The Day We Close the Last Book

Where Are Books Heading in the Age of AI?

Joo-Ho Lee · October 19, 2026

Prologue The Last Thirty-Two Volumes

In 1768, the first installment of the Encyclopaedia Britannica appeared in Edinburgh, Scotland. For more than 240 years afterward, its name was almost synonymous with “all the knowledge in the world.” A full set of Britannica, its gold lettering gleaming on the living-room shelves, was a sign that a family valued knowledge.

In March 2012, Britannica announced that it would stop publishing its printed encyclopedia. The thirty-two-volume 2010 edition would be the last. A fortress of knowledge weighing nearly sixty kilograms had moved from the bookshelf to the screen. Wikipedia had launched just eleven years earlier.

Britannica itself had not disappeared. Its writing and editing continued online. In one sense, this was a demonstration that knowledge could leave paper behind and reach more people. At the time, dictionaries and encyclopedias were seen as books for looking things up, so it seemed natural for them to give way to search. Novels, textbooks, and scholarly books, people believed, would follow a different path.

I suspect that boundary will not hold forever either. But the book I worry about is more than an object bound in paper. It is a form in which one person thinks at length and leaves a written record, and another follows that thinking all the way through. That can still happen with an e-book or a long piece of writing on a screen.

What happens, though, if gaining knowledge becomes ever easier while both the people who create and preserve it and those who read it deeply grow fewer? This essay is an attempt to imagine such a future. The outcome is not predetermined. Still, I think it is a landscape worth looking at before we become too comfortable with convenience.

1 A Textbook Just for Me

Imagine a first-year university student in the not-too-distant future. He is taking thermodynamics, but he does not buy a thick textbook. Instead, he asks an AI: “Explain why entropy in an isolated system does not decrease. Assume I only know high-school probability.” The AI remembers the concepts that confused him last week, draws examples from the soccer he loves, and gives him questions to check his understanding. If he gets an answer wrong, it explains again from a different angle. At two in the morning, or on the tenth repetition of the same question, it never tires.

This is also an old dream of education. In 1984, the educational psychologist Benjamin Bloom described research on the effects of one-to-one tutoring. In that research, the average tutored student performed better than about 98 percent of students in conventional classrooms. Of course, this does not mean that the same effect occurs in every setting. What mattered was the difference that teaching tailored to a student could make. Since giving every student a personal tutor was not feasible, textbooks served to carry one expert’s knowledge to thousands or tens of thousands of readers.

Textbooks have done that job admirably, but no single book can suit everyone. It is dull for those who already understand and difficult for those who do not. In fast-moving fields, new research accumulates while readers wait for a revised edition. A well-designed AI tutor can ease these limitations by adapting its explanations and finding and connecting up-to-date sources. Yet current sources do not automatically make an explanation accurate. What to select and how to teach it still need to be tested.

Some intriguing experiments have already appeared. A study conducted in a Harvard physics course in the fall of 2023 and published in 2025 found that students using an AI tutor designed around educational principles learned more in less time than students learning the same material in an active-learning classroom. In an after-school program run by the World Bank in Nigeria in 2024, students who combined AI tutoring with teacher guidance also made clear gains in English over six weeks. These are not proofs that textbooks or teachers are no longer needed. They do suggest that the routes students take to learning may change.

What, then, will happen to textbooks in specialized fields? At first, AI will be an aid that “reads” a textbook and explains it. Before long, some students may skip the book and ask the AI directly. Experts, too, may find it easier to leave behind papers and lecture materials and let AI do the explaining, rather than spend years writing a book. Some specialized textbooks will cease to have new editions.

But a textbook does more than answer questions. It decides what should be learned first and what should come later, and places before the reader problems they do not yet know how to ask about. Can studying by getting answers only to the questions I already have take over that role? What we may lose is not simply a book, but the effort to bring an entire field together from beginning to end.

2 A Flood of Papers and a Drought

Academia is in the midst of a flood. According to the US National Science Foundation, roughly 3.3 million science and engineering papers were published worldwide in 2022 alone. That is more than 9,000 a day. No researcher can easily keep up with every paper in an expanding field.

The change is striking in my own field, robotics. ICRA, one of its leading conferences, received 5,088 paper submissions in 2026. ICRA 2027, which I am helping to prepare, has received more than 8,000. Comparing the previous year’s official announcement with the figure I have learned through this year’s preparations suggests an increase of over 57 percent in a single year. I suspect that recent improvements in AI have played a significant part in this surge. AI itself opens up new research topics, while also easing the work of implementation, analysis, and writing papers. Submission counts alone, of course, cannot tell us how much of the increase is attributable to AI.

