What recommendations do to the people who get them
Recommender systems appear on 111 recent papers and personalized recommendation on 54, almost all in AI journals; 45 recommendation papers carry a diversity, fairness or popularity-bias keyword, and welfare, consumer surplus and addiction appear in almost none. Recommendation decides what billions of people see and buy. Who gains, and by how much?
What prompted this
Since October 2025, 513 papers carry a recommender or recommendation keyword (440 in AI journals): "recommender systems" on 111, "personalized recommendation" on 54, "sequential recommendation" on 47, "multimodal recommendation" on 38. Forty-five of them also carry a filter-bubble, diversity, polarization, fairness, welfare, popularity-bias or well-being keyword, led by "popularity bias" on 19 and "recommendation diversity" on 13.
Questions
- In platform experiments, how much consumer surplus do recommendations create relative to search alone, and how is it split between consumers, sellers and the platform?
- Do recommendations that optimize engagement reduce the variety of what people consume over months, and does the variety loss show up in satisfaction?
- When platforms are required to offer a non-personalized feed, as the EU now does, who switches, and what changes?
Data that could help
Platform experiments with randomized recommendation policies, consumption logs over long horizons, and the EU's transparency reports on feed options.
Next step
Run or obtain a platform experiment that randomizes recommendation strength and measure purchases, variety and self-reported satisfaction over three months.