We Need to Better Understand the Impact of Algorithmic Decision-Making

Markets are based on rules and institutions, and these are frequently altered by digital platforms. Three examples: Spotify links music producers and consumers through a largely new market mechanism. Alongside a new way of pricing, this specifically involves a comprehensive mechanism for music recommendation. Tripadvisor has introduced a comprehensive reputation and quality management mechanism. Uber changed the price-setting mechanism in the market for ad-hoc short-distance mobility by replacing time-or-distance based ride prices – customary for taxis – with prices flexibly determined by supply and demand.

Algorithmic decision-making is used in multiple ways on digital platforms. For example, Amazon uses a machine learning algorithm to recommend products that might be of interest to a customer. This recommendation is based upon data about other users’ purchasing behaviour. Machine learning is also employed on digital platforms to predict supply and demand and to automate price-setting mechanisms. Even the introduction of voice-driven terminal devices simply represents another new pathway for market access. An indication for the value that machine learning generates for digital platforms is that they dominate significant parts of top-level academic and applied research in this area and are willing to accept significant expenses to do so.

The benefits of algorithmic decision-making on digital platforms are often quite apparent. For instance, many consumers view product recommendations by platforms like Spotify or Amazon as simplifying their lives. From an economic perspective, this means a reduction in search costs. Opportunities also exist in areas that often are largely hidden from the users of the platforms. These include winnowing out dubious market participants from digital marketplaces, which often is only possible cheaply and efficiently through high levels of automation. What remains hidden in most cases are the risks – among them algorithmic discrimination.

Machine language algorithms are often calibrated on data of human behaviour. In the process, human misbehaviour may be incorporated into the algorithm and even be reinforced. Another problem area is so-called “selection bias” in the data used to calibrate the algorithms. If the calibration data includes few if any observations of social minorities, then the predictive quality of the algorithms for this portion of the population become expectably poorer. As a result, the minority may be systematically disadvantaged, for example, fully denying them market access or subjecting them to adverse automatic price-setting. However, as economists, we are only at a very early stage of understanding the impact of algorithmic decision-making on markets.

This question concerns scientists in different disciplines and decision-makers in various policy areas. Our aim is to better understand the impact of algorithmic decision-making on markets. Given the great importance and increasing scope of algorithmic decision-making, particularly through digital platforms, this becomes a matter of some urgency.