A private AI on your phone is good at “more like this” recommendations. Tell it three books or films you loved and why, and it will suggest titles with a similar theme, tone or pace, explain the connection, and keep your taste to itself. Its blind spot is anything released after its training cutoff, and small models occasionally invent a title or pin a book on the wrong author, so check that a suggestion exists before you buy it.
How do streaming and store recommendations work? #
Most recommendation engines combine two ideas.
Collaborative filtering looks at people whose history resembles yours and suggests what they watched or bought next. It needs a big pool of user data, so it runs on the company’s servers.
Content-based filtering looks at the item itself: genre, cast, author, tempo, keywords. It suggests things with similar traits. This can run on a device, since it only needs the catalog and your own history.
Either way, the service’s goal isn’t only to find you something good. It’s also to keep you watching, reading or buying on that platform, and your viewing history becomes part of your profile for ads and promotions. That’s why “Because you watched…” rows lean toward what the service owns or wants to push.
Why ask a private AI for recommendations? #
- It explains itself. You can ask why it thinks you’ll like something, and argue back.
- No platform agenda. It isn’t trying to keep you on one service.
- Your taste stays private. What people read and watch can reveal health worries, religion, politics or relationships. With an on-device model, your chat never leaves the phone.
- It works anywhere. On a plane with a folder of downloaded films, you can paste the list and ask which suits your mood.
Personal LLM runs open models such as Qwen 3.5 and Gemma 4 on your phone’s own chip, with no account. After the one-time model download, the only network use is ads in the free version when you’re online.
How do you get better recommendations from an AI? #
Vague prompts get generic lists. These habits get you somewhere interesting:
- Name three to five favorites and say what you liked. “The Martian for the problem-solving and humor” is far more useful than “The Martian”.
- Say what you didn’t like. “No bleak endings” or “nothing over 500 pages” cuts the list fast.
- Ask for a mix. “Two well-known picks and three lesser-known ones” avoids the same bestsellers everyone gets.
- Ask for the reason and a caution. “For each, one line on why I’d like it and one line on why I might not.”
- Ask for a table. Title, author or director, year, and the reason. The app renders Markdown tables cleanly.
- Keep a taste chat. Pin one chat with a system prompt that lists your favorites and dislikes, and add to it over time. Every future request starts from what it knows about you.
A prompt that works well:
I loved Station Eleven, Klara and the Sun and Never Let Me Go: quiet, character-driven, a bit sad but not hopeless. I didn’t enjoy fast-paced thrillers. Suggest five novels in a table with author, year and one line each on why they fit. Include at least two I’m unlikely to have heard of.
The same works for films, TV, podcasts, music and board games. For music, describe the feel (“warm, unhurried, good for cooking”) as well as artists you like.
How do you check the AI didn’t make a title up? #
Language models generate plausible text, and a plausible book title with a plausible author is easy to generate. Small models do it more often. Protect yourself:
- Ask for the author or director and year every time. A real title with the wrong author is a common slip; a fully invented one will usually fall apart when you search it.
- Search before you buy. A quick look in your library app or store confirms it exists.
- Use the larger model for obscure picks. Qwen 3.5 9B knows noticeably more titles than the 0.8B. You can switch models mid-chat from the chat header if a list looks thin.
- Lower the temperature. The Precise preset (temperature 0.3) sticks closer to what the model knows. Creative (1.2) is fun for brainstorming but more likely to wander.
What can’t an offline AI recommend? #
| Request | Works offline? | Why |
|---|---|---|
| “More like this” for older books, films, albums | Yes | Well covered in training data |
| Classics, genres, directors’ filmographies | Yes | Stable knowledge |
| This year’s new releases | No | After the model’s training cutoff |
| What’s on your streaming service right now | No | Changes constantly; no web access |
| Current bestseller or box office lists | No | Needs live data |
| Ranking a list you paste in | Yes | You supply the options |
The fix for the last few rows is to supply the options yourself. Paste the titles on your watchlist, the new releases from your library’s page, or the films downloaded to your tablet, and ask the model to rank them for you. We explain the cutoff problem in does offline AI know current events.
Can it help with reading, not just choosing? #
Yes. A few uses readers like:
- Book club prep. Paste your notes and ask for discussion questions.
- Catching up on a series. Ask for a spoiler-free recap of earlier books in a long series. Check it against a trusted summary if details matter, since memory of plot specifics is where models slip.
- Reading in another language. Paste a paragraph and ask for a translation and the tricky words explained. Our offline translation guide has more on this.
Frequently asked questions #
Can AI recommend books based on what I like? #
Yes. Give it a few favorites and what you liked about each, plus what you don’t enjoy, and it suggests titles with reasons. The more specific you are about tone, pace and themes, the better the list.
What’s the best AI for movie recommendations? #
For recent releases and what’s streaming now, a service or AI with live web data is better, because offline models don’t know anything after their training cutoff. For older films and “more like this” suggestions with explanations, an on-device model works well and keeps your viewing taste private.
Why did the AI recommend a book that doesn’t exist? #
Language models produce plausible-sounding text and sometimes invent titles or mix up authors, especially small models. Ask for the author and year, search each suggestion before buying, and use a larger model for niche requests.
Does an offline AI know new releases? #
No. It only knows what was in its training data, which ends before the model was released. Paste a list of new titles and it can still help you choose among them.