Dating Algorithms: The Compatibility Myth the Industry Won't Admit

The DII Take

The dating industry has a compatibility problem, and it is not the one it thinks. The problem is not that algorithms are bad at matching. The problem is that the industry has built its value proposition around a capability that relationship science says is structurally impossible to deliver. The strongest predictors of long-term relationship success - interaction quality, conflict resolution style, responsiveness under stress - are emergent properties of a specific relationship between two specific people. They cannot be predicted from individual profiles, no matter how detailed.

What algorithms can do is filter for baseline compatibility (age, location, dealbreakers) and predict initial attraction. These are useful functions, but they are modest ones, and the gap between what algorithms actually deliver and what marketing promises creates the disillusionment that drives app fatigue. The industry would be better served by honestly positioning algorithms as introduction tools rather than compatibility engines.

What the Research Actually Shows

The academic literature on romantic compatibility prediction is remarkably consistent in its central finding: pre-meeting characteristics are weak predictors of relationship outcomes. Finkel et al.'s 2012 review examined the full body of evidence available at the time and identified a critical disconnect. Dating platforms' algorithms focus on variables that are easy to measure before two people meet: personality traits, attitudes, interests, demographics, and stated preferences.

Relationship science, by contrast, has identified that the strongest predictors of relationship wellbeing - communication quality, responsiveness, the ability to navigate stressful circumstances together, and the ratio of positive to negative interactions (John Gottman's research suggests a 5:1 ratio predicts stability) - are properties of the relationship itself, not of the individuals within it.

The strongest predictors of long-term relationship success are emergent properties of a specific relationship between two specific people. They cannot be predicted from individual profiles, no matter how detailed.

A 2017 study by Samantha Joel, Paul Eastwick, and Eli Finkel, published in Psychological Science, applied machine learning to data from speed-dating interactions. Using over 100 variables about each participant, the researchers attempted to predict romantic desire. The results were sobering: individual-level variables (what a person is like) predicted some variance in general desirability (how attractive someone is on average to everyone), but essentially zero variance in unique compatibility (why person A is attracted to person B specifically). In the language of the researchers, romantic desire was dominated by 'relationship variance' - the idiosyncratic chemistry between two specific people - rather than 'actor' or 'partner' variance.

A massive 2024 registered report by Eastwick, Sparks, Finkel and colleagues, published in the Journal of Personality and Social Psychology, tested the predictive validity of ideal partner preference matching across 43 countries with over 10,000 participants. The study found that while preference-matching effects were statistically significant, they were small in magnitude. People do, to a modest degree, evaluate partners more favourably when those partners match their stated ideals. But the effect sizes suggest that ideal-preference matching explains a tiny fraction of the variance in romantic outcomes - far less than the dating industry's marketing implies.

The similarity question is equally deflating for algorithm designers. Popular wisdom and much of the dating industry assumes that similarity breeds attraction. Research by Tidwell, Eastwick, and Finkel (2013) distinguished between perceived similarity (believing a potential partner is similar to you) and actual similarity (objectively measured overlap in traits and attitudes). Perceived similarity strongly predicted attraction. Actual similarity, the kind that algorithms can measure, did not.

Research Finding	Implication for Algorithms	Key Study
Pre-meeting traits weakly predict relationship outcomes	Algorithms cannot reliably predict long-term compatibility	Finkel et al., 2012
Romantic desire dominated by 'relationship variance'	Unique chemistry is unpredictable from individual data	Joel, Eastwick & Finkel, 2017
Ideal preference matching has small effect sizes	Matching on stated preferences has limited value	Eastwick, Sparks, Finkel et al., 2024
Perceived, not actual, similarity predicts attraction	Algorithm-measurable similarity is a weak predictor	Tidwell, Eastwick & Finkel, 2013
Interaction quality predicts relationship success	Key predictors are emergent and unmeasurable pre-meeting	Gottman research programme

What Algorithms Can Actually Do

The research does not suggest that algorithms are useless. It suggests they are useful for different purposes than the ones typically marketed. Filtering is the most valuable algorithmic function. Eliminating clearly incompatible matches (based on dealbreakers, distance, age range, and non-negotiable preferences) reduces the search space and saves users time. This is genuinely useful, but it is a logistics function, not a compatibility function. No dating platform has ever marketed itself as 'a really good filter', yet that is the function that algorithms perform most effectively.

Predicting general desirability is another legitimate algorithmic capability. Machine learning can identify which profiles will be attractive to many people (based on photo quality, profile completeness, response rates, and similar signals). Hinge's 'Most Compatible' feature, powered by the Gale-Shapley algorithm, optimises for mutual interest probability rather than deep compatibility - a more honest framing that aligns with what the science supports.

Hinge's 2025 algorithm updates, which use deep learning to predict mutual compatibility and have reportedly produced double-digit increases in match rates, represent the industry's most sophisticated current approach. Notably, Hinge describes the improvement in terms of more matches and better recommendations, not in terms of predicting relationship outcomes. This distinction matters: better initial matching is a defensible claim; predicting relationship success is not.

