When AI starts reading for people

Books can disappear in many ways — burned, neglected or left unread. Now they can be bought, cut apart, scanned and turned into data. This unsettling process is Anthropic's Project Panama. The company plans to acquire vast quantities of physical books, remove their bindings, digitize their pages and discard the originals. Internally, the project was described as "our effort to destructively scan all the books in the world".

The controversy is not simply about whether a paper book can be cut up, nor only about copyright. It raises a fundamental question: what happens when the world's written inheritance is increasingly consumed by machines?

AI promises easier reading. It translates difficult passages, explains unfamiliar concepts, compares books, answers research queries and provides summaries in seconds. In education, it acts as tutor, translator, examiner, lesson planner and research assistant. It can give a child in a remote school access to resources once available only to students with expensive tutoring.

But the more efficiently machines perform the informational part of reading, the greater the possibility that humans will abandon the cognitive work that made reading valuable in the first place. The question is whether AI will help people read better — or render reading unnecessary.

The educational shortcut

AI's role in education is compelling, especially when educational resources are scarce.

China's iFLYTEK is an instructive example. In Zhejiang's Kaihua county, a school with only six classes and 72 students has combined online lessons from master teachers with local teachers and AI-assisted learning. In the Xizang autonomous region, AI teaching systems and digital courses are being used to connect students with educational resources beyond their isolated surroundings. The premise is simple: technology can partially bridge the gap in the geographical distribution of teachers and educational resources.

The company's intelligent assessment system collects examination papers, assists with grading, preserves original papers and analyzes students' learning patterns. Its guiding principle is revealing: "AI handles whether it is correct; teachers handle what is good."

The company's "virtual scientist" takes the idea further. Instead of simply giving students an answer, it uses heuristic questioning and visual question chains to encourage them to work through a problem themselves. An AI tutor that says "the answer is X" saves time; one that asks "why do you think it is X?" educates.

Schools have incentives to experiment with AI, which saves teachers time and gives students quicker answers. But the problem is that efficiency makes unnecessary precisely the skills education should teach.

The limits of reading machines

Reading difficult books is inefficient, and that's the point. Serious readers don't just absorb conclusions. They get lost, reread paragraphs, disagree with authors, change interpretations and sometimes discover the original question was misguided.

Falk Huettig, a senior investigator at the Max Planck Institute for Psycholinguistics in Nijmegen, the Netherlands, has spent years studying how literacy affects cognition. His 2026 book The Perks of Being a Bookworm synthesizes research showing that reading shapes memory, attention, language processing and reasoning, going well beyond the acquisition of factual knowledge. In a recent interview, Huettig warned against students relying on AI summaries. An AI condensation may deliver facts, he said, but it removes the linguistic structure and complexity through which readers exercise working memory and abstract thinking. He warned that avoiding complex texts means missing important forms of cognitive training.

This is why a summary of War and Peace is not War and Peace. The problem is not just Leo Tolstoy's word count. Reading a long novel requires remembering characters, holding competing views, detecting irony, tolerating ambiguity and constructing meaning gradually. A summary gives the destination without the journey.

Historian Zhao Dongmei of Peking University likens reading to "searching": entering a book with questions and confronting evidence, arguments and contradictions. Dai Jinhua, a culture researcher at PKU, argues for "reading thick before reading thin" — understanding a text's intellectual, historical and literary networks before reducing it to a few conclusions.

AI excels at producing "thin". It can condense 500 pages into 500 words, 50 words or five bullet points. This is useful, but a culture that always opts for the five bullets may eventually forget why the 500 pages mattered.

Skills atrophy without practice. If students never formulate arguments, remember complicated narratives, struggle with difficult paragraphs or construct sentences from scratch, they may finish assignments but become less adept at the cognitive activities underlying it.

The distinction is between assistance and substitution. When a student asks AI to explain a difficult chapter after reading it, AI can deepen comprehension. When the student asks AI to replace the chapter altogether, the educational transaction changes. The machine hasn't merely helped with reading. It has done the reading.

