AI did what World Book Day couldn't: it got British adults reading again.

By my own rough count I now read somewhere between ten and fifty times more words a week than I did before ChatGPT, Claude, Gemini and Copilot turned up on my laptop. Specifications, drafts, summaries of the drafts, the second draft, the explanation of why the second draft differs from the first. My optician has not said anything. She has started to look at me differently.

I assumed this made me unusual. Then I did the sum, and the sum says something stranger: at the top of that range I am not reading at all. Not in any sense a reading researcher would recognise. And if that is true for me, it is true for every heavy AI user in the country, which is now a surprisingly large group.

Half the country stopped reading. Then it started typing.

Two surveys, a year apart, tell the story between them. In April 2025 The Reading Agency reported that only 53% of UK adults read regularly, down from 58% in 2015. Nearly half (46%) said distraction stops them reading. Almost a quarter of 16 to 24 year olds said they had never been readers.

In April 2026 Ofcom reported that 54% of UK adults now use AI tools such as ChatGPT, Copilot or Gemini, up from 31% the year before. Three quarters of online adults read AI-generated search summaries at least sometimes. The government's own DSIT survey of 30,698 people puts generative AI use at 59% in the previous three months.

Every one of those interactions ends in text. A chatbot does not reply with a shrug. It replies with four paragraphs, a bulleted list and an offer to go further. The country that could not finish a novel has, without anyone planning it, taken up reading again at industrial scale. World Book Day has been trying to do this since the late 1990s and would have settled for a fraction of it.

The arithmetic.

The most reliable figure for adult silent reading comes from Marc Brysbaert's meta-analysis of 190 studies: 238 words per minute for non-fiction, 260 for fiction. Slower than most people assume, and that is the point.

Take a typical reader on half an hour a day. That is about 7,000 words a day, or roughly 50,000 a week. Multiply by my own range. Ten times is 500,000 words, which at 238 words a minute takes 35 hours. Fifty times is 2.5 million words, which takes 175 hours. A week has 168. At the top of the range I would need to read for longer than time exists, and at the bottom I would need a second full-time job to fit it in.

For scale, the complete works of Shakespeare run to about 884,000 words. Two and a half million words a week is just under three of them, every week, with no interval and no ice cream. These are estimates, and I have rounded freely, but the shape of the problem survives any amount of rounding. Something has to give, and it is not the week.

What you're doing instead has no name yet.

What gives is the reading. Not the volume of text in front of us, which is real, but the act of taking it in line by line. We already knew people do not read web pages: the Nielsen Norman Group estimated years ago that a visitor reads at most 28% of the words on a page, more likely 20%, in a scan shaped like an F. A chatbot pane is a web page that refills itself.

The research on screens points the same way. Delgado and colleagues pooled 54 studies and about 170,000 participants and found a small but consistent penalty for reading informational text on screens rather than paper, worse under time pressure. The more worrying finding sits underneath: screen readers are more confident they understood than print readers, and less likely to be right.

Then there is what AI does to the job of reading. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers who use ChatGPT or Copilot at least weekly and collected 936 real examples. Their conclusion was that thinking does not disappear, it moves: from doing the task to verifying, integrating and stewarding what the tool produced. The more people trusted the tool, the less of that they did.

So here is my working name for what heavy users are actually doing. We are reading like sub-editors. Scanning for the sentence that matters and the sentence that is wrong, and skipping the rest. Hannah Clarke, the HR business partner who became her building's unofficial AI helpdesk, does it on her own Copilot output and then on everyone else's. Sanjay Mistry does it fifty product descriptions at a time. Tomasz Wiśniewski does it on every email he polishes before it goes to his boss. None of them was trained for it. None of them is paid for it.

You are not reading more. You are editing more, without ever having been hired as an editor.

The good news, and it is real.

