Machine learning algorithms have obviously got a huge amount of potential as they allow organizations to crunch through vast amounts of data to make smarter decisions. They have also been accused, however, of worsening racial equality by using inherently biased data in their training datasets.
Airbnb has notably developed an AI-based smart-pricing tool that allows hosts to set the price of their property each day automatically. New research from Carnegie Mellon University explores whether the algorithm had any impact on racial disparities between Airbnb hosts. The analysis reveals that while the tool did indeed manage to narrow the revenue gap between White and Black hosts, it actually made things worse overall because fewer Black hosts were using the algorithm.
“The disparity in revenues earned by White and Black hosts has been the subject of a lot of negative publicity for Airbnb in the last few years,” the researchers say. “Our results show that a smart-pricing algorithm can be effective in mitigating racial disparities, but that effectiveness is limited by the extent to which the tool is adopted.”
Unintended consequences
The pricing algorithm was introduced by Airbnb in November 2015, with the researchers examining behavior on the platform from July of that year through to August 2017. Over 9,000 properties were randomly selected across seven American cities. Of those hosts, 2,118 had adopted the algorithm during the study period. The researchers were particularly interested in the average daily revenue per host, and broke this down by the race/ethnicity of the host.
The analysis showed that prior to the introduction of the algorithm, White hosts were earning around $12.16 more per day than Black hosts with similar properties in similar neighborhoods. This was primarily because demand for properties with Black hosts was around 20% lower than for White hosts, with this apparent racial bias well documented throughout the media at the time.
The introduction of the algorithm did indeed rectify this, with the analysis revealing that Black hosts benefited more than White hosts as it helped to level out demand for properties rented by Black hosts. The problem is that Black hosts were 41% less likely to actually use the algorithm than White hosts, so while the algorithm was beneficial for those who used it, because of the unequal distribution of users, it actually made the revenue gap wider between Black and White hosts.
“Given the much lower rate of adoption of the algorithm by Black hosts than White hosts, managers may want to devise strategies to encourage Black hosts to adopt the algorithm,” Shunyuan Zhang, the researchers conclude. “Otherwise, an algorithm that could reduce disparities may end up increasing them.”