From reading to parroting

The issue gets more complicated when AI enters the attention economy. Consider Byte-Dance's TikTok and Douyin. Algorithmic platforms can both encourage and undermine reading. TikTok's #BookTok community is a significant force in book discovery. According to NielsenIQ BookData and Media Control data cited by TikTok, more than 50 million books recommended by the #BookTok community were sold across major European markets in 2025. More than one-third of the respondents aged 16 to 39 in the surveyed markets discovered new books through the platform.

This is algorithmic reading's optimistic side. A teenager sees a video about a novel and buys it. A recommendation opens a door. But there is another possibility: the recommendation becomes the experience itself.

Books become plots. Novels turn into emotional reactions. Philosophy becomes "five lessons". History is reduced to dramatic anecdotes. A difficult work is pre-interpreted before the reader can form an opinion.

AI amplifies this compression. Why read the book when a chatbot can explain it? Why examine evidence when a model summarizes debates? Why wrestle with ambiguity when the machine offers smooth interpretations? This leads to pseudo-reading: the feeling of acquiring a book's knowledge without the experience of reading it.

Sune Lehmann, a cognitive scientist at the Technical University of Denmark, has drawn a clear line. Regarding AI-generated literature, he states: "I don't want to read anything written by AI. Period." The reason is not just technical. For him, fiction and poetry are ways of connecting with another human being and entering a universe created by another person. One may disagree with Lehmann's personal boundary, but his point highlights that reading is more than just information. It is an encounter. But then the question is how do you determine if your meeting is with AI or with a human, when AI becomes the prevailing content creator.

Algorithm tells you what to read

The point is algorithms are now part of the infrastructure through which people access information. A traditional library gives the reader shelves. A search engine provides a ranked list. A recommendation algorithm offers a personalized stream. A generative AI system goes a step further: it delivers interpretation without opening the source. An algorithm doesn't need to dictate thoughts. It only needs to influence what appears first.

If a book is recommended 10 times and another just once, the platform has not banned the second. If one interpretation tops every search while another is buried several pages down, no one has prohibited the second interpretation. Yet the user's intellectual environment has changed. AI intensifies this by merging recommendation and interpretation. Instead of merely saying, "You might like this book", the model may say, "Here is what this book means". That is a much deeper form of mediation.

The new educational inequality

AI can bridge gaps, but also create new divides. One obvious divide is access: who has a computer, reliable connection and AI subscription. A subtler divide is competence: who can use AI critically.

Orhan Agirdag, a professor in education at KU Leuven, introduces "prompting literacy" to describe this new capability. Drawing on French sociologist Pierre Bourdieu's concept of linguistic capital, Agirdag argues that differences in prompting proficiency may replicate educational inequalities. It's not simply about writing better prompts; it is about interacting effectively and critically with AI as a socially distributed skill.

This creates a curious hierarchy. One student uses AI as a debating partner, independently writing an essay before asking the model to challenge its assumptions, identify gaps and construct an opposing interpretation. Another simply types: "Write my essay. "Both have AI, but not the same education.

This distinction matters because AI can make the second approach appear successful. The outcome may be polished and plausible, earning the student good grades. But the intellectual process education should cultivate is outsourced. AI literacy must mean more than operating a chatbot. It must include knowing when not to use one.

Who owns the library of the future

That brings us back to Project Panama. The most revealing aspect of the episode is not that Anthropic wants books but that it was seeking high-quality text produced by humans. As AI-generated material spreads across the internet, training on the outputs of previous models risks recursive contamination. Human-authored books therefore become valuable as relatively stable reservoirs of language, argument and knowledge. Some people call text written after 2022"AI-polluted text".

That changes the paradigm of the library. For centuries, libraries collected books so that people could read them. In the AI economy, books become inputs for developing proprietary systems.

The legal question is narrower than the cultural one. In a United States copyright litigation, a judge ruled that Anthropic's training of models on lawfully acquired books could be fair use, while the acquisition of allegedly pirated books was a separate issue. Anthropic subsequently settled for $1.5 billion over pirated works.

