Your taste isn't enough anymore
Dating apps may be the first products to evolve their matching from chess ratings to psychological models, but they won't be the last.
Have you ever thought about how games match you with other players in zero-sum games, chess specifically? No, it’s not random. And most of the time, it’s not only dependent on the level you choose.
They use the Elo rating system, which dynamically calculates a player’s strength relative to other players based on the matches they’ve played. Imagine all players having an imaginary number that says how good they are at the game. If you win, your number goes up a little. If you lose, your number goes down a little.
But it’s not limited to that. It also matters who you beat or lose against. If you beat someone with a much higher number than you, your number jumps much more, indicating you must be really good. Unlike beating someone with a much lower rating than you, your number only goes up a tiny bit, as everyone expects you to win anyway.
It works the same way in reverse. If you lose to someone way better than you, it’s no surprise, your number barely drops. But if you lose to someone way worse than you, oops.
Dating apps used to have a similar rating system too. Tinder publicly mentioned that it used an Elo based ranking system for its users until it announced in 2019 that it no longer did, in favor of a more complex recommendation algorithm.
Early apps knew almost nothing about users, and they had to find a notion of desirability. Elo was a proven shortcut.
Yes, dating apps do rank users by “desirability”, even if it’s no longer a pure Elo score. It’s based on how much they’re swiped right on, and who is swiping right or left on them. Because, you know, being liked by someone who is swiped right on often is a signal of being desirable. And as mean as it sounds, vice versa. Being swiped left on by people who are rarely swiped right on can decrease a user’s ranking significantly.
Swiping, or sending likes, whatever you call it, isn’t just UX, it’s an incredible data generation machine. Instead of manually assigning scores, users trained the models.
You might have heard that Bumble founder and CEO Whitney Wolfe Herd confirmed a few months ago that swiping is going to be history soon. It feels like the optimization target is changing again. About a year earlier, she had also announced that they were working on a new concept focused on understanding users’ emotional attachment and relationship patterns. That includes their past experiences and dating behavior in general.
In simpler words, an app that tries to understand your dating patterns. How likely you are to ghost someone, fall for someone, how attached you get, how quickly you’ll lose interest, or whether you’re likely to commit.
Back then, it was about predicting attraction. Matching, and not really beyond. The new objective isn’t only learning users’ tastes, but also understanding how likely they are to make a relationship work, with the help of AI assistants.
We still don’t know what the end product will look like or how successful it’s going to be. But I can say that it sounds like the natural evolution of personalization.
Maybe, dating will simply be the first consumer product where this shift becomes obvious. It’s going from ranking desirability and predicting attraction to understanding the psychology behind our decisions.
But it won’t stop at dating.
If we step back and look at the bigger picture, products are moving from modeling what we are (skills, demographics, a CV, a bio) to modeling how we behave over time. Static traits to dynamic patterns.
Can’t this be applied elsewhere too?
Imagine recruiting. Not just matching a resume to a job description, but predicting the likelihood of a long-term click with a team. Do they tend to lose interest once the honeymoon phase is over? Do they thrive with autonomy or need structure? Do they confront conflict or avoid it?
Or education. Not just serving you the next topic, but understanding how you actually learn. Do you give up after one failed attempt or push through? Do you learn better from competition or collaboration?
In both cases, the question changes from “what can this person do” to “how will this person behave”.
That’s a much more intimate thing to model.
Preferences are easy to admit to, even to an algorithm. Patterns, motivations, attachment style, that’s the stuff we don’t always know about ourselves. Are we comfortable with products inferring it anyway?
Products used to model our preferences. Now, it looks like they’re trying to model us.
While we expect them to get better at predicting our choices, maybe they’ll get better at understanding the pattern behind our choices, and even influence them.