Sailing with Artificial Intelligence: Recommendation systems and digital markets

A- A A+

In this complex digital age, people’s lives are greatly simplified by recommendation systems – AI algorithms that make it easier for online users to find the things they like in a vast ocean of options. These systems have profound impacts on individuals’ decisions and market outcomes. On one hand, they provide a higher match value than when individuals autonomously engage in a costly search for their preferred items. On the other hand, they lead to substantial market concentration and prompt sellers to raise their prices. Although our research generally indicates a positive net effect, the complex influence of these AI tools on economic outcomes requires careful examination and understanding.

In today’s digital era, the invisible hand guiding our online choices is often virtual. Many decisions made while browsing are guided by an Artificial Intelligence (AI)-powered recommendation system (RS). These algorithms have revolutionised how we discover products, music, films and more. Imagine navigating a vast sea where the sheer volume of options is overwhelming: 353 million products on Amazon, 90 million songs on Spotify, 26 billion videos on YouTube, and billions of news stories on your social networks. RSs are modern-day compasses in this digital ocean, steering our decisions, and ultimately shaping our tastes.

But these systems are far more than mere digital assistants. They are sophisticated algorithms that learn from our behaviours and preferences, as well as those of others. They are the actors behind the scenes, crafting our digital experiences to be as personalised and effective as possible. Today the impact of RSs is staggering: 75% of movies watched on Netflix, 35% of products viewed on Amazon, and 60% of videos watched on YouTube result from these recommendations. No sector is completely immune to the spread of RSs, which also shape choices of financial assets (with ‘robo-advising’), holiday destinations, and which crops to plant on farms. RSs also appear in more niche applications, such as suggesting to undergrads which courses to take, helping scholars decide which papers or books to read next, and advising editors of academic journals which reviewers to engage with.

This isn’t just about convenience. It’s a fundamental shift in how we interact with the digital world and, behind it, a fundamental and inescapable part of our lives.

But, as ever, with great power comes great responsibility and controversy. The influence of RSs extends beyond helping and simplifying choices; it touches on deeper issues. Democracy, for example, where RSs curate the news we read. From an economic perspective, RSs deeply affect market outcomes, competition and consumer welfare. RSs are undoubtedly beneficial in many ways, especially in vast digital markets; they provide suggestions that are a much better fit than we could otherwise achieve without the support of the AI. But they also raise growing concerns, with fears that these systems might lead to unintended consequences like price inflation, market monopolisation and the amplification of pre-existing biases.

Let’s delve into this intricate web. When we log onto a site like Amazon or Netflix, the recommendations we see are not just random selections. They result from algorithms processing vast amounts of data to predict what we, as an individual, might like next, ultimately providing personalised recommendations. This process, known as collaborative filtering, is akin to a digital form of word-of-mouth, where the system learns from the collective preferences of users to make predictions. The difference, however, is that these systems rely on a sheer amount of data from other users and other items or products, and they can predict individual preferences for unknown products/items very effectively.

But here’s where the plot thickens: the data used to train these algorithms is a product of our interactions, which are themselves influenced by previous recommendations. This creates a feedback loop, where the algorithm’s suggestions can reinforce existing preferences, potentially leading to a narrower field of choices, and exacerbating market outcomes. For instance, a popular song on Spotify might become even more popular simply because it’s being recommended more often – a phenomenon known as the ‘rich-get-richer’ effect.

This has led to heated debates among policy-makers, economists and computer scientists. For instance, the European Digital Service Act (DSA) calls for mitigating the “negative effects of personalised recommendations”. The concern is that left unchecked, these systems might enhance the market power of already dominant players, skewing competition and potentially harming consumer welfare (by diminishing choice). To understand these benefits and risks that may come with RSs, our recent research adopts a not-so-common approach, simulating AI systems in controlled, synthetic environments that mirror real-world economic scenarios. This method allows us to dissect the complex interplay between AI algorithms and market forces with very precise detail, and in ways that would be difficult to achieve with empirical research (much of the data are not available to researchers in these cases) and theoretical investigations (modeling complex AI algorithms turns out to be impractical).

