The algorithm as matchmaker

Every machine learning system optimizes an objective function. The choice of function determines everything downstream. Dating apps have, without exception, chosen functions that reward engagement: swipes, matches, messages, returns. None has chosen, because none can measure, "produces stable marriages with high reported satisfaction at year ten." The measurable proxies are not the goal. They are correlated with the goal at very low resolutions and decorrelated at the resolutions that matter. The algorithm is therefore not failing at matchmaking. It is succeeding at a different task that is rhetorically dressed as matchmaking, and the gap between the rhetoric and the function is where the collective harm accumulates.

Tinder confirmed in 2019 that it had maintained, and largely deprecated, an internal "Elo score" that ranked users by desirability based on swipe behavior. The successor systems are functionally similar and operate under different names. Bumble, Hinge, and the OkCupid algorithms all maintain analogous scoring. Users are paired within bands of this score. This means that, in effect, the algorithm has imposed a caste system on the dating market — invisible to the user, unappealable, and self-reinforcing. A user with a low score sees low-score profiles, receives few matches, and never accumulates the engagement that would raise the score. The system is stable in the worst possible way: it locks the bottom in place.

Rudder's data from OkCupid showed that, in every year measured, Black women received the fewest responses and Asian men the fewest responses, controlling for self-reported attractiveness, education, and income. These are user behaviors, not algorithmic choices. But the algorithm trains on them. When the model learns that profiles of a certain type are systematically swiped past, it lowers their impression rate to preserve match probability, which further reduces their exposure, which further suppresses their match rate. The bias is in the users; the amplification is in the model. The collective effect is the construction of a romance market that operates with the prejudices of 1955 and the efficiency of 2025.

A machine learning system can act only on features it can measure. The features available to a dating app are: photo embeddings, text embeddings, demographic fields, swipe behavior, message metadata. The features that actually predict relationship success — kindness, humor in real time, emotional regulation under stress, shared moral imagination, the quality of attention paid to small things — are not in the feature set. They cannot be. The algorithm is therefore matching on a thin slice of the relevant variables and is, by construction, blind to the thick slice. It does this with the confidence of a system that does not know what it does not know.

The traditional matchmaker had a long memory and a reputation in the community. A bad match cost her. A good one earned her future commissions and social standing. This produced an alignment between her interests and the couple's that no algorithm has. The algorithm does not lose money on a divorce. It does not lose status when a couple it formed turns out to be miserable. It cannot, because it does not observe the outcome. The accountability gap is not a fixable bug. It is a feature of optimization at scale: the further the matcher is from the matched, the less the matcher can be held to the matched's wellbeing.

When a user first joins an app, the algorithm has no behavioral data and must rely on declared preferences and demographic priors. These priors are, in practice, drawn from population-level statistics — and they are drawn in ways that perpetuate stereotype. New Black women users are shown to a smaller pool. New older users are shown to a smaller pool. The cold start is therefore not a neutral seed; it is a re-application of the average user's revealed prejudice, applied to every new arrival before they have generated any data of their own. The discrimination is automated before the user has done anything.

Every recommender faces a tradeoff: show the user what they have liked before (exploitation) or expose them to new things (exploration). Dating apps lean heavily toward exploitation, because exploration costs engagement in the short term. The result is a feed that narrows over time, showing each user a tighter and tighter range of phenotypes that match their early swipe behavior. Users who, in real life, might have been pleasantly surprised by a partner outside their type are never given the chance, because the algorithm has decided early what their type is and has stopped offering deviations. The serendipity that the human matchmaker provided — the unexpected pairing that works — is algorithmically extinct.

Each swipe trains the model. The model determines what is shown next. What is shown next determines the next swipe. This is a closed loop with no external input correcting it. Over six months of use, the typical user's feed has converged to a narrow band of profiles that closely resemble each other. The user experiences this as "the app finally getting it." What is actually happening is that the user has been confined to a basin in the option space, with the rest of the population invisible. The narrowing is mistaken for personalization. It is in fact confinement.

Because the algorithm allocates visibility proportional to received attention, the top of the desirability distribution captures a disproportionate share of all matches. On most platforms, the top 10% of men receive over 50% of female right-swipes; the corresponding figure for women is less skewed but still concentrated. This is the romantic analogue of a Pareto distribution, and the algorithm is the mechanism that maintains it. A human matchmaker, knowing the community, would have actively redistributed introductions to give the less-attended a fair hearing. The algorithm does the opposite, by design.

No dating algorithm is published. No external audit is permitted. No regulator has the technical capacity to inspect the systems that mediate, at this point, a majority of new romantic pairings in the developed world. This is unprecedented. Every prior matchmaking institution — religious, familial, communal — was subject to social inspection. The current generation of matchmakers operates in a black box, defended by trade-secret law, and reports to shareholders rather than to the matched. The democratic deficit is significant and has not yet entered serious public debate.

The algorithm is, ultimately, a substitution of measurable proxies for the unmeasurable judgments that pairing has always required. This substitution is sometimes useful — a coarse filter is better than no filter when the search space is enormous. But the substitution becomes pathological when the proxy is mistaken for the goal, which is what the apps have done. Users now report a strange phenomenon: a partner who looks like a good match on paper and feels wrong in person, or feels right in person and looks wrong on paper. The dissonance is the gap between metric and reality, and the algorithm is fluent in the metric and blind to the reality.

The algorithm is not going away. It will become more sophisticated, more multimodal, more capable of approximating the human matchmaker's judgment. But "more sophisticated" is not the same as "well aligned." The next institution of romantic mediation will need to combine the scale of the algorithm with the accountability of the yenta — paid by the matched rather than by advertisers, audited by users rather than by shareholders, evaluated on five-year outcomes rather than on next-day returns. Such institutions are beginning to emerge at the margins: subscription matchmakers, vetted networks, community-embedded services. Their growth is the leading indicator of whether the algorithm gets demoted to a tool or remains, by default, the sovereign of the love market.

1. Rudder, Christian. Dataclysm: Who We Are When We Think No One's Looking. New York: Crown, 2014. 2. Finkel, Eli J. The All-or-Nothing Marriage: How the Best Marriages Work. New York: Dutton, 2017. 3. Ansari, Aziz, and Eric Klinenberg. Modern Romance. New York: Penguin Press, 2015. 4. Schwartz, Barry. The Paradox of Choice: Why More Is Less. New York: Ecco, 2004. 5. Fisher, Helen. Anatomy of Love: A Natural History of Mating, Marriage, and Why We Stray. Rev. ed. New York: W. W. Norton, 2016. 6. Bergström, Marie. The New Laws of Love: Online Dating and the Privatization of Intimacy. Cambridge: Polity Press, 2021. 7. Turkle, Sherry. Alone Together: Why We Expect More from Technology and Less from Each Other. New York: Basic Books, 2011. 8. Arendt, Hannah. The Human Condition. Chicago: University of Chicago Press, 1958. 9. Tolentino, Jia. Trick Mirror: Reflections on Self-Delusion. New York: Random House, 2019. 10. Weigel, Moira. Labor of Love: The Invention of Dating. New York: Farrar, Straus and Giroux, 2016. 11. Wade, Lisa. American Hookup: The New Culture of Sex on Campus. New York: W. W. Norton, 2017. 12. Rosenfeld, Michael J., Reuben J. Thomas, and Sonia Hausen. "Disintermediating Your Friends: How Online Dating in the United States Displaces Other Ways of Meeting." Proceedings of the National Academy of Sciences 116, no. 36 (2019): 17753–58.