Abstract
In the modern economy, consumption and other important economic choices are increasingly shaped by algorithms powered by artificial intelligence (AI). Despite their importance, relatively little is known about the economic effects of these algorithms, their normative desirability and any consequent economic policy prescriptions. This research proposal is based on two interlinked project components both studying the economic consequences of deploying popular classes of Al-powered algorithms. The first component focuses on Recommender Systems (RS). These are algorithms that estimate user preferences over a given set of items and use these estimates to match users to items. Amazon, Netflix, Youtube, Spotify and social media apps are popular commercial applications of this technology allowing to filter items (products, media, user posts) to produce personalized experiences. Another prominent application is the advertising business. The main goal of the first component is to conduct an experimental analysis of RS in the spirit of Calvano, Calzolari, Decicolò ad Pastorello (2020, American Economic Review). We plan to build simplified but reasonably realistic RSs and analyze the way their recommendations evolve in different economic settings. We also plan to analyze how firms' and consumers' behavior changes in the presence of RSs, studying in particular whether recommendations are biased and in what direction, whether the presence of RS intensifles or attenuates competition, and how it changes the competitive landscape, affecting the relative position of incumbents and entrants. The second component focuses on two classes of algorithms (Q-learners and Neural Networks) and on their deployment to assist bidders in the online advertising auctions where internet advertising is traded. Online advertising is a major financial engine for most digital platforms. Within this sector, we will focus on the search auctions used to sell ads on search engines, a market that accounts for approximately half of the total revenues of online advertising (i.e., $120 billion in 2020 in the US). The second component of the study will thus provide the first evidence on the use of AI algorithms in the multi-items auctions where digital ad is traded (according to both VCG and GSP auction mechanisms). It will use the same experimental approach described above for component 1, but in an auction setting. We will account for both individual and joint bidding, as shown to be crucial in Decarolis and Roviatti (202 1, American Economic Review). Compared to previous studies, a major feature of our work will also be to address the question of information availability by exploring how different algorithms (Q-learners vs Neural Networks) respond to the availability of more granular information and how the selling platform can distort the information it passes to the advertisers' algorithms in order to increase its revenues.
Results achieved
The research has led to the development of a novel economic analysis of Recommender Systems: AI-based algorithms that use feedback such as ratings, clicks, purchases, and other forms of user activity to predict consumers’ interest in products they have not yet experienced. Platforms routinely use these predictions to provide personalized recommendations, thereby guiding consumers toward selected products. Our analysis identifies a fundamental channel through which recommender systems can affect product-market competition. Algorithmic recommendations create personalized prominence: different consumers are directed to different products on the basis of the algorithm’s estimates of their match values, and consumers use these recommended products as the starting point of their search. This reshapes demand elasticities and, in turn, affects equilibrium prices. We show that algorithmic recommendations lead to systematically higher prices. Three distinct mechanisms drive this result: front-loading, reliance, and price salience. The front-loading effect arises because consumers who inspect a product first are, under personalized prominence, disproportionately those with high match values for that product, while low-match consumers are directed elsewhere. As a result, firms face a less elastic pool of first-time visitors and have stronger incentives to raise prices. The reliance effect arises when consumers understand that recommendations are informative. In that case, they infer that products not recommended to them are, on average, less attractive than randomly drawn alternatives. This lowers the expected value of continued search and makes consumers more captive to the first firm they visit, again softening price competition. The price-salience effect arises when prices are observed before consumers inspect products. In this case, search is directed: the order in which consumers inspect products depends on prices and possibly on other observable product characteristics. Firms then have an additional incentive to cut prices, because a lower price can both increase the probability of purchase conditional on inspection and move the product earlier in consumers’ search order. When recommendations take prices into account, this price-undercutting motive exists both with and without recommender systems, but it is weaker with them. Recommendations rank products using estimated match values as well as prices, so price becomes only one component of prominence. A price cut therefore has less influence on search order than it would under unassisted search. Recommender systems thus weaken this channel of price competition and exert additional upward pressure on prices. We also show that the resulting price increase can be large enough to offset the benefits that recommender systems generate for consumers through improved matching and reduced search costs. Thus, although algorithmic recommendations may help consumers find better-suited products more efficiently, they may also soften competition in ways that reduce consumer welfare. The paper has been circulated as a CEPR Discussion Paper and has received a Revise and Resubmit from the Review of Economic Studies. We are currently finalizing the revision and plan to resubmit it in the coming weeks.Project details
Unibo Team Leader: Vincenzo Denicolò
Unibo involved Department/s:
Dipartimento di Scienze Economiche
Coordinator:
Università Commerciale Luigi Bocconi MILANO(Italy)
Total Unibo Contribution: Euro (EUR) 67.473,00
Project Duration in months: 29
Start Date:
28/09/2023
End Date:
28/02/2026
