Using ChatGPT (and other AI) to choose what to read: prompts, mistakes, and how to verify

"What book do you recommend if I liked X?" is one of the most common questions in ChatGPT, Gemini, Claude, and similar copilots. The answer arrives in seconds, with a confident tone and sometimes a convincing synopsis. The problem is not speed: the model does not "look up" a bookstore in real time like a human at the counter. It generates probable text. In reading, that becomes titles that sound real, mixed-up authors, editions that do not exist, or translations credited to the wrong publisher. Using AI to choose what to read can help if you treat the reply as a draft idea list, not a confirmed order. This guide covers prompts that work, typical mistakes, how to verify before spending money, and how a recommender tied to verified records, like Librarius on mobile, differs.

Language models have seen millions of reviews, prize lists, and book conversations. That lets them steer you by genre, tone, era, or theme with reasonable accuracy when the question is open: "short novels with dark humor," "accessible essay on memory," "Latin American women authors from the last fifteen years." They can also summarize why two authors feel similar in style, suggest reading order within a known series, or adjust if you say you avoid graphic violence, tragic endings, or dense prose. In that mode, AI works like a conversational bookseller: it expands your search vocabulary and breaks the loop of repeating the same three names.

Limits show up in verifiable detail. Asking "Is there a Spanish edition of this title with this ISBN?" may return a made-up number or a plausible but false publication date. Asking "What is coming out this month in Spain?" mixes real new releases with hallucinated titles if the model lacks reliable catalog access. For reading habits, AI shines at conceptual exploration; it fails when you need bibliographic certainty. So split two moments: brainstorming in chat and checking against a primary source (bookstore, library, publisher catalog, or a work page in an app that only shows materialized books).

A good reading prompt does not ask for "the best book in the world." Describe real constraints: time available, format (print, ebook, audiobook), language, whether you read on transit or in weekend blocks, and what worked or failed recently. The more concrete the context, the less generic the list. Useful example: "I want a novel under 300 pages, hopeful but not naive tone, set in contemporary Europe; I am tired of thrillers; I read and liked X for its pace." Asking for three to five options with one line of "why it fits" forces justification, not just a list.

• Include two or three books you read and what you liked (pace, voice, theme), not just bare titles.
• Say what you want to avoid: hard spoilers, explicit romance, academic footnotes, second person.
• Ask for comparisons like "if you liked A, try B for tone, not plot."
• Limit the result: "authors with Spanish translation only" or "works before 2010" if you want accessible classics.
• Ask follow-up questions: "ask me three questions before recommending" to refine without a random long list.

Another trick is to assign roles: "act like a neighborhood bookseller who barely knows bestsellers" or "like a professor recommending rigorous popular science." That shifts model bias. Iterating in the same thread also helps: "from that list, drop anything that is part of a long series" or "more women authors writing in Spanish." If you use AI on your phone, save a screenshot or note of the final list; chats get lost and repeating the prompt is tiring. To log what you actually read and avoid duplicate suggestions, track your reading progress helps anchor states without turning reading into productivity theater.

Typical mistakes: invented books and false data

Track what you read

Create a free account

Free, no subscription

AI "hallucinations" are confident answers not grounded in facts. With books they show up often: title plus author who never published together, fake prizes, a "sequel" to a standalone novel, invented ISBN, Amazon link that does not match. Vendors document that models can invent quotes and references; in literary recommendations the symptom is subtler because many titles sound credible. A generic name ("The last winter of…") with a common surname is especially risky: you search online and get different works, none matching what the chat described.

Another mistake is confusing a namesake author or a different translation: the model recommends the English edition when you only read in Portuguese, or attributes a novel to the wrong homonym. "List bias" also happens: the same ten BookTok hits repeat even when you asked for something niche. If the assistant has web browsing, recent releases improve, but risk remains: it may misread a press note or mix review with publisher copy. Practical rule: any title you will buy must exist in at least two independent sources (bookstore, national catalog, Google Books with a clear edition, or a page in a reading app). If only the chat "confirms" it, treat it as rumor.