Our findings are both fascinating and nuanced. On the one hand, RSs undoubtedly make life easier, helping us sift through the digital clutter to find products or content that aligns with our interests. This saves time (and search costs) and enhances our experience by connecting us with items we might otherwise never have discovered. In more technical terms, the expected utility a user can obtain by participating in a market with the help of an RS can be increased by as much as 6% compared with a benchmark where the consumer must autonomously perform a costly search for preferred items. RSs are more effective at making decisions for people than people are.

On the other hand, we cannot overlook the impact of RSs on markets. These systems can significantly increase market concentration, turning certain products into ‘superstars’, sometimes leading to price hikes of up to 16%. In fact, in our research we show that with sellers of items in digital markets increase their prices once they realise that consumers’ behaviour is mediated by an RS. So, the increased utility comes at a real cost.

Although these developments raise legitimate concerns about consumer welfare, we show that the picture is not entirely negative. The improved match granted by RSs can lead to an overall increase in consumer welfare, even when factoring in the higher prices.

Our simulation approach enables us to delve further into the underlying causes of these significant effects of RSs. For instance, we have identified that the demand expressed by consumers assisted by RSs – what we term ‘algorithmic demand’ – differs markedly from the unassisted ‘human demand’. This difference is illustrated in Figure 1. By analysing the shifts and bending in algorithmic demand, we gain a deeper understanding of the functioning and impact of personalised recommendations.


Figure 1: Algorithmic and human demand


Source: Authors’ calculations

Note: The red line is the “algorithmic demand” mediated by the recommender system, the blue line is the “human demand”, dots indicate the equilibrium prices for this item

Our simulation method also allows us to examine environments featuring both exogenous and endogenous data generated by RSs. By comparing these two scenarios, we have discovered that the reliance of AI algorithms on endogenous data (i.e., market data they themselves contribute to generate) does influence market outcomes, such as increasing market concentration, though this is a secondary effect.

Interestingly, we find that the relationship between the amount of information RSs use and consumer welfare is not linear. Initially, as RSs access more data, consumer welfare improves due to better matching of products to consumer preferences. But past a certain point, negative impacts like price increases begin to outweigh these benefits. This discovery points to a need for a balanced approach in the data utilisation by RSs, suggesting areas for potential regulatory intervention. As with most regulations in the digital economy, striking the right balance between protecting consumer welfare while still promoting innovation is key.

Our research also investigates the darker aspects of RSs, such as the potential for platforms to manipulate recommendations for profit. This is particularly concerning when the platform deploying the RS also sells products that consumers might choose, creating a conflict of interest. While such manipulation could have detrimental effects, our research indicates that it also tends to lower prices for favoured and over-recommended products, thereby mitigating some of the adverse impacts on consumer welfare. This finding does however open new avenues for understanding the strategic behaviours of platforms and their implications for market competition.

RSs are akin to a double-edged sword. They are powerful tools that enhance our digital experiences, but their broader implications on markets and consumer welfare necessitate careful scrutiny. As Chat GPT sees it, “as we venture further into this digital ocean, understanding

and molding the role of RSs in our lives becomes a technological quest and a societal duty. This research is merely the first step in a longer expedition to decode the intricate influence of AI on our existence. Who knows, maybe AI and Recommendation Systems will eventually be savvy enough to suggest the best course for our research voyage, ensuring we don’t sail off the edge of the digital world map.”

Author: Giacomo Calzolari

Author’s Note: This piece is based on the results in Calvano, Emilio and Calzolari, Giacomo, Denicolò, Vincenzo and Pastorello, Sergio, Artificial Intelligence, Algorithmic Recommendations and Competition (December, 2023). Available here. ChatGPT, an AI language model developed by OpenAI, helped improve the language of this text at first draft.