Reading volume is one of the few things in education research that behaves itself. Cunningham and Stanovich showed in 1998 that how much a person reads predicts their vocabulary, general knowledge and verbal ability, and that the effect holds after you control for raw intelligence and reading skill. Exposure is the mechanism. More words in, more words available.

Which means the people who use these tools well, who ask tight questions, read the answer properly and push back when it is wrong, are getting something the school system never gave most of them: a daily supply of structured, edited, grammatically clean prose on exactly the subject they care about. Tomasz is reading more corrected formal English per week than an evening class could offer him. Hannah, who is dyslexic and found that AI was the first tool to make report writing feel fair, is now doing more reading than anyone in her building, on her own terms, at her own pace, in a font she chose.

Even Chloe Barnes, the Year 10 English teacher who tried ChatGPT only to see what her pupils see, has a lever here she did not expect. Her pupils are reading more words than any cohort before them. Just not the ones she set. The skill nobody is teaching them, reading machine prose with a raised eyebrow, is the one she is best placed to teach. It is close reading. She has been doing it with Priestley for years.

There is a compounding effect too. Better questions produce better answers, and better answers are worth reading properly, which produces better questions. The curators pull away from the skimmers, not because the tool is different but because what they are reading is.

The bill.

Start with the words themselves. A benchmark published in January 2026, YapBench, measured how much assistant models say beyond what the question needs and found roughly a tenfold spread between the terse ones and the talkative ones. The padding is restated prompts, generic caveats and boilerplate openers. A separate study of multi-turn conversations found that longer replies went with worse answers on five tasks out of six. A fair slice of the reading load is text that should never have been written.

Nobody has ever complained that a chatbot's answer was too short. The cost of that lands on the reader, and the reader is you.

Then there is what the engineers have started calling comprehension debt: the gap between how much code a system contains and how much of it anyone actually understands, which accrues at the review stage, because the volume outruns the reviewer. Swap "code" for "policy", "proposal" or "product listing" and it describes every office in the country. The drafted policy Hannah approved at ten past five. The listing Sanjay signed off in a batch. Clean, well formatted, superficially correct.

An MIT Media Lab study of 54 students, "Your Brain on ChatGPT", found that participants who wrote essays with the model struggled to quote from them minutes later and reported less ownership of the work. It is a small study, only 18 people did the follow-up, and it measured writing, not reading. Treat it as a hint rather than a verdict. The hint is that passing text through yourself is not the same as keeping it.

And the people are tired. A February 2026 survey of 1,250 US corporate workers found 37% reporting AI fatigue and 31% saying their workload had grown since the tools arrived, against 16% who said it had shrunk. US figures, commissioned by a careers site, so hold them loosely. But nobody who reviews Copilot output for a living will find them surprising.

So what for the UK.

Britain measures adult reading in books. The Reading Agency, the National Literacy Trust and World Book Day all count the same thing, and none of them counts the Copilot pane. The biggest shift in adult reading habits in a generation is invisible to the bodies whose job is adult reading. Somebody should ask Ofcom to put a word count in next year's survey.

Employers have a sharper problem and, unusually, an existing law to hand. The Display Screen Equipment Regulations have obliged UK employers since 1992 to assess workstations, plan breaks and fund eye tests for staff who use screens. A tenfold rise in words read per week is a DSE change. Hannah can take that to IT and to her board without needing an AI policy first, which is handy, because her housing association still has not written one.

For Dr Nadia Rahman and anyone else building these systems, output length is a cost line nobody owns. Teams track tokens spent and seconds of latency and never the minutes a human spends reading the result. An eval that scores an answer on correctness alone will keep rewarding four paragraphs where one would do. Nadia can fix that in an afternoon of prompt and eval work. Most teams never will, because the person paying the cost is not in the room.

Schools last. Chloe's Year 10s are reading more machine prose than any cohort before them, and no part of the English curriculum asks them to read it critically. That gap will not be closed by a ban. It will be closed by a teacher putting a ChatGPT answer on the board and asking the class where the writer lost their nerve.