Streaming Apps Now Use AI to Fix Their Own Bad Recommendations

Every time you open a streaming app, a 'recommender system' quietly decides what to show you. When it gets it wrong, you scroll forever and give up. Until now, the people fixing these systems mostly relied on averages — like 'this show was clicked 4% of the time' — which tell you something is broken but not why, or for whom. A new paper describes an AI system that digs into the details instead.

The system is called AURA. It uses AI agents — software that can read, reason and take actions on its own — to comb through millions of real viewing sessions. The agents flag patterns: maybe the app keeps pushing sequels to people who never finish them, or buries a genre that a certain group loves. Then, armed with the company's own code and training data, the agents propose and write fixes directly into the recommendation software.

Researchers tested it on production data from two large consumer platforms at a major media-streaming company (the paper doesn't name it). The authors say the same approach should transfer to online shopping, where 'you might also like' recommendations drive sales. The bigger idea: instead of humans guessing at improvements, the system diagnoses itself and repairs itself, continuously.

The catch: the results are early and reported by the company building the tool, with no outside verification yet. And there's a risk in letting AI rewrite the code that decides what billions of people see — safeguards are mentioned but not detailed. Also, optimizing for clicks mostly means more of what you already watch, not necessarily better TV.