Can Gemini, Claude or Perplexity recap a book without spoilers?

01The one thing they share

A general assistant does not open your copy of the book. It answers from what it learned in training, from anything in the conversation, and, when search is on, from web pages. "Stop at chapter 12" is an instruction it tries to follow, not a filter on the text. That is the core problem explained in Can ChatGPT recap a book without spoilers?, and it applies to every assistant on this page.

02How does each assistant build an answer?

The table describes how the products work by their makers' own descriptions. It is not the result of a test.

03Why Perplexity is the riskiest for spoilers

Perplexity's help center says that when you ask a question "it uses advanced AI to search the internet in real-time", and it shows the sources it used. That is a strength for facts and a weakness for spoilers. The pages that rank for "[book title] recap" or "[character] explained" are written by people who have finished the book or the series: plot summaries, fan wikis, review threads. An answer stitched from them can mention a later death or reveal even when you asked it to stop at chapter 12.

If you use Perplexity for a book you are in the middle of, keep questions to things a spoiler cannot ruin (the author, the publication history, a historical setting), and avoid "what happens" questions.

04Gemini and Claude: fewer web pages, same blurry line

When they are not searching, Gemini and Claude answer from what they learned. For a popular novel that usually means whole-book summaries and discussions, so the model has to separate "before chapter 12" from "after" in its own memory, and that line is blurry. They can also mix up books in a series, or a book with its film or TV adaptation.

Both makers warn about accuracy in plain words. Anthropic says Claude "can write things that might look correct but are very mistaken". Google's Gemini privacy hub says "Gemini can make mistakes" and asks you to take care. A recap that sounds confident is not a recap that is right.

05Is any assistant safer than the others?

Not in a way you can count on. The safety comes from what you give it and what you ask:

• No search is safer than search, for a book you have not finished.
• Your own notes or text keep the answer to what you have read, because the assistant works from what you paste instead of its memory. Whatever you paste is sent to the company that runs the assistant, and you should only paste what you have the right to share.
• Narrow questions ("remind me who Marcus is, as of chapter 12") are safer than broad ones ("what's going on in this book").

06A prompt you can copy for any assistant

I'm reading [title] by [author], [format or edition]. I have just finished chapter [number], "[chapter title]". Do not search the web. Recap only what happens up to the end of that chapter. Do not mention, hint at or foreshadow anything later, including twists, deaths, relationships or what happens in sequels or adaptations. If you are not sure whether something happens before or after this point, leave it out and say so. Then list the characters I have met so far, one line each.

For Perplexity, which searches by design, replace "Do not search the web" with a narrower question that does not need plot at all, or use a different tool for recaps.

07What should you check in the answer?

1. Names you do not recognise. A new name is either an error or a character you have not met yet. Stop reading the answer.
2. Events you do not remember. Check against your book before trusting them.
3. Words like "eventually", "later" or "little does she know". Foreshadowing is a spoiler in disguise.
4. Cited sources (Perplexity). If a cited page is titled "ending explained" or "full summary", assume the answer used it.

More checks are in How to check an AI recap, and safer wording is in the spoiler-safe question bank.

08When should you use a page-locked tool instead?

When the twist matters, when the book is part of a long series, or when you are coming back after a long break. Page-locked tools limit what they use to your place in software, not just in the prompt:

• In the store's own app: Kindle Story So Far and Ask this Book, Google Play Books Book insights, and Spotify Recaps for audiobooks.
• From your own copy: Magic Bookmark finds your place from a page photo, a typed page number or the last phrase you heard, for its supported titles on iPhone.

Even a page lock does not make every sentence right. It stops the tool from using later pages; it does not stop it from misreading earlier ones. Why that happens is explained in Why AI book recaps get things wrong.