A group of researchers from Princeton and the University of Washington studied how large language models behave when user interests conflict with platform incentives. In particular, they considered when a sponsored option is worse, more expensive, unnecessary, or potentially harmful, but it gets recommended anyway. Their main case study is booking a flight. Across twenty-three models, all but eight recommended the more expensive sponsored option more than half the time. Their paper also demonstrates how AI advertising is framed as a kind of advice—a helpful assistant becoming strangely enthusiastic about a product you never asked for.