Matchmaking software can sort thousands of profiles faster than any person. That does not mean it can identify your future partner with scientific certainty. “Compatibility” may describe a shared-interest filter, a behavioural ranking, a questionnaire score or a marketing label—and those are not interchangeable.
The useful version of synthetic chemistry is modest: help people discover relevant possibilities and explain enough of the process for them to keep control. The risky version turns incomplete signals into a destiny score.
What a matching model actually sees
A system may receive stated preferences, profile text, answers to prompts and behavioural signals such as likes or skips. It can look for patterns linking those inputs with an outcome chosen by its designers. The outcome might be a click, a reply, a longer conversation or a reported match. Each target rewards different behaviour.
A model optimised for replies is not automatically optimised for respect, safety or long-term satisfaction. It may also confuse platform activity with romantic success: someone who opens an app frequently produces more observable data, but that does not make their relationships healthier. Cupid still has not earned a machine-learning certification.
Research puts a limit on the crystal ball
A 2017 study in Psychological Science applied random-forest models to data from two speed-dating studies. Romantically unattached participants completed more than 100 self-report measures before four-minute dates. The models predicted some variation in people’s general tendency to desire others and to be desired, but they could not predict the relationship-specific desire one participant would feel for a particular partner before the two met.
That finding is not proof that every contemporary recommender is useless. The study concerned initial attraction in structured, opposite-sex speed-dating events, not every orientation, culture, online conversation or long-term relationship. Its practical warning is that knowing many traits about two people does not necessarily reveal the chemistry unique to their encounter.
AI can also influence how people present themselves. In a small OzCHI experiment, 48 participants aged 23 to 33 rated ten fictional dating profiles. Believing that AI helped write the text did not significantly change attractiveness ratings, but it lowered perceived trustworthiness. Because this was a convenience sample evaluating fictional text in a survey, it should not be used to predict how all daters will behave. It does support a simple design principle: assistance and authorship should be understandable.
Seven questions for a compatibility claim
- What is being predicted? A click, a reply and a satisfying relationship are different outcomes.
- What information is used? Look for a clear account of profile, questionnaire and behavioural inputs.
- Can inputs be corrected? A mistaken inference should not quietly shape every future recommendation.
- Is the score explained? “92% compatible” is unhelpful without a definition, comparison group and known limitations.
- Was the system evaluated? Ask whether evidence comes from offline tests, live behaviour, user surveys or independent research.
- Who may be missing? Training or evaluation data may not represent every age, culture, gender, orientation or relationship goal.
- Can you choose another route? Manual browsing and editable filters should remain available.
The voluntary NIST AI Risk Management Framework recommends treating AI risk as an ongoing process of governing, mapping, measuring and managing. It does not certify a matchmaking service. It does, however, offer a good test for ambitious claims: are limitations monitored, affected people considered and failures correctable?
Use rankings as introductions, not verdicts
A recommendation can be valuable without being prophetic. Let it widen discovery, then use conversation to test pace, humour, boundaries and mutual interest. Do not interpret a high score as consent, identity verification or a safety check. Those require separate evidence and human judgment.
Our guide to AI dating signals explains how feedback loops can misread behaviour. For a repeatable product assessment, follow the EmberGF review method. And if a system uses emotionally persuasive language, read When a Bot Says “I Missed You” before treating simulated warmth as proof of mutual feeling.
Synthetic chemistry is most credible when it admits what it cannot know. A model can rank an introduction; only the people involved can discover what happens next.