Pornhub · Preference Assignment

In 2020, technology journalists working with internal sources at Pornhub — the largest pornographic-video platform then in operation — published investigations (in Wired, the New York Times, and elsewhere) examining the platform's content moderation and recommendation systems. The recommendation-system finding, less reported than the parallel content-moderation scandal of the same period, is the one this case treats. The finding, in its most-cited phrasing, was that the platform's algorithm "unilaterally assigns sexual preferences to each individual user without their knowledge." Specifically: the engine, after fewer than one hundred video views, categorized users into preference profiles that were then used to shape every subsequent recommendation. The categorization was performed without the user's awareness; the resulting profile was not visible to the user; the profile could not, in practical terms, be revised.

The investigation's significance is the affective domain in which the operation was occurring. The specific operation it documented was old news — recommendation algorithms have inferred user preferences for as long as they have existed. The preferences in question were sexual; the inferences were being drawn from the user's least-articulable response patterns; the resulting profile was determining, in significant measure, what the user would subsequently encounter as their own desire. Pornhub's recommendation engine instantiates the Desire Engine operating on the most intimate possible variable. The mechanism is identical to the mechanism deployed by every other major recommendation platform; the variable is different. This is the case Preciado's pharmacopornographic regime anticipated. Preciado argued, in 2008, that contemporary capital had achieved a degree of integration with sexual and pharmacological subject-formation that no earlier regime had approached. The pornographic platform, in Preciado's frame, is the regime's clearest operational instance.

Strip the case back to its mechanism and it is plain: the platform reads patterns in the user's micro-responses and, on the strength of them, serves content its model predicts the user will respond to. The content is predicted, and the user never independently sought it. The user's experience of "my preferences" is, in this case, the platform's prediction of the user, fed back to the user as content. The Pornhub case is the cleanest demonstration that the apparatus's Desire Engine operates in the same way regardless of the domain in which it is deployed. The mechanism the engine implements on a user's sexual response is the mechanism it implements on a user's news preference, fashion preference, political preference, and food preference. The platform-economic apologia that recommendation algorithms "give users what they want" survives only as long as one does not press on what want means when the variable being predicted is the user's least-articulable response.

The least-cited finding from the investigation was probably the most damning one: users in the platform's own testing said they were satisfied with the recommendations the engine assigned them. The engine works, in the sense its operators meant it to work, and the satisfaction the engine produces is part of how we know it does. This is the harder thing to sit with. There is no overriding of the occupant's desires against their will: the apparatus produces the desires the occupant subsequently experiences as theirs, and the experience is good.