What AI Still Gets Wrong About Books

AI gets books wrong in five reliable ways: it invents quotations, merges editions and adaptations into one remembered blur, leaks endings while trying not to, defaults to generic literature-class readings, and delivers all of it with unearned confidence. Knowing the failure modes is half the defense; knowing which are fixed by architecture — and which only by your own judgment — is the other half.

We build AI reading software, so this is partly a list of our own homework. Here's each failure, why it happens, and what actually fixes it.

Invented quotes and merged details

Ask a general model for a line from a novel and you'll often get something book-flavored: right voice, wrong words — or a film adaptation's line wearing the novel's citation. The cause is architectural: the model holds a compression of many texts about the book, not the book. The fix is retrieval — answering from the actual pages makes quotes checkable and editions singular. Any tool discussing a specific book without its text in hand is reciting from a blur.

The ending leak

Told "don't spoil past chapter 4," a model must reconstruct chapter-4 knowledge by inference from whole-book memory — and inference fails invisibly: emphasis shaped by later events, characters weighted by their endings. This one has no prompt-level fix; only a structural boundary — retrieval physically limited to your position — makes the leak impossible. It's the failure mode our whole product thesis hangs on.

The generic reading

Ask what a book "means" and general models converge on the same safe essay: themes of identity, critique of society, ambiguity as richness. It's not wrong; it's uninformative — the mean of every study guide ever written. The fix here is only partly architectural: grounding helps (a companion citing your pages must engage specifics), but the rest is prompting craft — pressure questions ("weakest claim," "strongest objection," "what would X say") force positions where "analyze this" invites mush.

False confidence, and what stays your job

All the above arrives fluent and sure — the register of expertise without its accountability. Grounding fixes confidence about the text (claims come with receipts); nothing fixes confidence about interpretation, because interpretive judgment isn't a fact to retrieve. The stable division of labor: trust a grounded companion on what the book says, treat what the book means as an argument between you — and treat any tool that blurs that line, in either direction, as unfinished.

Frequently asked questions

Why does AI make up quotes from books?

General models store a compression of everything written about a book — editions, summaries, adaptations, discussions — not the text itself, so "quotes" are reconstructions in the book's style. Retrieval-grounded tools fix this by answering from the actual pages, making every quote checkable.

Can AI hallucinations about books be fixed?

Four of the five main failures — fake quotes, merged editions, ending leaks, ungrounded claims — are fixed by architecture: retrieval from the real text plus a hard position boundary. The fifth, generic interpretation, is only partly fixable; pressure questions and your own judgment do the rest.

How do I know if an AI reading tool is grounded?

Two tests: ask for a quote you can verify against your copy (grounded tools quote exactly), and ask how the book ends mid-read (grounded, bounded tools refuse). Passing both means answers come from your pages; failing either means memory is in the loop.