If AI Can Summarise Every Book, Why Read One Yourself?

Geoffrey Chen

How a Person Builds a World · Season One: “Reading, Memory and Judgment” · Article Eleven

For a 400-page book, AI can produce ten key points, a chapter outline, character relations and a central argument within seconds. The reader can continue: “What is the author really saying?”, “Which ideas matter most?”, “How does this differ from another thinker?” If the answers are accurate enough, does spending many hours with the book amount to an inefficient attachment?

“AI has no soul” is not an adequate reply. Summaries have always existed. Contents pages, reviews, lectures and other people’s accounts help us decide quickly. AI changes the cost, speed and degree of personalisation. It can recompress the same work around each new question and generate indefinitely many versions.

The important issue is not whether summaries have value. It is which part of reading they can replace, and which part they cannot.

A summary can solve a selection problem without solving understanding

There are far more books than an individual life can read. A summary is first a navigation device. It can help decide whether a work is relevant, locate chapters for closer attention, explain background terms and provide a structural check after reading. Refusing every summary does not create depth. It can simply waste finite time.

A map, however, is not the same knowledge as travelling a route. A summary retains what it treats as important and deletes repetition, pacing, examples, hesitation and local deviation. For an operating instruction, such compression may be the point. In philosophy, history and literature, what is deleted may be part of how judgement occurs.

The ending of a novel can be stated in one sentence. A reader’s judgement of its characters forms through the order of disclosure, the creation of sympathy and the destruction of early impressions. A philosophical argument is also more than a set of conclusions. Why an objection is handled first, where a concept narrows, and which example carries the transition may determine how the conclusion should be limited.

A summary provides the terrain after arrival. Reading lets us experience why the route arrived there.

Summarising a long work remains technically difficult

The fluency of generative AI can conceal the difficulty of the task. Books contain causal relations across chapters, long character development, recurring concepts and distant foreshadowing. If a system divides the work into pieces and repeatedly merges intermediate summaries, a relation lost during early compression may be impossible to recover later.

The researchers who created the BookSum dataset noted that long-form narrative differs from short-document summarisation through long-range causal and temporal dependencies. See the BookSum paper. Later work on BooookScore compared workflows such as hierarchical merging and incremental updating and emphasised the difficulty of evaluation, including contamination of established book datasets in model training.

Evaluation is itself a bottleneck. A summary may be coherent while omitting a decisive qualification. It may cover the principal events and incorrectly connect two causes. Research on long-document factual consistency has developed measures specifically for cross-span verification because short-summary metrics do not transfer cleanly. LongDocFACTScore is one such effort.

“The AI had access to the whole book” is therefore not a sufficient guarantee. Accessing, compressing, preserving relations and generating a reliable account are different capabilities.

The greatest risk may not be an obvious error

If a system changes a character’s name, the reader may notice. Scope changes are harder to detect. A source reports a result for a particular sample, condition or dosage; the summary converts it into a general conclusion. An author proposes one possible explanation; the summary presents it as the claim. A work preserves conflict; a summary chooses one side in order to become clear.

A 2025 study in Royal Society Open Science compared summaries generated by ten large language models with source scientific material and found a tendency towards broader generalisation in multiple models. Requests for accuracy did not uniformly remove the bias. See “Generalization Bias in Large Language Model Summarization of Scientific Research”. The finding concerns particular models, prompts and scientific texts and should not be converted into a claim about every book summary. It identifies a particularly dangerous form of compression: a summary sounds clearer because the source’s qualifications have been smoothed away.

Research on LLM summarisation of medical evidence has likewise shown why automatic metrics can fail to capture harmfulness, factual inconsistency and human preference. An omitted or distorted attribute can expand a conclusion beyond its evidence.

In literature, the risk may be interpretive closure rather than factual error. The system must provide a “central theme”, while a work’s value may lie in themes that never reconcile. The more confident the summary, the easier it is to mistake one defensible interpretation for the work itself.

Personal reading trains the ability to decide what matters

The least visible cost of a ready-made summary is that it performs the first judgement of importance for the reader. The material has already been selected and ranked. Even if the summary is entirely accurate, the reader has not had to decide which paragraph deserves pause, which example changes the claim or where rereading is necessary.

That selection is not incidental labour. It is part of reading competence. While engaging with a complex text, the reader continually updates hypotheses: What do I think the author is trying to establish? Does this evidence support or break my account? Is a term changing? A summary turns this dynamic process into a product.

Writing one’s own summary is also different from receiving one. Experimental research found that generating a summary after a delay could improve the accuracy of students’ judgements about their own comprehension. The act required reconstruction rather than mere familiarity. See “Summarizing Can Improve Metacomprehension Accuracy”. The study does not imply that everybody must summarise everything. It shows that part of a summary’s educational value can lie in making it.

When AI supplies the product immediately, the labour saved may be the labour through which judgement would have formed.

Reading gives a text a chance to resist the reader

Questions posed to AI normally begin from the user’s existing interests. The system extracts relevant material and makes the book easier to fit into an established frame. The important work of a book can be to ask a question the reader was not prepared to pose.

If somebody asks only how a work supports an existing research theme, a chapter that appears irrelevant may never be seen—even if it would have changed the direction of the research. Slowness, detour and discomfort allow a text not to obey the reader’s current objective.

This does not require reading every book word by word. Different purposes warrant different depths:

  • To assess relevance, begin with a reliable human or AI summary.
  • To find a fact, locate the source and inspect its context.
  • To understand an argument, read the sections that carry its reasoning.
  • To evaluate literature, experience its time, voice and structure.
  • To quote publicly and accept responsibility, return to verifiable text.

The useful design is not summary versus original. It is a summary that continues to point into the original rather than closing it.

AI is best used as scaffolding around reading

Before reading, AI can explain background and terminology. During reading, it can compare passages, ask questions and trace a character or concept. Afterwards, it can test the reader’s reconstruction. It can also create accessible entrances for readers working in an additional language.

A good system should display its basis. Which chapters support each key claim? Where do rival interpretations exist? Which statements are inference rather than source? It should make it easy to open the relevant passage, rather than producing a feeling of completion through fluent prose.

Readers can change their questions too. Instead of asking only for a summary, ask for the contradictions hardest to compress, passages supporting and resisting the same interpretation, voices likely to disappear, and verifiable locations in the source. AI then does more than reduce words. It helps the reader decide where to increase attention.

Reading forms a person capable of judgement

From a Sustenesis perspective, the meaning of a book is not a package moving from textual input to summary output. It forms among authorial arrangement, readerly time, prior experience, questioning, rereading and later use. AI can participate in those relations. It cannot give a reader the same structure without the reader undergoing any of them.

If the aim is to know the broad subject of a book, a reliable summary may be enough. If the aim is to decide whether to continue, AI can save substantial time. But when a book’s value lies in changing a question, training attention, exposing contradiction or forming durable judgement, summary cannot replace reading, because what is removed is not merely information. It is the process by which the reader changes.

We should continue to read books ourselves not to prove that humans are more diligent than machines, and not because every work deserves complete reading. The reason is simpler. Some knowledge can be reported to us; some judgement forms only while language continues to resist what we expected to find.

AI can tell a person where a route leads. Reading determines whether, after travelling it, the person has acquired an ability to recognise roads.

Primary sources and further reading

Continue reading: Explore the How a Person Builds a World series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.