Legality is not cultural wisdom. A book can be legally destroyed but still be culturally valuable. A physical book doesn't just contain words. It has provenance, marginalia, typography, illustrations, ownership marks and a history of readers. Most books are not rare, but scholars cannot predict which forgotten edition, obscure pamphlet or marginal note will become significant.

Digitalization preserves information but changes the access control. When millions of books feed proprietary models, the public library undergoes a remarkable transformation. The public may own the shelves, but a private company owns the intelligence trained on their contents.

Educating or informing

AI's educational promise meets its deepest risk here. The technology that brings world-class lessons to a child in a remote school can also mediate that child's understanding of the world. The algorithm that helps discover forgotten books can influence which forgotten books are "rediscovered" and which visible books disappear into obscurity. The AI tutor that patiently explains a difficult concept can also become the default authority.

In a 2024 paper coauthored by Huettig and Morten Christiansen, a professor of psychology at Cornell University, the authors argue that large language models may exacerbate declining literacy but can also support it when used critically and when educational systems encourage productive interaction rather than passive consumption. They conclude that the future of literacy is too important to be left entirely to tech companies.

That last point is crucial. The question is who shapes the relationship between AI and human intelligence. If AI becomes the intermediary between people and knowledge, then decisions about training data, retrieval, recommendation, ranking, moderation and answer generation shape the architecture of knowledge. The old librarian pointed toward the shelves; now algorithms decide what to take from the shelves. And it does not need to screen explicitly to exercise that power. Ranking is enough.

Keeping the reader in the room

None of this makes a case for keeping AI out of education. On the contrary, iFLYTEK's example demonstrates why AI has a legitimate educational role. It expands access to teachers, reduces routine work, analyzes learning patterns and helps students in places where educational resources are limited.

The virtual scientist offers perhaps the best model: AI should not always supply answers; it should push students to think harder. The same principle should govern reading. AI can clarify difficult texts without replacing them, challenge a student's argument rather than compose it, recommend sources without becoming the source, and expose competing interpretations instead of flattening them into one polished answer.

Students ought to read a difficult chapter before asking for a summary; the AI, in turn, ought to solicit their interpretation before offering its own. They can then compare its explanation with the original text and identify what has shifted. Teachers, likewise, can require learners to confront contradictory arguments rather than let the machine harmonize them. The best educational AI may thus be the one that knows when not to make learning effortless.

That principle also offers a way to understand Project Panama. The deepest concern is what happens when vast cultural resources become inputs into privately controlled systems that influence how the rest of society encounters knowledge.

For centuries, knowledge was produced through a long chain of human activities: reading, writing, arguing, teaching, editing, publishing, reviewing and remembering. AI is now entering every link in that chain. It can read, summarize, recommend, interpret and even write a book based on what it has absorbed from millions of others.

And algorithms decide what knowledge millions of people are most likely to encounter. At that point, the issue is who controls the algorithms between human beings and knowledge as their training data, ranking mechanisms and interpretive priorities are not necessarily transparent.

That is the real risk of AI controlling knowledge production. The danger is not a robot sitting in a library and forbidding people to open books. It is something quieter. People may continue to have access to every book in the world — while increasingly encountering those books through summaries, recommendations and interpretations through purposely designed filters.

Nothing is banned. Nothing is burned. Nothing is formally screened. But the route from the original text to the human mind has acquired a new gatekeeper. Project Panama offers a powerful metaphor for this transformation. Physical books are cut apart so that machines can absorb their contents more efficiently. The next question is whether the human reader will gradually be cut out of the process in the same way.

AI should help us read more, not read instead of us. It should help us ask better questions, not eliminate the need for questions. It should expand access to knowledge without becoming the final authority over what knowledge means.

The original promise of mass literacy was to put knowledge in the hands of ordinary people. The AI age could fulfill that promise on an extraordinary scale — or quietly reverse it, if people become dependent on machines to decide what knowledge matters, what it means and what deserves their attention.

The answer should be to ensure that humans continue to read, question, compare and disagree.

liyang@chinadaily.com.cn