The most revealing figure in the new research on reading is not the one about decline. It's the split inside a single study of more than 26,000 students in grades 7 to 12 in China: AI use lifted homework scores by 18 per cent and cut completion times by 30 per cent — and within six months it lowered monthly exam scores by 20 per cent.
Both numbers are real. That is the trap. The tool that makes the work look better is making the learner worse, and the homework is the only place we can watch it happen in real time.
Last week, this column followed MIT's reckoning with what AI is doing to education; the week before, the synthetic-science thread, where machine-generated output crowds out the slow verification science depends on. This week the thread turns inward, into cognition and reading itself — into what happens to a mind that never has to hold a difficult paragraph long enough to understand it.
The tool that makes the work look better is making the learner worse, and the homework is the only place we can watch it happen in real time.
The OECD has a name for the result. Across 91 countries and 760,000 fifteen-year-olds, average reading scores fell from 476 in 2022 to 461 in 2025. The share of "hasty readers" — students who race through a text and answer quickly but wrongly — nearly doubled between 2018 and 2025. In the United States, roughly ten per cent of children now read quickly but inaccurately, twice the rate of seven years ago.
But here the sources split, and the split is the interesting part. The technology story says screens and AI made us hasty. The schooling story says we were already hasty, and AI only industrialised it.
Consider the scene in The Spectator: a mother and son on a train south from Durham, three hours spent on a single comprehension question — why does the author compare the road to a gray silk ribbon? By King's Cross she is shouting. "I can't think when you're shouting at me." "You've been thinking for hours." "Mum, it's a stupid question."
The boy was right, and the diagnosis that follows is not about ChatGPT. It is about a curriculum that requires the endless analysis of excerpts and never the reading of whole books. Find the similes, the sibilance, the subordinating conjunction. The novelist Katherine Marsh made this argument in The Atlantic in 2023, before AI was the subject du jour: assessment pressure taught teachers to teach to the test, and so inspiring a love of story gave way to parsing passages.
The academic Carl Hendrick's formulation is sharper still. What we have is not a generation less cognitively capable but a generation raised in a culture of skills-based reading that implicitly tells them what you read matters less than how you read it. Some of them are past hasty. A teacher friend describes bright students who can say the words aloud, can scan for a simile, but cannot retell the story or anticipate what comes next. Not hasty readers. Zombie readers.
Now add AI, and the hollowing accelerates. OECD evidence shows over 40 per cent of students use AI to help draft school writing weekly or daily. The English teacher and writer Alex Quigley warns that AI-generated class reading saves time while stripping out the complexity and sophisticated vocabulary that actually teach anything — the short, hollow read that fits on a PowerPoint slide.
The performance data agree. A study of 1,000 students in Turkey found that those with unrestricted AI access did worse on closed-book tests than those with no access or with guardrails. Meanwhile 78 per cent of Australian children have used an AI assistant, 91 per cent of them for learning or translation.
What is being lost is not information. It is the struggle — and the struggle is the mechanism. The brain almost quadruples in weight from birth to adulthood without adding a single new neuron, the University of Tasmania neuroscientist Matt Kirkcaldie notes; it is all physical change driven by learning. His description of the AI reader is exact: "They might read the AI-generated text and understand what it says, but they never went through the struggle of working through the ideas and how they might fit in the overall picture."
That struggle builds cognitive reserve — the resilience that delays the onset of dementia and buffers stroke and brain injury, according to the neuroscientist James Vickers. Skip it, and you skip the construction.
Maryanne Wolf, whose Reader, Come Home is this year's Princeton Pre-Read, reports a study in which young readers of a short story retained its plot better on paper than on a Kindle. There is something in a book's concrete, spatial dimension that helps us. Her indictment of our collective experiment is worth reading twice: no internal review board would have permitted what our culture did with "a complete, quasi-addictive set of attention-compelling devices without knowing the possible side effects." The result is "decreasing attention spans, pushed beyond their cognitive limits by a flotsam of distractions and information that will never become knowledge."
That last clause is the whole problem. Information that never becomes knowledge. Byung-Chul Han, in Psychopolitics, puts the same idea in starker terms: storage and retrieval are not the same as remembering and interpreting. Big data, he argues, is elementary knowledge only, and the second Enlightenment — the one that elevates data over reason — is "an epoch without reason." James Marriott's new book carries the thought to its political conclusion: as reading fades, we may face a post-literate future resembling the pre-literate past — "mystical, tribal, and autocratic."
There is a second erosion, less discussed and harder to fight. The writer in Vuink notes that AI's tendency to atrophy sustained mental effort has been clear for some time; what is new is that worrying about AI is now sapping concentration too. On September 8, the researcher Jacob Coxon announced on X that he was quitting Anthropic, writing that both Anthropic and OpenAI are "gambling with our lives" and that the people building AI "earnestly believe that it could kill us all