Apple TV's Search Bar Now Learns What You Like to Watch

Key points

• Apple ML Research published a system that personalises Apple TV search results after every single keystroke, even with just one or two characters typed.
• The system combines two AI models: one that reads the meaning of words across languages, and one that learns from what millions of users actually watch.
• It works across multiple languages, making it useful for Apple TV's global audience.
• Apple hasn't announced a release date, so it may already be running quietly or still moving through testing.

You type the letter "S" into the Apple TV search bar. The app has maybe half a second to guess whether you want Severance, Slow Horses, or that Spanish thriller you watched last month. It's a genuinely hard problem, and Apple ML Research has published details of how they're trying to solve it.

The challenge is called incremental search, meaning the app ranks results after every single keystroke, not just when you finish typing. One or two characters give almost no clue about intent, so the system needs extra help from what it already knows about you.

Apple's approach pairs two AI models.

The first, TextEmb, is a multilingual text encoder, essentially software trained to understand the meaning of titles across different languages. It was fine-tuned on data about which shows people tend to watch together: if fans of one thriller often watch another, the model learns those titles are related, even if the words look nothing alike.

The second, IdEmb, skips words entirely. It treats every show and film as a unique ID number and learns patterns purely from clicks and viewing behaviour. Think of it as the "people who watched this also watched" logic baked into search ranking itself.

When you search, the system builds a quick profile from your recent viewing history and blends both signals to reorder results.

This is Apple's second published search overhaul in as many days. On 14 July we reported that Apple rewrote its Music search engine to handle misspelled and phonetically typed queries. The TV system feels like the same thinking applied to a harder version of the problem, where letter-level ambiguity is even more acute.

Should you worry about the privacy trade-off?

Yes, to a degree, though Apple has long said it processes personalisation data on-device where possible. The published research doesn't spell out exactly where this computation happens. If on-device privacy matters to you, it's worth checking Apple's privacy settings under your Apple TV account. The paper focuses on ranking quality, not data handling.

The practical upside is real regardless. If you half-remember a title and type a three-letter guess, a personalised system has a much better shot at finding the right show than one that just matches letters. Apple TV carries a large multilingual catalogue, so cross-language training matters: someone watching content in both English and Italian should see relevant results whichever language the title is stored in.

The honest read here is that the ranking improvement is probably genuine, but the privacy question Apple leaves unanswered is the one that will actually matter to most people who care about it.