. It comes from young people themselves. They are acutely aware of what they risk when they offload and as the table below shows they are by far the most concerned party when it comes to the risk of cognitive degradation.

A recent British study found that the greatest fear among young people is not whether they will find a job when they finish. It is that they will stop being able to think and produce independent work.

4. AI creates a collective action problem

AI exposes the narrow and outdated measurement tools of our education systems. In a performance culture, results determine how success is measured. In secondary school, results mean the assignments you submit. For a long time this has been a teacher's best basis for assessing learning. It no longer is.

One student in the study said she was proud to receive a middling grade in a difficult subject, because she was actually learning the material and doing the work herself, whilst all her classmates were using AI and receiving top marks. The same student said, after a brief pause: "But I'm not sure how much longer I can keep this up. I do want to get into a master's programme." Students are caught in a dilemma. They understand the harmful effects of AI, yet see no alternative. They are trapped in a coordination failure created by the system and the technology.

Reading the quotes, it is unsettlingly reminiscent of the dynamics we have long observed among young people and social media, where a majority admit they would rather be free of the platforms, but nobody dares go first, so everyone remains trapped in the same loop. This is what the makers of The Social Dilemma have called a "collective action problem": a fundamental coordination failure that cannot be resolved collectively without intervening in the system driving the behaviour.

For many students it is no longer a question of morality. It is game theory, and it resembles unregulated doping. If you do not use AI, you risk falling behind those who do. This is precisely the mechanism that the Danish EVA survey from February 2026 documents empirically: only 21 per cent would use AI on an assignment if they believed no classmates were doing so. That figure rises to 40 per cent when they believe most classmates are.

5. AI erodes trust

The finding that concerns Winthrop most is not about grades or skills. It is about trust, and the situation is poor. This matters, as Winthrop notes: trust is something you do not miss until it is gone.

In the report, more than half of teachers say AI has made them more distrustful. They no longer believe that what students hand in is their own work. This makes teaching nearly impossible. You cannot help someone if you do not know what they understand and what they do not.

Half of students, meanwhile, do not trust their teachers either. They believe teachers are secretly using AI to mark assignments and prepare lessons, and respond with criticism and suspicion, even when AI use is perfectly reasonable. Most students today are considerably more proficient with the technology than their teachers, and they experience it as unfair that the rules do not apply equally to everyone.

There is also, in Winthrop's research, the example of a university student who turned up at a professor's office hours to contest a grade, explaining that the professor was wrong, because ChatGPT had given a different answer.

Trust has broken down at three levels simultaneously: teachers do not trust students, because they cannot assess what is genuinely the student's own work. Students do not trust teachers, because they believe teachers are hiding their AI use. And both groups trust ChatGPT more than they trust each other.

AI education must be designed

There are, fortunately, stories of schools that have embraced AI with success, confirming that the problem is not AI itself, but our current uncontrolled deployment of it as a general intelligence. AI without pedagogical framing is genuinely harmful to learning, but under the right conditions, recent research shows that AI can actually advance it.

An international study from 2026 identifies a central distinction and points to a possible path forward for responsible AI use in education. The study identifies two primary ways students can use AI: as cognitive delegation, where AI does the thinking for them, or as a cognitive partner, where AI is used strategically to free up capacity for a more demanding next step.

The first path degrades critical thinking. The second actually strengthens it.

The study finds that students who treat AI as a partner they are accountable to and must evaluate critically develop stronger analytical thinking than those who use it as an answer oracle. It also finds that students who strategically delegate routine tasks to AI, such as summarising, structuring and initial drafts, and use the freed capacity to engage more deeply with the difficult parts, actually learn more, not less.

Two examples of AI-integrated education

One example of AI in practice is Alpha School in Austin, Texas, which has attracted significant attention for an approach that combines motivation, learning and social development. Geoffrey Hinton, one of the founding figures of modern AI, recently called it one of the best things to have come out of the technology, which carries some weight, given that he is otherwise deeply ambivalent about what he helped create.

Alpha School's model is markedly different from anything most of us recognise as school. The day begins with two hours of intensive screen-based learning, where pupils are guided by a personalised AI mentoring platform that awards points for measurable progress. The rest of the day is given over to practical activities: assembling furniture, solving a Rubik's cube, juggling, running a small Airbnb rental. Test results place students at the 90th percentile across subjects.

Alpha School is, however, a private institution with a fee-paying intake, and strong exam results are not in themselves evidence of deep learning. The model raises the same question the Brookings report poses: are students genuine cognitive partners with AI, or are they delegating their thinking away? That question calls for more documentation before the model can be offered as a scalable answer.

A third example is Kūlia Academy in Honolulu, a middle school of 150 pupils and Hawaiʻi's first with a dedicated AI and data science programme. Pupils do not sit in front of chatbots. They learn to code manually before they touch AI at all. They study rising sea levels by going out and measuring them along Oahu's coastline, then use AI to analyse what they themselves have collected. There is no homework, but a longer school day with plenty of outdoor activity.

Kūlia's Year 6 pupils achieved the highest results of any public middle school in Hawaiʻi in 2025: 75 per cent in mathematics, 80 per cent in English. On average, pupils progressed 2.4 grade levels in mathematics and 2.6 in English within a single year.

Racing against the unwiring

The solution seems straightforward enough: integrate AI actively, and make deliberate choices about when it adds value and when it causes harm. That distinction is what is missing from the current debate. The question is not whether AI should be used in education, but when AI advances learning rather than dismantling it.

As long as current chatbots remain freely accessible, we are also obliged to AI-proof our existing systems. There are many ways to approach this, and collaboration with students themselves is essential. Winthrop describes one example where a student panel red-teams assignment briefs before they are distributed. If they can be hacked with a chatbot, the brief is not good enough. She also recommends a renaissance for oral presentations and in-class written work, currently the only assessment formats that actually measure what we want to measure.

AI literacy itself needs to go deeper than standard source criticism and prompt guidance. It means giving students and parents an open, non-judgmental conversation about what AI is, what it is not, and what it does to a person.

The student who took her middling grade with pride knew exactly what she was in danger of losing. The question is whether we, as teachers, parents and policymakers, are equally honest with ourselves. We are racing against time. The great unwiring is already under way.