AI shopping agents swing wildly on product picks
researchers test six ai models as shopping agents, find tiny context shifts change purchase decisions drastically. this affects e-commerce and ai reliability.
researchers at the wharton school tested six current ai models, including mini variants and frontier-level models, as personal shopping assistants. they used the aces simulator to show the ai agent a product page and analyze the image. the study found that even tiny shifts in context, such as adding a single external source, swung purchase decisions by wide margins. wirecutter had the strongest influence, increasing the probability of picking the fitbit inspire 3 by 90 percentage points for claude opus 4.8. the order of sources and delivery method also affected product picks. the study's findings impact the use of ai in e-commerce and highlight the need for more reliable ai decision-making.