Facebook’s response to Dehaye illustrates another consequence of the extreme asymmetries of knowledge at play. The company insisted that access to the requested data required it to surmount “huge technical challenges.” As behavioral surplus flows converge in machine-learning–based manufacturing operations, the sheer volume of data inputs and methods of analysis moves beyond human comprehension. Consider something as trivial as the case of Instagram’s machines selecting what images to show you. Its computations are based on varied streams of behavioral surplus from a subject user, then more streams from the friends in that user’s network, then more from the activities of people who follow the same accounts as the subject user, then the data and social links from the user’s Facebook activity. When it finally applies a ranking logic to predict what images the user will want to see next, that analysis also includes data on the past behavior of the subject user. Instagram has machines doing this “learning” because humans cannot.

Shoshana Zuboff, The Age of Surveillance Cap…, loc. 1546