Algorithmic discrimination ------------------------------
Joy Buolamwini of MIT and Timnit Gebru of Microsoft Research New York studied the accuracy of facial-recognition software from three major companies in the field: IBM, Microsoft and Face++. Their results were comparable, for better and for worse (Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification, Proceedings of Machine Learning Research 81, 2018). While all three programs achieved a success rate of around 97% for light-skinned people, this fell to 83% for those with darker skin. A similar disparity emerged when distinguishing women from men. Combining the two characteristics produced striking results: nearly 100% accuracy for white men, but barely 70% for Black women!
That is not enough to accuse the designers of racism or sexism, but it is enough to remind us that artificial intelligence is precisely that—artificial—and can therefore reproduce certain biases in its own way, particularly through our choice of the data on which it is trained.
-
Smart data -------------------------
The rise of Big Data has brought artificial-intelligence algorithms back into the spotlight. Big Data even provides fertile ground for them, since machine learning, used by today's AI systems, feeds on vast amounts of data.
But AI is also a tool for harnessing Big Data. Such data are not limited to behavioral information cross-referenced for commercial purposes. Genome sequencing, particle accelerators and astronomy all produce torrents of data on a scale far beyond that generated by observing human activity. A special feature in issue 181 of Tangente, and another, more technical one in Tangente SUP 77–78, explore the mathematics of Big Data.
AI can optimize the algorithms used to store, classify and analyze these vast volumes of data. It can even detect phenomena that no one had thought to look for. In statistics-based methods, AI algorithms used for interpolation or extrapolation lead to adaptive forecasting models that become more accurate as they are tested against real-world data.
-
Preventing cancellations -------------------------
Marketing aims not only to attract new customers but also to keep them from leaving. Four years ago, SFR introduced a strategy for monitoring customers’ online activity—the pages they visit and the keywords they enter. According to Les Échos, these data enabled the mobile operator to anticipate more than 80% of potential subscription cancellations. By contacting these customers before they had made their decision and presenting them with a targeted offer, SFR would reportedly retain three-quarters of those who would otherwise leave.
This is an example of predictive marketing, which uses algorithms to analyze vast amounts of data to predict not the statistical behavior of a mass of customers but the behavior of individual customers. It is a promising field, but it must of course be accompanied by careful ethical thought about what companies may legitimately know about their customers.