Why do AI book recaps get things wrong, and how do page-locked tools reduce it?

01The short answer

Every company that ships an AI recap says it can be wrong. OpenAI says ChatGPT "can produce incorrect or misleading outputs". Anthropic says Claude "can write things that might look correct but are very mistaken". Google's Play Books help page says "Generative AI can make mistakes", and its launch post adds that Book insights "may make mistakes". These are not boilerplate. They describe how the technology works.

02Six reasons AI recaps go wrong

1. Memory instead of the text

A general chatbot answering "what happened in chapter 9?" does not read chapter 9. It writes from patterns it learned from summaries, reviews and discussions of the book. Popular books are discussed as a whole, so its memory of chapter 9 is mixed with chapters 10 to 40. The result can be a scene that is roughly right but in the wrong order, or a detail that belongs to a later chapter.

2. The wrong place

Edition differences move page numbers. Some books restart chapter numbering in each part. Audiobook tracks rarely match chapters. If the tool guesses your place, everything after the guess is at risk. This is why page-locked tools put so much effort into finding your place: Kindle, Google Play Books and Spotify use your position inside their own apps, and outside-the-store tools ask for evidence, such as a photo of the page or the last phrase you heard.

3. Web pages written by finishers

When a chatbot searches the web, the pages it finds for a novel are mostly plot summaries, fan wikis and "ending explained" articles. Perplexity searches the web for every answer by design, and OpenAI says ChatGPT's search is on for all models by default. Those pages are written for people who have finished the book, so they carry both spoilers and other people's interpretations.

4. Blending books, sequels and adaptations

Series books share characters and places, and film or TV adaptations change scenes. A model can merge them: a character's later title appearing too early, a scene from the adaptation presented as if it were in the book, or a relationship from book three showing up in a recap of book one.

5. Compression

A recap turns hundreds of pages into a few paragraphs. Something has to be dropped, and the AI decides what. It can drop the scene you needed, or merge two conversations into one. Shorter is not safer: the more a recap compresses, the more it relies on the model's judgment of what mattered.

6. Names and confident tone

Large casts, similar names and characters who use nicknames or titles are hard for any reader and for any model. And a wrong recap reads exactly like a right one. OpenAI notes ChatGPT can sound confident even when it is wrong.

03What does a page lock fix?

A page lock means the software only uses the book's text up to your position. Google describes Book insights as "designed to give spoiler-free answers by only referencing the text up to your current reading position". Spotify says Recaps "only cover the portion of the book you've already heard". Kindle's Ask this Book answers up to your current reading position by default. Magic Bookmark says its recap "stops exactly where you are" and its cast list is "built only from what you've actually read".

That addresses causes 1, 3 and most of 4: the tool works from the actual text rather than memory of the whole book, it does not need web pages, and later books in a series are out of reach.

04What does a page lock not fix?

• Finding your place (cause 2). In-store tools know your position exactly. Outside tools depend on the photo, page number or phrase you give them; a blurry photo or a different edition can still put the line in the wrong place.
• Compression (cause 5). The recap is still written by an AI that chooses what to keep.
• Names (cause 6). A model can still mix up two characters within the pages you have read.
• Coverage. Page-locked tools only work for books they support. Kindle covers thousands of popular fiction titles, Google says more than a million English-language ebooks, Spotify a limited selection of audiobooks, and Magic Bookmark a catalog of supported titles led by romance, fantasy, romantasy, YA, sci-fi and historical fiction.

So a page lock lowers the risk of spoilers and of invented later events. It does not make every sentence right.

05How do you protect yourself?

1. Prefer a page-locked tool when spoilers matter: inside your store's app, or from your own copy with a tool such as Magic Bookmark.
2. Give your place two ways when using a chatbot: chapter number and chapter title or last scene.
3. Treat unfamiliar names as a stop sign. They are either errors or spoilers.
4. Check one detail against the book. If one claim is wrong, distrust the rest.
5. Report errors with the tool's feedback buttons; Google and others say feedback helps.

Step-by-step checks are in How to check an AI recap, and how the lock works is drawn in How page-locked recaps work.

06Did we measure error rates?

No. We have not run accuracy tests, and we do not repeat error rates from sources we cannot verify. This page explains why errors happen, based on how the makers describe their own tools.