Machine learning cannot predict chemistry. The same method predicts relationship quality.

Before anyone met anyone, the researchers collected more than 100 self-reported traits and preferences from every participant. Age, values, attachment style, what they said they wanted in a partner, what they said they could not stand. Then they sent everyone on a series of four-minute dates and asked a machine learning model one question: which of these people will want each other?

On that question, the model returned nothing.

The study, and what it was actually testing

The paper is Is Romantic Desire Predictable? Machine Learning Applied to Initial Romantic Attraction, by Samantha Joel, Paul Eastwick and Eli Finkel, published in Psychological Science in 2017. Two speed-dating studies, unattached participants, more than 100 measures each, and a round-robin design where every person met every opposite-sex person for four minutes.

The clever part is how they split the outcome into three pieces, because the split is what makes the result mean something.

Actor variance is your general tendency to like people. Some of us walk into a room predisposed to be interested; some of us are harder to impress.

Partner variance is your general tendency to be liked. Some people are widely desired, and this is not a mystery to anyone who has been to a party.

Relationship variance is the part everyone actually cares about. It is how much a specific person wants a specific other person, over and above those two general tendencies. It is the word chemistry, given a number.

The models did fine on the first two. Random forests predicted 4 to 18 percent of actor variance and 7 to 27 percent of partner variance. Modest, but real, and in the direction you would expect.

On relationship variance they were, in the authors' words, "unable to predict relationship variance using any combination of traits and preferences reported before the dates."

Not weakly. Not with a small effect that needs a bigger sample. They had over a hundred variables chosen because decades of prior research said they mattered for mate selection, and the specific-pair signal was not there to find.

They had more than a hundred variables, chosen because decades of research said they mattered, and the chemistry signal was not there to find.

Be precise about what this does and does not show

Four limits, stated plainly, because a study is only useful if you know its edges.

It measured initial attraction after four minutes, not whether two people would build a good life together. Those are different questions and this one answers the first.

It tested self-reported traits and preferences collected in advance. It did not test whether anything is predictable from how two people actually behave once they are in a room. That is a live question and this study leaves it open.

It was speed dating, a format with its own peculiarities: brief, structured, and with everyone knowing the clock is running.

And a null result is a harder thing to interpret than a positive one. The honest reading is that this particular kind of information, gathered this way, does not carry the signal, rather than that the signal cannot exist anywhere.

One more thing worth knowing, and it cuts in the study's favor. The paper was later re-examined by an independent auditor, Florian Pargent, under the ERROR program, which pays reviewers to hunt for mistakes in published work. He found some: a couple of transcription slips, undocumented handling of missing data, and methodological choices he would have made differently. The verdict was that none of these affect the core conclusions and that no correction was recommended. What protected the paper was that the authors had validated their models on an independent sample rather than trusting the first fit.

To be clear about what that audit is and is not, since it is doing real work in this post: an ERROR report is a commissioned error check published by the program, not a peer-reviewed reanalysis and not a replication. Nobody re-ran the experiment. What it establishes is that someone was paid to find problems in this specific paper, looked hard, and did not find any that moved the result.

The claim this lands on

There is a specific promise buried in the phrase "our matching algorithm," and it is worth saying out loud, because once it is out loud you can check it. The promise is that a form you fill in before meeting anyone can predict who you will fall for.

That promise had already been examined. In 2012 the same Eastwick and Finkel, with Benjamin Karney, Harry Reis and Susan Sprecher, published a 64-page review of online dating in Psychological Science in the Public Interest. Their assessment of the matching claims was blunt: there was no compelling evidence that any online dating matching algorithm actually worked. Finkel's diagnosis of why is the useful part. Developers, he argued, built on the information easiest for them to collect, similarity in personality and attitudes, rather than on what relationship science had found actually predicts how a relationship goes.

The 2017 study is what happens when you take that diagnosis seriously and test it properly. Collect the easy information, all of it, more than a hundred measures. Run a method designed to find patterns humans would miss. Ask it the specific-pair question.

Nothing.

The same researcher found the opposite, three years later

Here is the part that changes this from a debunking into a direction.

In 2020, Samantha Joel published a second study, in the Proceedings of the National Academy of Sciences, pooling 43 longitudinal studies and more than 11,000 couples. Same lead author. Same statistical method, random forests. Different question: not who will want whom, but what predicts whether a relationship is good once it exists.

That time the method worked. The qualities of the relationship itself predicted satisfaction roughly two to three times better than any facts about the individuals in it. We went through that study in detail in what makes a relationship last.

Put the two side by side, because almost nobody does.

The same researcher, using the same tool, found that the moment of attraction between two specific people is not predictable from anything you can put on a form, and that the quality of a relationship between two specific people is quite predictable from what happens inside it.

Before the match: noise. After it: signal.

What an app can honestly claim

None of this means dating apps are useless, and Joel said so herself. Her read was that sites are genuinely valuable for narrowing the field, but that they "don't let you bypass the process of having to physically meet someone."

That is a real service and it is not a small one. Introducing you to people you would never otherwise cross paths with is most of the value online dating has ever delivered. Filters that respect what you actually need are part of that. So is not wasting your evening.

What an app cannot honestly claim is the other thing, the part the marketing leans on hardest: that its model knows, in advance, which two people will click. On the current evidence nobody knows that, including us.

We think the category has this exactly backwards, and the reason is not mysterious. Selling prediction is how you justify a subscription for the search. If the promise is that a better algorithm is one payment away, the search is the product, and the search never has to end. We wrote about where that incentive leads in why most dating apps make money when you stay single and about how the resulting feed makes people feel in why a feed of strangers makes you feel like inventory.

The evidence points somewhere else. If the predictable part of a relationship is the part that happens after two people meet, then the honest place to build is after the match. That is the whole reason Bloom spans the relationship rather than stopping at the introduction.

What to do with this on a Tuesday night

Three things follow, and none of them require trusting us.

Stop optimizing your profile for a machine. There is no model reading it that can tell whether you and a particular person will click, because that model does not exist. Write for the person, not the sorter.

Meet sooner. If the signal only appears in the room, weeks of texting are not gathering evidence, they are postponing it. This is the practical upshot of the whole literature.

Judge the relationship, not the résumé. The thing that turned out to be predictable was how two people treat each other once they are together: commitment, appreciation, whether each believes the other is staying. Those are things you can only observe, and things you can build.

Chemistry stayed unpredictable across a hundred variables and two studies. What people do with it afterward did not.

Bloom is pre-launch, starting in Utah along the Wasatch Front. The only thing to do today is join the waitlist and grow with us.