There are also signs that AI is widely used in academic writing. A team led by Dmitry Kobak at the University of Tübingen in Germany analyzed changes in the vocabulary of abstracts in the medical database PubMed. They estimated that at least 13.5 percent of abstracts in 2024, roughly one in seven, had been written with help from large language models. One clue was a sudden rise in the use of particular words, such as “delve.” This is evidence of writing assistance, not evidence that the research itself is false.

Making text easier to produce, however, makes it easier to generate weak manuscripts as well as to communicate careful research. Researchers have more papers to read, and reviewers must sift through more claims to find those they can trust. Verification struggles to keep pace with production.

This flood may eventually take a different form. Small experimental findings or routine analyses may be shared through databases, code, and brief research notes rather than lengthy papers. Knowledge would not disappear; the form of its record would change. A paper is both a letter announcing a result and a record of the methods used and the people responsible for its claims. Those functions will not become unnecessary.

The drought I worry about does not arise from a fall in the number of papers. As reporting results becomes easier, questions that take a long time and offer no assurance of success may be pushed aside. Even as more manuscripts appear, less research may be devoted to wrestling patiently with genuinely new problems. That kind of drought can happen in the middle of a flood.

3 Knowing the Answer Before We Try

In 1928, the British bacteriologist Alexander Fleming noticed something unusual in a dish where he was growing staphylococci. Mold had appeared, and bacteria were not growing around it. This was the starting point of the discovery of penicillin. The history of science is full of such moments. An unexpected result, a failed experiment, or a strange number turning up in an unlikely place became a doorway into a new world. Chance mattered, but so did the presence of someone willing to look closely at an anomaly rather than simply throw it away.

Today, the range of scientific outcomes we can predict before trying something is expanding rapidly. AlphaFold greatly improved our ability to predict a protein’s three-dimensional structure from its amino acid sequence, and about 200 million protein structure predictions were subsequently made publicly available. In 2023, Google DeepMind’s GNoME predicted 2.2 million new crystal structures, identifying about 380,000 as computationally stable candidates. That same year, the autonomous A-Lab at Lawrence Berkeley National Laboratory in the United States reported that robots had mixed, heated, and analyzed materials to synthesize 41 new compounds in 17 days. Critics questioned the novelty of the compounds and the validity of the analysis. Even so, the direction is worth watching: AI and robots performing physical experiments and changing their next attempts in light of the results.

These tools let us rapidly explore possibilities that people could never examine in full. We are getting faster at picking the low-hanging fruit. It does not follow that we will soon run out of fruit altogether. One discovery may reveal branches we could not see before.

There were signs, however, that achieving the same gains required more effort even before AI. In a study published in 2020, Stanford’s Nicholas Bloom and his colleagues estimated that the number of researchers needed to double chip density every two years had grown more than eighteenfold since the early 1970s. A 2023 Nature study by Michael Park and his colleagues analyzed citation relationships among papers and patents and reported a declining share of work that changed the direction of existing research. The absolute number of such contributions, however, remained broadly stable. Rather than showing that major discoveries are disappearing, these findings ask whether the growing volume of research is producing a corresponding number of shifts in direction.

We do not yet know how AI will change this pattern. Simulations are constrained by the laws and assumptions we put into them, while physical experiments can yield results beyond those expectations. If AI enables more experiments, unexpected discoveries may actually become more common.

What concerns me more than the limits of the tools is how we get used to using them. Might we choose only problems that are easy to predict, dismiss unexpected results as noise, or abandon questions because they are hard to compute? If the way research is evaluated and rewarded also favors quick, easy gains, sustained attempts to reach the higher branches can lose their place. The better we become at prediction, the more consciously we must protect the value of looking beyond what we expect.

4 An Age When Becoming a Writer Is Harder

On February 20, 2023, the American science-fiction magazine Clarkesworld closed its submissions. That month, it had received about 700 submissions it considered human-written, alongside 500 it judged to be machine-generated. Ironically, the magazine’s mascot is a robot. After adjusting its procedures, it reopened submissions on March 12. This was not a moment when literature’s doors closed forever. It did show how suddenly the work of keeping those doors open could multiply.

