AI That Knows Your Taste May Not Be Worth the Cost

Every time you open a streaming app, it builds a quick sketch of who you are — what you like, what you skip, what you binge at midnight. There are two main ways to build that sketch. One adds up numbers from your viewing history (cheap and fast). The other uses a large language model — the same tech behind chatbots — to write a few sentences describing your taste in plain English, then reads that text back into the system. The second approach is significantly more expensive to run.

So the researchers asked a practical question: when is that extra cost actually worth it? They built four versions of the profiling system, crossing two choices. First, numbers versus AI-written summaries. Second, whether the system treats your recent behavior separately from your long-term history — a show you binged last week versus a genre you loved two years ago. Then they tested all four on a real production dataset from a live streaming service, not a small lab experiment.

Critically, they didn't just ask whether you clicked. They also measured things like variety and novelty — whether recommendations felt fresh rather than repetitive. The takeaway: AI-written profiles are not uniformly better. Their value depends on the kind of viewer (heavy versus casual), on what quality means to the business, and on how far back the "recent" window reaches. Blanket advice to use AI profiles everywhere isn't supported by the data.

Why you should care: at the scale of millions of users, those extra AI calls cost real money, and that cost eventually shows up in subscription prices, ad load, or fewer features. The honest catch is that this is one production dataset from one service. Results may not transfer to other apps, and the paper is a research comparison, not a how-to guide for your favorite app.