Your Streaming App May Soon Use AI Profiles to Surprise You

Ever feel like your streaming app keeps pushing the same genre at you, over and over? That's the trade-off every recommendation system faces: play it safe with things you'll probably like, or take a risk and show you something new. A group of researchers say you don't have to pick one or the other — you can decide per person, in real time.

Here's the setup. One method is the familiar "collaborative" approach: it looks at what millions of other viewers with similar habits watched and suggests overlaps. It's cheap and reliable but tends to keep you in a rut. The other method uses a large language model — the same kind of AI behind chatbots — to write a rich, written profile of your tastes and recommend from that. It's better at finding genuinely new things, but it's more expensive to run and not worth doing for everyone. So the researchers built a gate: a lightweight sorting system that looks at basic, instantly available signals about each user and asks, "Is this someone who'd benefit from the AI profile?"

On a real dataset of movies, TV shows and sports content, that gate sent just 12.5% of users down the AI route. It allowed a small 5% dip in how relevant recommendations were, and in exchange, the amount of fresh, unexpected content in the top ten picks rose 6.5%. Simple rules of thumb and random switching did worse. Notably, the benefit came from choosing the right users, not from the AI writing profiles itself.

The catch: this is a research result on one dataset, not a product you can use today. The gains are modest, and the AI profiles cost extra computing. There's also a privacy question — an AI-written profile of your viewing habits is a detailed file about you. Whether it's worth that trade is a call the companies will make.