In 2018, an engineer from San Francisco named Jake wrote an algorithm that turned his life upside down. Not for a startup, not for money. He was just tired of Tinder. Hundreds of matches, dozens of dates, not a single one he wanted to see again. Jake sat down and wrote a program that analyzed his own visual preferences. Two weeks later, the algorithm suggested a profile of a woman he ended up living with for three years. This story isn’t about magic. It’s about how data science works when applied to the most human of all tasks.
How a neural network sees what you don’t
When you swipe through photos on a dating app, you think you’re the one in charge. You look, you judge, you decide. But the truth is, your brain makes a decision in 300 milliseconds. That’s less time than it takes to blink. The rest of the time, your brain is busy retroactively coming up with justifications for a decision your subconscious already made.
A neural network works differently. It doesn’t make hasty decisions. It doesn’t analyze just one photo, but a hundred. It finds patterns you don’t even suspect in yourself. Maybe you’re attracted to people with a certain eye shape. Or you unconsciously gravitate toward those who share similar facial features to your own. You don’t see it—the neural network does.
Professor Michael Kosinski from Stanford, known for his work on digital footprints, conducted a large-scale study in 2024. His team trained a neural network on the visual preferences of several thousand people, then tested the accuracy of its predictions. The result: 78 percent accuracy. Compared to 34 percent for random matching. A 2.3-fold difference. And that’s only on visual data—without considering interests, values, or personality.
The 128-dimensional space of your attractiveness
Technically, it looks like this: each face you rate is transformed into a point in 128-dimensional space. No, that’s not a metaphor. FaceNet—a neural network published by Google in 2015 with 99.63 percent accuracy—converts an image into a numerical vector. Your brain does the same thing; you just don’t know how to describe it.
The difference between you and the neural network is this: it remembers every single decision you make. Every swipe. Every profile you lingered on for half a second longer than usual. Every photo you zoomed in on to see the details. You don’t remember any of that. Your conscious mind filters and distorts subconscious signals, molding them into socially acceptable patterns. The neural network doesn’t filter.
That’s why algorithmic matching often turns out to be more accurate than intuitive matching. Not because algorithms are smarter. But because they’re more honest. They’re not embarrassed that you’re attracted to types that aren’t conventionally considered attractive. They don’t care that you’re ashamed of your true preferences. They just work with the data.
Why 78 percent is the limit
78 percent isn’t magic. It’s the accuracy ceiling for the current generation of algorithms that work only with visual data. The next step is incorporating behavioral patterns, communication styles, and value systems. When systems learn to analyze not just how you look, but how you communicate, accuracy will climb to 90 percent and beyond.
Unison is already doing this: their algorithm takes into account not only visual preferences but also thousands of other parameters—from messaging style to decision-making patterns. The result: a conversion rate to real-life meetings of over 20 percent. For comparison, Tinder’s is around 5 percent.
So AI chooses a partner for us?
No, it doesn’t choose. It filters. And that’s a huge difference. Imagine you’re looking for a needle in a haystack. AI doesn’t say: here’s your needle. It says: here are ten straws among which your needle is most likely to be. Instead of going through the whole stack, check these. Everything else is up to you.
MIT Professor Sandra Wachter, a specialist in human-computer interaction, puts it this way: technology doesn’t replace human decision-making. It compresses the search space. Instead of a thousand options, 98 percent of which are obviously incompatible, the algorithm shows you ten, 80 percent of which have real potential. The time saved isn’t measured in hours—it’s measured in months of your life.
The future: from quantity to quality
Venture capital firm Andreessen Horowitz, which invested $20 million in an AI dating startup in 2023, framed its strategy this way: we’re not investing in technology, but in bringing back human connection. It sounds paradoxical, but there’s deep logic to it.
Today’s dating apps make money off your loneliness. The longer you search, the more they earn. That’s a conflict of interest that doesn’t exist in the quality matching model. When an app earns from successful matches rather than time spent inside, its goals align with yours. And that changes everything.
Jake, who started this story, never turned his code into a startup. When asked why, he shrugs: I wrote it for myself. But this engineer from San Francisco, without knowing it, upended an industry worth over five billion dollars. His algorithm is still running—he just doesn’t know it.
Read also: Konrad Lorenz, imprinting discovery, Nobel Prize