The issue is not simply whether AI can write a good novel. When manuscripts arrive in quantities far beyond anyone’s capacity to read and select them, editors’ time and readers’ attention run short first. In that situation, people tend to turn to names they already know. A famous author’s name becomes a guide to choosing among countless pieces of writing.

But every famous author was once unknown. They built a name by submitting short stories to magazines, entering competitions, and publishing a first book with a small press. If those entry points become clogged with an unmanageable volume of manuscripts, editors lose the room to discover new voices. New routes to readers may open, but the fact that anyone can easily put writing out into the world does not mean that everyone can find an audience.

Even after finding readers, making a living is another matter. A 2022 survey commissioned by the UK’s Authors’ Licensing and Collecting Society, or ALCS, found that the median annual writing income of authors whose main occupation was writing was about £7,000. That was income from writing, not their total income. Adjusted for inflation, it was about 60 percent lower than in the 2006 survey. Earning a living as a writer was already difficult before AI. Adding large quantities of cheaply produced text may put further downward pressure on the value of human writing, at least in some areas.

There will still be readers who seek out work written by people. It may even become more prized precisely because it carries a human perspective and lived experience. Yet recognizing that value and creating conditions in which the next generation of writers can keep writing are different things. If literature sells only the books of established names while those writing their first books cannot stay afloat, it may look abundant on the surface while slowly drying up within.

5 A Generation That Does Not Read Books

In 2025, researchers at the University of Florida and University College London analyzed the American Time Use Survey, drawing on data from more than 236,000 Americans. The proportion who read something for pleasure on a given day fell from about 28 percent in 2004 to about 16 percent in 2023, a decline of roughly 40 percent. Leisure reading in this survey included e-books and audiobooks, not just printed books.

Where did that time go? The researchers did not settle on a single cause. Alongside the growth of digital media, they identified economic pressures, reduced leisure time, and differences in access to books and libraries as possible explanations. Still, we can feel in everyday life how short videos and summaries compete with time spent reading. A fifteen-second clip delivers laughter and surprise faster than a page can. A three-minute summary video tells us the plot of a three-hundred-page book.

Knowing the plot is not the same as the experience of following a whole book. Reading asks us to remember an earlier sentence as we move to the next, stay with passages we do not immediately understand, and hear out arguments we dislike. If we spend less time doing this, long texts may become increasingly difficult and unfamiliar. For children, making time and space to read may matter as much as telling them to pick up a book.

The way we seek advice is changing too. People turn to self-help books for knowledge, but also for “the words I need to hear right now.” Now there is an AI that asks about my situation, listens to my circumstances, and offers advice that seems tailored to me. A 2025 report analyzing examples of use found online also identified counseling and emotional support as prominent uses. It was not a representative survey of how frequently all users use AI, but it does show people treating it as more than a search box.

An AI that answers me alone can feel closer than a book written for thousands of readers. That is both its appeal and a cause for concern. If I receive only explanations I am likely to enjoy, will I have fewer encounters with unfamiliar ideas and less practice in staying with them as I read?

I do not think books are about to disappear. But for people growing up in a world where AI is much smarter than it is today and woven into every corner of life, reading a long text may occupy a different place. Books may remain, while the habit of developing our thinking through them narrows into a taste shared by only some.

6 An AI With Nothing Left to Learn

At this point, a reasonable objection arises. What does it matter if people read and write fewer books? Why not let AI organize knowledge and even make new discoveries for us? It is a fair argument. But it is worth remembering what today’s AI grew from. Human records in books, papers, newspapers, and internet forums were an important part of its foundation.

Stack Overflow, where programmers ask and answer one another’s questions, shows how that foundation might change. Its accumulated public questions and answers also served as material from which AI learned to code. According to one analysis, the number of new questions per month fell from about 108,000 in November 2022, when ChatGPT appeared, to roughly a quarter of that two years later. The decline had begun earlier, so it cannot all be attributed to AI. Even so, when questions once asked in public move into private conversations with AI, fewer records may accumulate for other people to find and read.

A new programming language can also be learned from official documentation, code, and the results of running it. The decline of a single question-and-answer site does not mean that AI stops learning. Yet accounts of where real users got stuck and which explanations helped them understand offer something different from official documentation. The path by which these varied experiences return to the public store of knowledge may be narrowing.

We also need care when filling the gap with AI-generated text. In a study published in Nature in 2024, indiscriminately feeding model-generated data back into the training of successive generations caused rare and distinctive content to disappear and the quality of outputs to break down. The researchers called this “model collapse.” It does not mean that all synthetic data is harmful. We should distinguish this process from approaches that preserve original data, verify generated results, and add real-world observations.

