Investigating Misinformation in Online Marketplaces: An Audit Study on Amazon

Search and recommendation systems are ubiquitous and irreplaceable tools in our daily lives. Despite their critical role in selecting and ranking the most relevant information, they typically do not consider the veracity of information presented to the user. In this paper, we introduce an audit methodology to investigate the extent of misinformation presented in search results and recommendations on online marketplaces. We investigate the factors and personalization attributes that influence the amount of misinformation in searches and recommendations. Recently, several media reports criticized Amazon for hosting and recommending items that promote misinformation on topics such as vaccines. Motivated by those reports, we apply our algorithmic auditing methodology on Amazon to verify those claims. Our audit study investigates (a) factors that might influence the search algorithms of Amazon and (b) personalization attributes that contribute to amplifying the amount of misinformation recommended to users in their search results and recommendations. Our audit study collected ~526k search results and ~182k homepage recommendations, with ~8.5k unique items. Each item is annotated for its stance on vaccines' misinformation (pro, neutral, or anti). Our study reveals that (1) the selection and ranking by the default Featured search algorithm of search results that have misinformation stances are positively correlated with the stance of search queries and customers' evaluation of items (ratings and reviews), (2) misinformation stances of search results are neither affected by users' activities nor by interacting (browsing, wish-listing, shopping) with items that have a misinformation stance, and (3) a filter bubble built-in users' homepages have a misinformation stance positively correlated with the misinformation stance of items that a user interacts with.