How AI movie recommendations actually work — and when to distrust them

“AI movie recommendations” now means at least three different machines wearing the same label. Knowing which one you’re talking to explains both why the picks feel magical on some apps and why they feel like a horoscope on others. This is the honest tour — including the failure modes, because we build one of these things and the failure modes are where the real work is.

The three machines

• Collaborative filtering— “people who liked what you liked also liked…” This is Netflix-style recommendation and it’s genuinely powerful at scale, but it needs millions of users, drifts toward popularity, and can’t tell you why it picked anything.
• The mood quiz— most new “AI what-to-watch” apps. You type “something tense but not bleak,” a language model maps that to genres and vibes, and you get plausible titles. Fine for strangers; shallow for you, because it knows tonight’s mood but nothing about your history.
• The taste-profile reader — a language model that reads your actual diary (ratings, rewatches, written notes) and reasons about it in plain language. This is the newest kind, the one Kinolog is, and the rest of this piece is about what makes it work or fail.

Signals aren’t equal — good systems weigh them

A five-star rating with a two-line note is not the same evidence as a thumbs-up you tapped in a half-second onboarding quiz. The systems that feel perceptive are the ones that rank their evidence: a dated, written logoutranks a quick swipe; a “never again” is absolute; wanting to see a film (a watchlist save) is weaker than having loved one; and your graded verdicts on past picks — hit or miss — correct the machine itself. Flatten all of that into “likes” and you get the horoscope.

Your written notes are the highest-octane fuel

The genuinely new thing language models unlock: your own sentences. “The ending destroyed me — patient dread, immaculate sound design” contains more usable taste than fifty star-ratings, because it says why. A good recommender mines those recurring patterns and can even quote your words back when explaining a pick — which is also your best lever as a user: write the occasional real note and every future recommendation sharpens.

The two sins: hallucination and over-claiming

Language models invent — confidently. An AI recommender that pipes raw model output at you will eventually recommend a film that doesn’t exist, misdescribe one that does, or re-suggest something you rated last month. The fix is architectural, not hopeful: every title must be verified against a real movie database before you see it, and your library and never-again list must be enforced in code, not by politely asking the model. If an app can’t tell you it does this, assume it doesn’t.

The subtler sin is false intimacy. Three ratings in, an app announces it “knows your taste” — it doesn’t, and the picks prove it. Honest systems say so instead: early picks labeled as early reads, claims about your taste gated behind enough real data, the occasional pick that openly admits “I’ve barely seen you rate comedy — this one’s here to find out.” Distrust any recommender that’s never uncertain.

How to actually evaluate one

• Ask why.Every pick should carry a reason grounded in your history — named films, your ratings, ideally your words. “Because you like thrillers” is filler.
• Check the memory. Does it ever re-suggest what you’ve seen or sworn off? Once is a bug; twice is the architecture.
• Look for the receipts. A system confident in its picks should show its hit rate against your own verdicts — even when unflattering.
• Test the cold start. Day one, does it perform humility or fake intimacy?

That’s the standard we hold Kinolog to — TMDB-verified titles, exclusions enforced in code, reasons that cite your actual diary, and a stats page that keeps score honestly. If you try it and the picks miss, say so with one tap: correcting the machine is the fastest way to teach it.