Friendship has always been shaped by the infrastructure through which it finds its conditions of possibility. The Roman forum, the medieval guild, the eighteenth-century coffeehouse, the twentieth-century suburb with its cul-de-sacs and PTA meetings—each of these was an infrastructure that sorted people into proximity and created the conditions under which friendship could form. What is new about the present moment is not that infrastructure shapes friendship formation; it always has. What is new is that the infrastructure making the sorting decisions is an algorithm, and the algorithm's optimization target has nothing to do with friendship.
Algorithm-mediated friend discovery is the process by which social platforms suggest, surface, and connect potential friends based on behavioral and demographic data. When Instagram's "Suggested for You" shows you accounts, when Facebook's "People You May Know" generates a list, when TikTok's For You page puts a creator in front of you whose content resonates—these are algorithmic interventions in the friendship formation process. The algorithm is not a neutral matchmaker. It is a product feature optimized to increase the time you spend on the platform, the number of accounts you follow, and the engagement activity (likes, comments, shares) you generate. These optimization targets do not overlap neatly with what produces good friendship.
What the algorithm knows about you is a behavioral profile: what you click on, how long you look at things, what accounts you follow, what you post, when you're active, what geography your device resolves to. From this profile it infers what kind of people might hold your attention. The inference is often accurate—people who like similar content, who are in similar social networks, who are at similar life stages do often make good friends. But the inference is made on the basis of engagement probability, not friendship quality. The algorithm cannot distinguish between someone who would keep you addicted to their content and someone who would become a genuine close friend. From the algorithm's perspective, these are the same: both generate engagement.
The result at the collective level is that algorithmic friend discovery systematically privileges certain kinds of similarity while being blind to dimensions of compatibility that don't produce engagement data. It can tell that you and someone else both love a niche music genre because you both engage with that content. It cannot tell that you and someone else share a quality of relational attention, a capacity for honest conflict, or a way of holding difficulty that would make the friendship genuinely sustaining. The surface similarities are legible to the algorithm; the deep compatibilities are not. Algorithmic friend discovery therefore systematically produces friendships organized around shared taste and shared demographic rather than shared depth.
There is also an ideological problem with algorithmic friend discovery that is underexamined: the algorithm reproduces the social structure it operates within. If your existing social network is racially homogeneous, the algorithm's analysis of your mutual-connection graph will suggest people who are also racially homogeneous. If your consumption patterns reflect class-specific cultural preferences, the algorithm will suggest people with similar class positions. The algorithmic suggestion does not disrupt social stratification; it reflects and reinforces it. The dream of the internet as a technology for expanding social horizons beyond the constraints of geography, class, and ethnicity runs directly into the reality of algorithms that optimize for sameness because sameness produces predictable engagement.
The deeper problem is the opacity of the process. You do not know why a particular person appeared in your "suggested friends" list. You do not know what data about you was used in the calculation, what data about them was used, or what the weighting was. You cannot audit the algorithm's model of who you should know. You simply receive its output as a naturalized social environment—these are the people in your feed, this is who the platform thinks you should connect with—and most users have no basis for questioning whether the suggestions are good ones by any criterion other than "this feels like someone I might like."
What is at stake here is not just individual friendship quality—it is the composition of social networks at the population level. If algorithmic friend discovery is the primary mechanism by which adults form new friendships, and if that algorithm consistently produces networks organized around consumption similarity and demographic clustering, the aggregate effect is a social world of increasingly homogeneous pods: people who all see the same content, share the same references, occupy the same class and cultural position, and have been pre-sorted by a machine that was optimizing for watch time. The costs of this at the level of democratic culture, cross-cultural understanding, and social solidarity are not trivial. The algorithm did not set out to balkanize society. But balkanization is what engagement optimization produces, because the content and people most likely to hold your attention are most likely to be the ones least challenging to your existing worldview.