Verification is a short ritual that saves money and frustration. Step one: search title plus author at an online store in your country or the public library catalog. No results means stop. Step two: check that the synopsis matches what the chat told you; sometimes the model blended two works. Step three: check length and format; a "short book" that turns out to be volume one of four kills the weekend plan. Step four: if translation matters, confirm translator and publisher on the record, not in the chat summary. Step five: read a recent human review at the store or use writing book reviews as a tone reference, not absolute truth.

1. Search store or library with exact title and author surname.
2. Match cover, year, and page count to what was promised.
3. Check availability in your language and format (audio included).
4. Check price and lead time; if it is import, add weeks of wait.
5. Decide: buy, reserve, add to to-read, or drop.

If AI gave five titles, you do not need to verify the four you already know. Focus on the unknowns. For clubs, also verify enough copies on loan or a price that does not exclude half the group. When the goal is to swap instead of buy, verification includes used copy condition; that is another flow, but the "it really exists" rule stays the same.

A reader friend knows your taste and lends the copy: high trust, low scale. A store algorithm ("customers who bought X") optimizes sales, not necessarily long-term satisfaction. A general AI assistant optimizes fluent answers. None alone guarantees the book exists, is available in your language, or fits your current pace. Academic recommender systems distinguish collaborative filtering (patterns from many users) and content-based filtering (work metadata); reading apps usually mix both with signals from your history. The ethical and practical difference is whether the system can recommend something it cannot show you afterward.

ChatGPT does not owe you a book page: it owes you a continuation of text. So the fair comparison is not "AI vs human" in the abstract, but "free text vs verified catalog." When you want discovery inside a universe where every suggestion opens cover, author, and reading state, you need a product that couples search, ranking, and book UI. That is where recommender systems inside reading apps matter, not the generic browser chat.

In Librarius, the "What to read now?" AI recommender lives in the iOS and Android app. The web route `/recomendador-ai` redirects to download the app with that deep link; it is not an open web chat. You write in natural language what you want (between eight and five hundred characters): mood, genre, author, tone. The system does not reply with endless prose: it finds real candidates, ranks them, and returns up to three books with pages in the app.

The pipeline, as implemented on the server, works like this. First, from your text it derives bounded search queries (text, author, title). Second, it collects candidates from the Librarius local catalog and external sources (Google Books and ISBNdb) already materialized as usable rows, not loose names. Third, a language model gets explicit instructions: it may only choose identifiers from that candidate list; never invent title or ID; prioritize fit with your intent using category and description, not a single word in the title; avoid repeating titles from your recent history unless you ask. Fourth, chosen works are enriched with cover and record data so you can add them to your library or keep reading. If there are no reasonable matches, the screen says suggestions are unclear and invites you to rephrase, instead of filling with inventions.

That does not make the recommender a literary oracle: ranking is still AI over the candidate subset found at that moment, with quota and response time limits. But it changes the contract versus ChatGPT: each result should open a work the app can display. In the same family is the taste-based recommender (`/recomendador-gusto` on web, also via app): it uses your favorites and ratings, explores the catalog with similar rules, and proposes up to ten titles, refreshing when your taste signal changes. One answers a specific text question; the other answers "what fits me in general." Both target verifiable discovery, not floating name lists.

• Prompt AI: best when you have a concrete need ("something short and optimistic for summer").
• Taste recommender: best when you already marked favorites and ratings and want continuity.
• General chat: best for genre maps, known series order, and questions without immediate purchase.
• Manual verification: mandatory if the recommendation came only from chat and you will buy.

A sensible routine mixes layers. Use the general assistant to widen horizons on a bored afternoon: ask for thematic maps, compare styles, prep club questions. Use store or library verification before any purchase born in chat. Use Librarius when you want the suggestion to land on a page, cover, and reading states you can share with your club or profile. If you already run collective reading, combine individual discovery with how to choose a book for the club: AI can propose, the group votes.

Avoid delegating human negotiation to AI: "Are we reading something dense this month?" is club conversation, not a prompt. Also avoid asking chat to simulate availability at your municipal library; search its catalog. If invented answers frustrate you, shrink prompt ambition ("only books that won Prize XYZ between 2010 and 2020" is more checkable than "only forgotten gems from micro presses"). Reading stays slow; AI speeds the list, not the pages.