No dating platform has ever marketed itself as 'a really good filter', yet that is the function that algorithms perform most effectively.

Facilitating interaction quality is where the real opportunity lies. If the strongest predictors of relationship success are emergent interaction properties, then platform design should focus on creating conditions for high-quality early interactions rather than predicting compatibility from profiles. Features like voice notes (which Hinge has promoted), video dates, structured conversation prompts, and Esther Perel's collaborative prompt collection (launched on Hinge in June 2025) all move in this direction. They help users have better first interactions, which is where genuine compatibility signals actually emerge.

The AI Matching Promise

The advent of AI-powered matching has reinvigorated compatibility claims. Match Group's AI assistant, announced in early 2025, uses voice interviews and behavioural analysis to recommend matches. Newer platforms like Fate and Volar market 'agentic AI' matching that goes beyond traditional algorithms. Hinge's 2025 deep learning updates reportedly produced double-digit increases in match rates.

The scientific question remains the same: can any system, however sophisticated, predict the emergent chemistry between two specific people before they interact? The research to date suggests the answer is no, or at least not to a commercially meaningful degree. AI may improve filtering precision, surface non-obvious patterns in preference data, and enhance the quality of early interactions through coaching and conversation support. These are genuine advances. But they are advances in introduction quality, not compatibility prediction.

A 2025 study by Eastwick and colleagues at UC Davis, building on the 2017 Joel et al. machine learning work, explored whether natural language processing of messaging data could predict relationship outcomes better than profile data. Preliminary findings suggest that interaction-level data does contain more predictive signal than pre-interaction data, supporting the theoretical argument that compatibility is an emergent property of specific relationships. If this research direction matures, the implication for platforms is clear: the data that matters most for matching quality is generated after the match, not before it. Platforms that collect and analyse interaction data systematically will build progressively better recommendation systems.

The risk for the industry is that AI matching claims repeat the cycle of overpromise and disillusionment that characterised the first generation of compatibility algorithms. If users are told that AI can find their ideal partner and the experience falls short, the trust deficit deepens. If users are told that AI can help them have better first dates with more compatible initial matches, the promise is modest, deliverable, and commercially sufficient.

Research by Finkel and Eastwick on virtual versus in-person first dates (published 2023) found that virtual dates were surprisingly comparable to in-person dates across multiple outcome measures when controlling for date length. In-person dates were generally longer, but the quality of interaction per unit of time was similar. This finding is relevant to the AI matching discussion because it suggests that the medium of interaction matters less than the quality of the interaction itself. AI tools that improve interaction quality - through better conversation prompts, compatibility-relevant questions, and real-time coaching - may produce more meaningful improvements than AI tools that attempt to pre-select compatible pairs.

What This Means for Product Design

The relationship science literature points toward several product design principles that the industry has been slow to adopt. Design for interaction, not prediction. Platform features should maximise the quality and depth of early interactions rather than attempting to pre-select compatible pairs. Voice-first features, video introductions, structured conversation starters, and facilitated first-date experiences all create conditions for genuine compatibility signals to emerge.

Reduce choice, do not expand it. The paradox of choice literature (discussed in DII's separate analysis of Schwartz's theory applied to dating) suggests that more options reduce satisfaction. Algorithms that curate small, high-quality sets of recommendations outperform those that present unlimited options, even if the curated set's 'compatibility' predictions are imperfect.

Measure outcomes, not just activity. Hinge's 'We Met' feature, which allows users to confirm whether a match led to a real-world date, represents the most important data feedback loop in the industry. Platforms that systematically collect post-match outcome data can refine their recommendations in ways that move closer to genuine compatibility prediction over time, even if the theoretical ceiling for such prediction remains low.

Users do not need to believe that an algorithm has found their soulmate in order to pay for a dating subscription. They need to believe that the platform will introduce them to people they would not otherwise have met, that those introductions will be filtered for basic compatibility, and that the platform will facilitate a high-quality interaction.

Be honest about what algorithms do. The platforms that frame their algorithms as sophisticated introduction tools rather than compatibility engines will build more sustainable user trust than those that perpetuate scientifically unsupported claims. The commercial implications of this honesty are not as threatening as operators might fear. Users do not need to believe that an algorithm has found their soulmate in order to pay for a dating subscription. They need to believe that the platform will introduce them to people they would not otherwise have met, that those introductions will be filtered for basic compatibility, and that the platform will facilitate a high-quality interaction.

These are promises the science supports, and they are commercially sufficient to justify premium pricing. The gap between honest positioning and current marketing is not a gap between 'good product' and 'bad product' but between sustainable trust and eroding credibility. In an industry where user trust is already fragile, as documented in DII's consumer insights coverage, the honest positioning may prove to be the more commercially valuable one.

More articles by DII Editorial