How do you check whether an AI book recap is right?

01Why do AI recaps need checking?

Every recap on the tools we cover is written by a language model. OpenAI's help center says ChatGPT "can produce incorrect or misleading outputs" and lists "Fabricated quotes" and "Overconfident answers" among the problems. Anthropic says Claude "can write things that might look correct but are very mistaken" and that users "should not rely on Claude as a singular source of truth". A page lock limits which pages the model may use; it does not check the model's work.

02Six checks that take a few minutes

1. Confirm the place. Before reading the recap, check where the tool thinks you are: the page number, the chapter, or the phrase it matched. A wrong bookmark is the fastest route to a spoiler. In a page-locked app, move the bookmark and try again.
2. Spot-check one name. Pick a minor character the recap mentions and find them in the pages you have read. If you cannot, treat the name as a possible error or a spoiler.
3. Spot-check one event. Pick a specific claim ("she left the city at night") and flip back to the scene. You only need to confirm it happened, not reread.
4. Check relationships. Who is related to whom is where recaps slip most easily, because it is spread across many pages. Compare with a cast list if your tool has one.
5. Watch the tense. Lines that look forward ("this will matter later", "little does he know") are a sign the text knows more than you do.
6. Ask for the location. Ask the tool where in the book something happened. A page-locked tool should be able to place it before your page; a vague answer is a reason to check.

03What are the red flags?

• A character or place you do not remember at all.
• A scene that sounds like a film or TV adaptation rather than the book.
• An event from a later book in the series.
• Very confident wording about motives the book has left open.
• Exact quotes you cannot find in the text.

04What should you do when you find an error?

• Stop reading that answer if it might be a spoiler.
• Check your bookmark and regenerate if your tool allows it.
• Narrow the question. Ask about one chapter or one character at a time.
• Report it through the app's feedback option, if it has one. Anthropic's help center, for example, points to the thumbs-down button for unhelpful answers.
• Fall back to the book. Skim the last chapter you read, or your own notes.

05A worked example of the checks (no real book)

Here is how the six checks play out on an invented recap of an invented novel, so no real book is spoiled.

You are at the end of chapter 14 of a fantasy novel. You ask for a recap and get:

"Mara reaches the capital and meets the Archivist, who agrees to teach her. Her brother Tam, now a captain of the guard, warns her to leave the city. Later, the Archivist's true loyalty becomes clear."

• Check 1, names. You know Mara, the Archivist and Tam. Good.
• Check 2, titles. "Now a captain of the guard": you do not remember Tam being promoted. Either the recap is wrong or this happens later. Stop and check.
• Check 3, order. Did Tam's warning come before or after the Archivist agreed to teach her? Flip back to check one scene.
• Check 4, hedging words. "Later, the Archivist's true loyalty becomes clear" is foreshadowing. That sentence is a spoiler in disguise, or an invention.
• Check 5, one detail against the book. Find the scene where Mara meets the Archivist and confirm the agreement.
• Check 6, source. If this came from a chatbot with web search on, assume it read wikis and ending explainers. If it came from a page-locked tool, check that your page or position was set correctly.

Result: two warning signs in three sentences. You would treat the recap as unreliable from "now a captain" onward.

06Which tools make checking easier?

• Tools that show what they used. Perplexity cites the web pages behind an answer, which tells you when it read an "ending explained" page. Merrilin says its answers include source citations to the text.
• Tools with feedback buttons. Google Play Books lets you rate Book insights answers with thumbs up or down and send detailed feedback.
• Tools that know your place. Kindle, Google Play Books and Spotify use your position in their own apps; Magic Bookmark asks for a page photo, page number or last phrase heard. Fewer guesses about your place means fewer chances to drift past it.

Why recaps go wrong in the first place is explained in Why AI book recaps get things wrong.

07Keep a two-line note as your own check

The fastest way to check any recap is to compare it with something you wrote yourself. When you stop reading, jot two lines: the last thing that happened and the last new name you met. When you come back, read your note first, then the recap. If the recap disagrees with your note, trust your note and look at the page. It takes ten seconds and catches most wrong-place and wrong-name errors.

08Is a page-locked recap more reliable than a chatbot's?

On spoilers, the design is safer: a page lock stops the model from using later pages, while a chatbot's limit is only your instruction. On accuracy, no maker we cover publishes accuracy figures, and we did not test any. Both kinds of recap should get the same quick checks. See Can ChatGPT recap a book without spoilers? and the AI reading companion checklist.