As “big data” is making our machine learning algorithms smarter, the wealth of online information appears to have an opposite effect on society, engendering incredulity and parochialism. The two phenomena may not be as contradictory as they would seem at first glance, as they both stem from biases caused by simplified models of reality.
Machine learning methods train mathematical models in the hope of discovering regularities that may be translated into decisions in the real world. Learning is made possible by the fact that the machines are taught to look at the data through the lens of a specific model with limited representation capabilities. Indeed, the “no free lunch” theorem ensures that no learning is possible without making prior restrictive modeling assumptions: An inductive bias is necessary to learn . This approach has been extremely successful at identifying statistically significant patterns for tasks that permit the collections of large data sets, such as speech or image recognition and translation.
But the choice of the right model is crucial, and this is often painfully clear when machine learning interface with real-world decision and policy making. In fact, the bias caused by the model’s underlying assumptions, which may be hidden from the users of the algorithms, has frequently yielded nefarious consequences, e.g., when assessing credit worthiness or evaluating teachers’ performance . A particularly sticky problem is the inability of certain models to account for bias inherent in the selection of a given data set. This was reported, for instance, in the Beauty.AI pageant, in which the algorithm appeared to be discriminating against minorities.
The same phenomenon seems to be at work in the creation of today’s society of isolated, and yet connected, personal models of the world, as the wealth of available information makes it possible to select sources that only reinforce one’s beliefs. Without a feedback mechanism grounded on facts (not “alternative” ones), these personal, simplified and biased, models of reality foster, and are in turn supported by, irrational urges .
Politics appears to have played an important role in the public distrust of established conventional sources of information and in the specialization of isolated antagonistic models. In fact, as argued in , in the face of an ever more complex reality, today’s nations, via their politicians, have retreated into simplified models of the world that contrast good and evil players and put absolute faith into the power of financial markets. The continuous distortion of reality necessary to make events conform to such models has created widespread wariness for information that comes from outside one’s own views of how the “machine” works.
The over-reliance on biased models trained on selected data for decision making in the political and personal spheres is an important threat to the survival of peaceful and democratic societies that no technological advance can counteract — not even the “supernova” of . To put it as in :
“… to get our basic bearings we need, above all, greater precision in matters of the soul. The stunning events of our age of anger, and our perplexity before them, make it imperative that we anchor thought in the sphere of emotions; these upheavals demand nothing less than a radically enlarged understanding of what it means for human beings to pursue the contradictory ideals of freedom, equality and prosperity.”
 S. Shalev-Shwartz and S. Ben-David, Understanding machine learning: From theory to algorithms, Cambridge university press, 2014.
 C. O’Neil, Weapons of math destruction: How big data increases inequality and threatens democracy, Crown Publishing Group (NY); 2016 Sep 6.
 P. Mishra, Welcome to the age of anger, The Guardian, Dec. 8, 2016.
 A. Curtis, HyperNormalisation, BBC.
 T. Friedman, Thank you for being late, Farrar, Straus and Giroux, 2016.