For most of human history, the matchmaker was a person — a mother, an aunt, a village elder, a professional yenta — who held in her head a working model of the available pool, the temperaments involved, the family situations, the deep compatibilities and the obvious mismatches. She had skin in the game. If she paired badly, her reputation suffered and her future commissions dried up. The match was, in the broadest sense, accountable. The algorithm is not.
The algorithmic matchmaker — the ranking and recommendation system inside every major dating app — is a different kind of entity entirely. It is optimizing for an objective function, and that objective function is not "produce stable, satisfied couples." It cannot be, because such an outcome is unobservable at the scale and speed required. The function is, in every documented case, some proxy for engagement: probability that the recommended profile receives a swipe, probability that a swipe produces a match, probability that a match produces a message, probability that the user returns tomorrow. Each of these proxies is correlated with relationship quality at the low end and decorrelated, or anti-correlated, with it at the high end. A profile that everyone swipes on is not the profile of the partner most people should marry. It is the profile of the partner most people will engage with on a screen, which is a different thing entirely.
At the collective level, the algorithmic matchmaker is doing something the human matchmaker never did: aggregating and acting on the revealed preferences of millions of users, then redistributing visibility according to those preferences. This produces feedback loops. Profiles that match the current aesthetic consensus receive more impressions, get more matches, train the model further, and become more visible still. Profiles that deviate are buried. Over a decade, this has measurable effects on the visible distribution of partners: a narrowing of who is shown to whom, a hardening of phenotypic and demographic sorting, and the algorithmic amplification of biases that the human matchmaker, embedded in community and accountable to it, would have softened.
The most consequential of these is the racial filter. Rudder's OkCupid data, published in 2014, showed that user behavior on the platform produced a stark racial hierarchy of response rates — and that the algorithm, in optimizing for matches, had internalized this hierarchy and was redistributing visibility accordingly. Black women and Asian men sat at the bottom of every response-rate chart. The algorithm did not invent the prejudice. It encoded it and made it operational at scale, in ways that no individual user fully saw because each user only experiences their own feed. The collective effect is the construction of a romantic market that is, in measurable terms, more segregated than the underlying society — the inverse of what online dating was supposed to do.
A second consequence is the elo-score regime. Most apps maintain an internal desirability score for each user, computed from the swipe behavior of others. Users are shown profiles roughly matched to their own score, on the theory that this maximizes match probability. The effect, over time, is the construction of tiers — invisible, unappealable, internally consistent — through which users circulate. A high-tier user is shown other high-tier users and rarely sees the rest. A low-tier user is shown other low-tier users and rarely sees the rest. This is not how human matchmakers worked. The yenta was perfectly capable of pairing a kind man of modest looks with a brilliant woman of modest looks. The algorithm cannot do this, because the algorithm does not know what "kind" means and does not have the social context to override the scoring.
A third consequence is the legibility problem. The algorithm operates on features it can measure — photographs, response times, message lengths, swipe patterns — and is blind to the features that actually predict relationship success: shared values, complementary temperaments, the way two people laugh together. These are not in the feature set, not because they do not matter but because they cannot be captured by the available instrumentation. The matchmaker is therefore optimizing in a feature space that systematically excludes the things that count, and is doing so with computational confidence that the human matchmaker never claimed.
The collective implication is that we have replaced an accountable, embedded, low-throughput pairing institution with an unaccountable, disembedded, high-throughput one — and we have done so without any public deliberation about whether this was a good trade. The algorithm matches more people faster and worse. It encodes existing prejudices and projects them at scale. It selects for screen-legibility rather than for partnership. And it does all of this behind a black box that the user cannot inspect and the regulator does not understand.
The corrective is not to abolish the algorithm but to demote it. Treat its outputs as one input among many, not as the answer. Restore the broker functions that algorithms cannot perform — the friend who introduces, the cousin who vouches, the colleague who sets up the dinner. The yenta is not obsolete. She was simply outcompeted, briefly, by a system that scales but does not see, and the rebuilding of pairing institutions that can both scale and see is one of the open civilizational projects of the coming decades.