Finkel Reviewed Every Major Dating Algorithm Study. The Verdict Was Blunt

In 2012, a team of psychologists led by Northwestern's Eli Finkel did something almost nobody in the dating industry had bothered to do: they read the actual science behind online dating algorithms, all of it, and asked whether any of it held up. Their paper, published in Psychological Science in the Public Interest, remains the most thorough academic audit of online dating algorithms ever conducted. The conclusion was blunt — no algorithm tested could reliably predict which two people would form a successful relationship, and the mathematical case for why one plausibly could was weak to begin with.

That finding is now over a decade old, but it hasn't been overturned. It's been quietly confirmed by everything that's happened since: bigger datasets, more computing power, and — per the platforms' own disclosures — no meaningful jump in reported relationship outcomes. This is a deep dive into what Finkel's team actually measured, what their review does and doesn't prove, and what it means for anyone currently deciding how to date in 2026.

What the Finkel Review Actually Set Out to Measure

Finkel and his co-authors — Paul Eastwick, Benjamin Karney, Harry Reis, and Susan Sprecher — weren't running a new experiment. They were doing something arguably more valuable: a systematic literature review of every major claim online dating sites made about their online dating algorithms, checked against the existing body of relationship science.

Their scope covered three separate claims dating sites commonly make:

• Access claim: that sites expose you to potential partners you wouldn't otherwise meet.
• Communication claim: that sites let you learn about a partner through messaging before investing in a date.
• Matching claim: that a proprietary algorithm can identify which two people are especially compatible, from a pool of possibilities, better than random chance or self-selection would.

The first two claims held up reasonably well. It's the third one — the algorithm itself, the actual engine sold as the product's core value — that the review dismantled.

The Mathematical Problem With "Compatibility" Before the Data Even Arrives

Before testing any specific algorithm, Finkel's team made a simpler point that's easy to miss: relationship science had already identified which variables predict long-term relationship success, and most of them are not knowable from a profile.

Similarity in attitudes and values matters a little. Personality traits matter a little. But the strongest predictors — how a couple communicates under stress, how they support each other during hardship, the timing and context of when they met — only exist once two people are already interacting. A questionnaire filled out alone, before either party has met the other, structurally cannot capture them.

This is the paper's central mechanical argument: an algorithm fed only individual-level data (what you say you like, how you answer a set of static questions) is trying to predict a property that emerges from interaction, not from two separate profiles stacked next to each other. It's a category error baked into the product before a single match is ever generated.

What the Review Found When It Checked the Actual Evidence

Finkel's team then looked at what evidence the major dating sites had published to support their matching claims at the time. The finding: none of the sites reviewed had published peer-reviewed, methodologically sound evidence that their algorithm-generated couples fared better than couples matched at random from the same pool, or better than couples who met on their own.

Where studies did exist, they suffered from a common flaw: they measured whether people responded well to a match's profile — attraction to a photo, interest in a bio — rather than whether the relationship that followed was actually happier or more durable. Predicting initial interest and predicting long-term compatibility are different problems, and the algorithms were, at best, only modestly useful at the first one.

This distinction matters more than it sounds. An algorithm that reliably produces someone you'll swipe right on is not the same as an algorithm that reliably produces someone you'll still want to be with in three years. The industry has spent the years since optimizing hard for the first metric, because it's the one that's measurable in real time and correlates with engagement. The second metric takes years to observe and doesn't show up on a quarterly earnings call.

What This Does Not Prove

It's worth being precise about the limits of a thirteen-year-old review, because overclaiming here would repeat the industry's own mistake.

• The review does not prove that no algorithm could ever meaningfully assist matching — it argues that the algorithms examined, built primarily to predict initial attraction, weren't built to solve the harder problem.
• It does not prove online dating fails to produce good relationships. Later research — including Michael Rosenfeld's long-run couple tracking, covered in our deep dive on his stability data — finds couples who meet online do about as well as couples who meet any other way.
• It does not distinguish between a purely computational matching engine and a model where a human or AI agent gathers a fuller, ongoing picture of a person over time rather than a one-time questionnaire.

The gap Finkel's team identified is specifically about static, self-report data feeding a black-box score. It's a narrower target than "matchmaking can't work" — and that distinction has only become more relevant with what's launched since.

