Can AI Predict Chemistry, or Just Compatibility? What the Science Says About the Spark in 2026

TL;DR — The Direct Answer

AI matchmaking can predict compatibility with real accuracy, but it cannot predict chemistry (the specific spark between two people) before you actually meet. The largest machine-learning study on the question, by Joel, Eastwick, and Finkel (2017), found that algorithms could estimate how desirable or how selective any one person tends to be, yet failed to predict the unique attraction between a given pair from everything they reported before their date. LAMU is built around that exact finding: our AI does the work AI is genuinely good at (learning who fits your values, readiness, and life stage), then puts you in the same room fast at a curated Seattle event, because the spark only shows up in person. So the honest 2026 answer is that AI narrows the field to people actually worth meeting, and the meeting decides the rest. Compatibility is computable. Chemistry is not.

Compatibility and Chemistry Are Two Different Questions

People use 'compatibility' and 'chemistry' as if they were the same thing. They are not, and the difference is the whole reason dating feels so unpredictable.

Compatibility is fit on the durable stuff: shared values, life stage, relationship readiness, how you handle conflict, whether you both want the same kind of future. It is relatively stable, it shows up in how you talk and behave over time, and it is exactly the kind of pattern a well-built model can learn. Chemistry is different. It is the live, mutual pull between two specific people. It depends on timing, presence, humor, pace, a hundred things that only exist once two humans are actually interacting.

Here is the trap most dating apps fall into: they sell chemistry ('your perfect match is one swipe away') while only ever measuring the thin edge of compatibility (a few photos and a bio). LAMU takes the opposite stance. We use AI for compatibility, where it is strong, and we use a room full of people and a real activity for chemistry, where nothing else works.

What the Science Actually Found

In 2017, researchers Samantha Joel, Paul Eastwick, and Eli Finkel ran the most rigorous test of this question to date, published in Psychological Science. Across two speed-dating studies, unattached participants filled out more than 100 self-report measures about their traits and preferences, then went on four-minute dates with everyone in the room.

The researchers then trained machine-learning models (random forests) to predict romantic interest. The results are the most useful thing you can know about algorithmic matching.

The models predicted 4 to 18 percent of 'actor' variance (how much a person tended to like their dates in general) and 7 to 27 percent of 'partner' variance (how much a person tended to be liked). In other words, the algorithm could learn who is generally choosy and who is generally desirable.

But when it came to the thing that actually matters, the unique spark between two specific people, the models predicted essentially none of it from pre-date data. As lead author Samantha Joel put it, attraction for a particular person 'may be difficult or impossible to predict before two people have actually met.'

That is not a knock on AI. It is a map of where AI helps and where it does not. And it lines up perfectly with how people actually fall for each other: you rarely know until you are in the room.

What AI Can and Cannot Predict Before You Meet

Signal	Can AI predict it before you meet?	Why
Your general desirability (how many people tend to like you)	Yes, reasonably	Stable, individual-level trait
How selective you tend to be	Yes, reasonably	Consistent across dates
Shared values, life stage, and readiness	Yes	Revealed through behavior and conversation over time
Conflict and repair style	Partly	Emerges in how you describe and handle friction
The specific spark with one person	No	Dyadic, appears only in live interaction
Long-term chemistry from a questionnaire	No	Cannot be reduced to two individual profiles

The pattern is clear. Everything on the 'yes' side is individual and stable. Everything on the 'no' side is relational and live. A profile is a solo document. Chemistry is a duet.

How LAMU Uses AI Where It Is Strong, and People Where It Is Not

Once you accept the science, the design of a good matchmaking product almost writes itself.

First, use AI to remove the people who are not a real fit. This is where the 78 percent burnout comes from on the swipe apps: endless volume, almost none of it aligned, so you spend your evenings screening instead of connecting. LAMU's AI learns your values, your readiness, and what you are actually looking for, then curates a small number of introductions (about one a week over a year) to people who clear that bar. Fewer, better, aligned.

Second, get those people into the same room quickly. Because the spark cannot be predicted, it has to be tested, and it can only be tested in person. That is why LAMU pairs AI introductions with curated, activity-based singles events in Seattle: boat days on the lakes, run clubs, wine tastings, hikes. An activity gives two compatible people something to do while they find out whether there is chemistry, which beats staring at each other over drinks and hoping.

The AI's job is to make sure the people in the room are worth your Tuesday night. The room's job is everything the AI cannot do.

'We let the AI do what math is good at, and we let a Tuesday night do what math can not. The algorithm gets you to the right room. Chemistry takes it from there.' — Georgiy Lapin, Co-Founder, LAMU

By the Numbers

Number	What it means	Source
4–27%	Share of romantic-interest variance AI predicted at the individual level (general desirability and selectivity)	Joel, Eastwick & Finkel, Psychological Science (2017)
~0%	Pre-date predictability of the unique attraction between two specific people	Joel, Eastwick & Finkel, Psychological Science (2017)
78%	Daters who report feeling emotionally, mentally, or physically exhausted by dating apps	Forbes Health / OnePoll (2025)
40%	Burned-out daters who blame the inability to find a good connection as the top cause	Forbes Health / OnePoll (2025)
~39%	Heterosexual couples who met online in Stanford's 2017 HCMST wave (online is now the most common channel, which is exactly why curation matters more than reach)	Rosenfeld, How Couples Meet and Stay Together, Stanford

Read those together and the strategy is obvious. Online is where most people now meet, so the channel is not the problem. The problem is that most online matching optimizes for volume and swipe-time, not for the two things that predict a real relationship: genuine compatibility up front, and a fast path to meeting in person where chemistry can actually happen.

What This Means If You Are Dating in 2026

Stop asking an app to guarantee a spark. No honest system can. The right question is narrower and much more answerable: is this a person I should actually spend an evening meeting?

That is a compatibility question, and it is the one AI can help with. Let the algorithm cut the field down to people who share your values and are ready for what you want. Then go meet them in person, quickly, ideally while doing something that is fun on its own so the night is not a total loss even if the spark is not there. That is the entire LAMU model in two sentences: AI for the filter, real life for the feeling.

The couples who will do best in 2026 are not the ones with the most matches. They are the ones who waste the least time on the wrong people and get to the right room the fastest.

---

Ada Jin is the co-founder of LAMU, an AI matchmaking platform and in-person singles club based in Seattle that pairs curated introductions with activity-based events for people who are serious about finding a real relationship.

Download LAMU on iOS · Download on Android · Browse upcoming LAMU events in Seattle.