In a 2024 study, the research organization Epoch AI estimated that if the expansion trends of the time continued, nearly all publicly available human text suitable for training could be put to use between 2026 and 2032. That is not the date when writing disappears or AI progress ends. It is a forecast that improving performance simply by collecting more public text may become harder. Other routes to learning, including experiments, observation, and verified synthetic data, remain open.

Still, I wonder whether the diversity of the records people leave can be easily replaced. Those records contain frustrations we have not yet named, the sensibilities of people who have lived through a particular time, and curiosities far removed from efficiency. Discovering a new material through experiments and understanding the lives people lead and the questions they carry are not the same problem.

Here I am reminded of the ancient episode known as the burning of books and burying of scholars. According to the Records of the Grand Historian, an order under Qin Shi Huang called for the burning of various privately held books, while exempting books in the keeping of court scholars. The details of the episode and the subsequent loss of knowledge remain debated. Even so, the image of power seeking to restrict who could own and read books has endured.

We cannot simply equate today’s AI with that period. There are many AI systems, and libraries and internet archives remain. We should nevertheless be wary of a situation in which our routes to knowledge narrow to a handful of services and users stop seeking out original texts or leaving records for one another. No one needs to take books away by force if convenience leads us to give up reading and writing for ourselves.

This is the paradox I most want to convey. If people do little beyond asking AI for answers, while doing less to create new questions and records, the foundations of knowledge on which AI depends may weaken too. Answers become easier to obtain, while the questions that could produce the next answers grow rarer.

7 What Stagnation Might Look Like

Technology does not advance in every direction at once. Today’s passenger aircraft may be much safer and more efficient than those of decades ago, yet their cruising speed has not increased very much. Speed alone cannot establish that aviation technology as a whole has stagnated. Even so, it is worth reflecting on how one familiar direction of progress can remain still for a long time.

Now imagine something similar in the world of knowledge: a world where AI’s capabilities keep growing, but the range of questions we ask grows narrower.

On the surface, that world is very comfortable. Whatever we want to know, AI answers immediately. Illnesses are detected early, goods are inexpensive, and endlessly entertaining videos flow across our screens. We get better and better at solving familiar problems.

Questions that require a long commitment, however, keep being pushed back. Research whose immediate use is hard to explain, books whose readership is uncertain, and ideas that question what everyone takes for granted struggle to find time. This is not just a question of whether AI can generate such questions. Someone still has to regard them as important, test them in practice, criticize the results, and leave a record for the next person.

If the habit of following another person’s thought through a long text weakens, and those creating new writing struggle to keep going, fewer people may be available to do that work. Answers abound, while our ability to examine them and choose a new path grows weaker.

Perhaps no one will recognize it as a crisis. Daily life is still comfortable, and the screens still glow. Stagnation does not arrive with a roar. It can arrive before we even notice what we have stopped asking.

Epilogue Before We Close the Last Book

When Britannica printed its final edition in 2010, few people knew that those thirty-two volumes would be the last. Endings often pass without our knowing they are endings. Perhaps some book we are reading now will turn out to be the last in its field.

The future pictured here is a possibility, not a prophecy. And I hope that possibility never becomes reality. I write in the hope of being wrong because noticing what might happen before it happens gives us a chance to find another path.

AI is a remarkable tool. It can create explanations just for me, assist with countless experiments, and listen to me at two in the morning. I am not arguing that we should reject it. I am asking that we also care for the world of knowledge around it. We still need people who ask the first question, take another look at a failed experiment, think an idea through and write it down, and read that writing and argue with it.

Protecting tomorrow’s progress may therefore begin with surprisingly modest acts. Read a long piece to the end. Write down your own thoughts, however imperfectly. Ask one more question of the answer AI gives you. Sometimes deliberately step away from the path a calculation recommends, as Fleming did when he looked again at his dish instead of simply throwing it away. And preserve room for research and writing so that the people who do these things can keep doing them over many years.

The day we close the last book may come. But putting that day off, even a little, begins with turning one more page today.

Sources and Further Reading

Harvard AI tutor · ICRA 2026 submissions · Kobak on AI-assisted writing · Park on research disruption · ALCS author earnings · UCL reading trends · Stack Overflow questions · Nature model collapse · Epoch AI training data

The ICRA 2027 figure reflects information the author learned during conference preparations and is distinct from the publicly announced ICRA 2026 statistics.