Why This Finding Held Up for Over a Decade

If the science was this clear in 2012, why did compatibility scores become more prominent, not less, over the following decade? Two structural reasons, both documented independently of Finkel's work.

First, choice overload research — most famously Iyengar and Lepper's original jam study, and later applied directly to dating contexts — shows that people want a number to anchor a decision when facing a large set of options, even when that number carries little predictive weight. A compatibility percentage doesn't need to be accurate to be useful for reducing the anxiety of choosing; it just needs to look authoritative. We've gone deeper on what a compatibility score is actually built from elsewhere.

Second, and more simply: the business model rewarded a different outcome. Match Group's own investor disclosures track "payers" and engagement, not relationship formation. A score that keeps you swiping is commercially useful whether or not it's scientifically valid. We've covered that incentive structure in detail in our look at the Match Group 10-K filings.

What Changed the Conversation This Year

Finkel's critique sat mostly in academic circles for over a decade — cited by researchers, largely ignored by the industry it was reviewing. That changed in July 2026, when Justin McLeod, the founder of Hinge, raised $18 million for Overtone, a new matchmaking service built on the explicit premise that swiping and static compatibility scores were never going to solve this problem. Match Group itself was among the investors.

The significance isn't that Overtone is live — it isn't yet, it's rolling out later this year in select locations. The significance is that the person who spent a decade building and refining one of the industry's most sophisticated matching algorithms is now building something that doesn't use one in the traditional sense at all. That's not a rival's marketing claim about Finkel's finding. That's the finding, agreed with, by the person best positioned to know whether it was true.

Where Agent-Mediated Matching Actually Addresses the Gap

Finkel's specific objection was to a one-time questionnaire feeding a static score. An approach where an AI agent gathers an ongoing, detailed briefing — not five checkbox questions but an actual account of what you want, what's failed before, and what your real constraints are — is a structurally different input, even before any matching happens.

It still can't observe the interaction dynamics Finkel's team identified as the strongest predictors — nothing can, before two people actually meet. But it closes part of the gap by working from richer, ongoing information instead of a frozen profile, and by making fewer, better-reasoned introductions instead of optimizing a score across a large pool. That's the model neverswipe runs on: no swiping, no static score, an agent that explains the reasoning behind each introduction rather than hiding it behind a percentage.

What This Means If You're Deciding How to Date Right Now

The practical takeaway isn't "algorithms are useless, do it yourself." It's narrower and more useful than that:

1. Treat any compatibility percentage as a UI convenience, not a scientific claim — the underlying research never supported that level of precision.
2. Weight your own judgment about interaction — how someone actually communicates on a date — more heavily than any pre-date score, because that's exactly the variable the algorithm can't see.
3. Favor approaches that gather deep, ongoing information about you over ones that reduce you to a handful of static answers, since that's the specific mechanism Finkel's team found lacking.
4. Don't mistake "the algorithm found you a match" for "the algorithm knows you'll be compatible" — those are different claims with very different evidence behind them.

Frequently Asked Questions

What did the Finkel study actually conclude about online dating algorithms?

That no algorithm reviewed had public, methodologically sound evidence it could predict long-term relationship compatibility better than chance, largely because the strongest predictors of relationship success only emerge once two people interact — something a static profile can't capture.

Is the Finkel review out of date?

It was published in 2012, but no major dating platform has since published peer-reviewed evidence overturning its core finding. Subsequent research on choice overload and company disclosures on engagement metrics are consistent with, not contrary to, its conclusions.

Does this mean online dating doesn't work?

No. Separate research, including Rosenfeld's long-run tracking of how couples meet and stay together, finds couples who meet online do about as well as those who meet any other way. Finkel's critique is specifically about the predictive power of matching algorithms, not about online dating as a category.

How is AI matchmaking different from the algorithms Finkel reviewed?

The critique targeted static, one-time questionnaires feeding a compatibility score. Agent-mediated approaches that gather an ongoing, detailed briefing about a person and make a small number of reasoned introductions are working from a different, richer kind of input, though no method can observe interaction dynamics before two people actually meet.

Why does this matter more now than it did in 2012?

Because the founder of Hinge has now built a new matchmaking company premised on the same conclusion Finkel's team reached — that swipe-style matching algorithms were never going to solve compatibility — backed in part by Match Group. The academic argument now has industry agreement from inside the field that built the original model.