ChatGPT Spoiled My Book, So I Built an AI That Can’t

I was partway through a book and asked ChatGPT a question about it. The answer was helpful, and it also told me how the book ends.

That is the whole origin story of BookBuddy. Not a clever idea, just a question I wanted answered without losing the rest of the book.

Why a chatbot spoils books

A general chatbot has no idea where you are in a book. It does not have your copy, and it does not know your page. When you ask about a character, it answers from everything it picked up about that book: summaries, reviews, discussions, the ending included. Nothing in that answer knows that you are on page 140.

You can try to fix this with the prompt. "I'm on chapter 12, no spoilers please." Sometimes it works. But the chatbot still has to guess where chapter 12 ends in your edition, and it still has to hold back what it knows while writing a helpful answer. The better it knows the book, the more there is to hold back.

The problem is not that the model is careless. It is that the model is answering from the whole book, and you have only read part of it.

An AI that only reads what you have read

So BookBuddy does not start from the model's memory of the book. It starts from your copy.

• The book is in the app. You upload an EPUB, or connect a tolino bookshop account and the DRM-free books in your library come across. You read in BookBuddy's own reader, so it always knows your page. With tolino, your reading position syncs too.
• Every passage it reads ends before your page. When the AI looks something up, it searches, quotes and summarizes only the text before your bookmark. That limit is in the code that fetches the text, not a request in the prompt.
• It is told not to fill the gaps from memory. A model may still know the book from its training. Its highest-priority rule is never to reveal, hint at or infer anything beyond the text it was given, even if it knows it from elsewhere.
• Ask about the ending and it says no. Then it offers what it can talk about instead: what has happened so far, who someone is, what a passage means.

Quotes in an answer are meant to be the book's exact words, and tapping one opens the book at that passage, so you can check.

What it cannot promise

The limit on the book text is hard. The rule about the model's own memory is an instruction, and a language model follows instructions well, not perfectly. So every answer has a thumbs down, and "spoiler" is one of the reasons you can pick. If it ever happens to you, that is the button.

Why this is the whole product

Most AI book tools summarize the whole book. That is useful after you have finished it. While you are still reading, the most useful thing an AI can do is answer your question and stop at your page.

That is the one thing BookBuddy is built around: a reading companion that is always on the same page as you.