Researchers: Your Own AI Should Pick Your Restaurants, Not Yelp

Every time Netflix suggests a show, Spotify builds a playlist, or Yelp ranks restaurants, a recommendation system is at work. These systems are usually described as helpful guessers: they watch what you do and pick what you'll probably like. But a new position paper from a group of researchers argues that description hides the real story. Platforms don't just guess your taste — they also decide which options you're even allowed to see, what information backs up those options, and how the choice is explained to you.

The authors say the next big problem isn't better guessing. It's control over information: who gathers it, who sees it, and who decides what gets shared. Their proposed fix is something they call Personal Agent-Mediated Recommendation, or PAMR — in plain terms, a personal AI assistant that represents you. Instead of asking one platform for ideas, your agent would search across many sources, filter and combine the results, and decide how much of your personal data to reveal along the way. The shift, they argue, is from the platform picking items for you to your agent managing evidence on your side.

To test the idea, the researchers ran a small proof-of-concept study on hard restaurant recommendation tasks using Yelp data. Under the same underlying AI model, letting the agent choose its sources and control what it disclosed produced the best balance of usefulness, traceability, exposure of private data, and cost. That's a promising early signal, though it's one experiment, not proof this works at scale.

The honest catch: this is a position paper, not a product. Nothing here is shipping, and the authors are arguing for a direction rather than announcing a tool. Real barriers remain — someone has to pay for your agent, it needs access to a lot of your personal data to be useful, and a single assistant that shops for you could become just as powerful and self-interested as the platforms it replaces.