Talk to your Book: why reading AI should use your notes, not the open web

Asking a general chatbot about a book you just finished feels magical—until the answer cites a chapter that doesn’t exist or summarizes a popular take that isn’t yours. Large language models optimize for plausible prose. They do not know which sentences you underlined unless you give them that context.

Talk to your Book flips the contract: the chat’s job is to help you think with your captures—quotes, OCR pages, and comments—not to replace the book with a web essay. That design choice is about trust as much as features.

The failure modes of ungrounded reading AI

• Hallucinated citations — fake page numbers and tidy “key ideas” you never marked
• Popular-summary bias — answers drift toward what the internet says about the title
• No personal emphasis — the model can’t know which argument mattered to you
• False confidence — fluent wrongness is worse than a short “not in your notes”

What grounded chat needs from you

Context quality is capture quality. If your library is empty, no product can invent faithful answers. Aim for notes that are:

• Specific — a claim, definition, or example, not “chapter was good”
• Located — book + chapter so the UI can keep you oriented
• Commented — one line on why you saved it
• Selective — enough signal that retrieval has something real to find

Good questions to ask

1. “What did I capture about X in this book?”
2. “Contrast the notes under chapter 3 vs chapter 7 on Y.”
3. “Quiz me on the definitions I saved for tomorrow’s seminar.”
4. “Where did I write that the author’s claim depends on Z?”

Honest limits

Grounded chat cannot recover an argument you never captured. It should not claim DRM storefront highlights it never saw. Cross-book chat on AI Max is powerful for synthesizing your library—and still bounded by what exists on your shelf. Treat fluent answers that ignore your notes as a failure mode, not a feature.

Bottom line

Reading AI earns trust when it stays inside your highlights. Capture deliberately, ask concrete questions, and keep the chrome quiet enough that the page—not the chatbot skin—stays the point.