How the ranking loop actually works
Every major app runs a version of the same system: it predicts how likely you are to like a given profile, and how likely that profile is to like you back, then shows you the pairings with the highest joint probability. Older systems used a single desirability score in the style of chess Elo; current ones use collaborative filtering, which is closer to how streaming services recommend films, people who swiped like you also swiped on these. Nobody at the company writes a rule about race or age. The model simply notices patterns in aggregate behaviour and optimises for them.
Where bias enters the loop
Three places. Input bias: the historical swipe data reflects the population's existing preferences, including well-documented racial ones: OkCupid's own published data showed Black women and Asian men receiving markedly fewer responses than average. Feedback bias: profiles that get shown less get liked less, which the model reads as further evidence they are less desirable, so it shows them even less. Proxy bias: even with race removed from the model, features like neighbourhood, education, photo style and language cluster tightly with it, so the model reconstructs the pattern anyway. The result is a system that faithfully amplifies the average user's bias.
The other biases nobody names
Age falls off a cliff at the visible thresholds, 30, 40, 50, because filters are set in round numbers, not because attraction changes overnight. Height is filtered explicitly on some apps and heavily correlated with likes on all of them. Body type is punished by photo-first interfaces, which is precisely why prompt-first and photo-optional formats exist. Income and class leak in through job title, education and the backgrounds of your photos. And there is a straightforward pay bias: paid boosts and priority likes buy distribution, so ranking partly reflects spending, not fit.
What the platforms have tried
Removing ethnicity filters was the visible move, several apps dropped them after 2020, though it is a cosmetic fix when proxies remain. Some apps now inject exploration, deliberately showing you profiles the model is unsure about, which both improves the data and gives under-shown users a chance. Prompt-led designs like Hinge's shift some weight from photos to writing. Photo-blind and voice-first formats attack the problem at the interface rather than the model. Grindr and HER made a different bet, building for a specific community so the aggregate preference the model learns is not the general population's.
“The algorithm does not have opinions about you. It has opinions about everyone else's behaviour, which is worse.”
What you can actually do
You cannot out-argue a ranking model, but you can change what it has to work with. Loosen your own filters, especially the round-number age and distance ones, because tight filters shrink the pool the model can pair you into. Swipe deliberately rather than in long bursts, indiscriminate liking teaches the system nothing about you and, on several apps, lowers your standing. Reply quickly and finish conversations; engagement signals feed back into distribution. Consider a niche or community app where the base population already matches what you want, since that changes the data the model learns from. And treat a long dry spell as a signal to change format, photo-first to prompt-first, or general to niche, rather than to swipe harder.
The honest limit
Fairness in dating is not a solvable engineering problem, because personal attraction is not something anyone can be required to distribute evenly. What platforms can be held to is narrower: not amplifying bias beyond what users bring, not making distribution purely a function of spending, being transparent that ranking exists, and giving under-shown users genuine exposure rather than a quietly throttled feed. Judge an app on those four, not on whether it promises to be neutral.