What Happens When You Feed 11,000 Relationships Into a Machine Learning Model
In 2020, a team of 86 researchers pooled 43 separate longitudinal datasets — covering more than 11,000 couples — and handed the combined data to a machine learning model with one job: determine which of over 2,000 measured variables actually predict how satisfied two people will be in their relationship. The result was a landmark study in machine learning relationship satisfaction research, and its findings were striking not for what the algorithms discovered, but for what they could not.
Led by psychologists Samantha Joel and Paul Eastwick, the study — published in the journal Proceedings of the National Academy of Sciences — found that individual traits such as personality, attachment style, depression levels, and demographics explained at most 21 percent of variance in relationship satisfaction. Even the strongest relationship-specific predictors, factors measured within the couple dynamic itself rather than in either individual alone, consistently failed to forecast whether a couple's happiness would rise or fall over time. For a field that has spent decades accumulating data on human behaviour, the finding is both sobering and instructive.
Why AI's Failure to Predict Love Is a Critical Lesson for Data Professionals
For developers, data scientists, and IT decision-makers, this study functions as more than a curiosity piece from the social sciences. It cuts to the heart of a fundamental assumption baked into most modern machine learning pipelines: that given enough data and sufficient computational power, patterns will emerge that allow reliable prediction of complex outcomes. The relationship satisfaction study suggests this assumption has hard limits — particularly when the system being modelled is inherently dynamic, contextual, and shaped by variables that resist measurement.
The challenge is not simply one of data volume. The researchers in Joel and Eastwick's collaboration had more than 11,000 couples and over 2,000 variables at their disposal — a dataset that would satisfy most enterprise analytics teams. Yet even with modern machine learning techniques applied to this rich corpus, the ceiling on predictive accuracy remained stubbornly low. As researchers noted in their findings, the dyadic context — the emergent quality of how two specific people interact — accounted for far more variance than any individual characteristic, and it proved the hardest to model.
This resonates with challenges faced across industries. In cybersecurity, for example, behavioural analytics platforms promise to predict insider threats by modelling individual user attributes. In HR technology, algorithmic screening tools claim to forecast employee performance and cultural fit. In financial services, credit models attempt to predict default risk from personal profiles. Each of these applications rests on a similar premise to the one tested by Joel and Eastwick's team — and each faces analogous limitations when the outcome depends fundamentally on relational or contextual dynamics rather than individual traits alone.
"The lesson is not that machine learning is useless — it is that we must be precise about what kind of problems algorithms can and cannot solve. Predicting emergent human behaviour from static individual variables is one of the hardest challenges in data science, and overconfidence here carries real costs."
— Samantha Joel, lead researcher on the relationship satisfaction studyBreaking Down the Numbers: What 2,000 Variables Actually Revealed
Understanding the structure of the study helps contextualise its conclusions. The researchers analysed data from longitudinal studies — meaning couples were surveyed not once but multiple times over extended periods. This is methodologically significant: longitudinal data captures change over time, which is precisely what the model was challenged to predict. Rather than simply measuring how satisfied a couple reported being at a single point, the model was asked to forecast whether that satisfaction would increase or decrease.
| Variable Category | Examples | Predictive Power |
|---|---|---|
| Individual Traits | Personality, age, attachment style, depression levels | Up to 21% of variance |
| Relationship-Specific Factors | Perceived partner responsiveness, conflict patterns, intimacy | Stronger but still limited |
| Dynamic/Emergent Patterns | How satisfaction changes over time, contextual interactions | Most predictive, hardest to model |
| Demographic Variables | Income, education, ethnicity, relationship duration | Low predictive power |
The predictive strength of relationship-specific variables — such as how each partner perceived the other's responsiveness — was higher than individual traits alone. However, even these relational predictors failed to reliably forecast the trajectory of couples' happiness. This suggests that relationship satisfaction is what data scientists would call a high-noise, low-signal environment: the signal is real but so heavily obscured by contextual and temporal complexity that extracting it remains beyond current algorithmic capabilities at scale.
The Broader Algorithmic Limits This Research Exposes
Indicative variance explained (R²) by machine learning across domains — illustrating where predictive AI performs well vs. struggles with complex human behaviours.
The implications of this machine learning relationship satisfaction study extend well beyond romantic partnerships. Research published by the journal Nature Human Behaviour has similarly found that predictive algorithms consistently underperform human expectations when modelling complex social outcomes. The common thread across these failures is the same: real-world human behaviour is shaped by emergent, contextual dynamics that resist reduction to static feature vectors.
For policy professionals and regulators — particularly those working in the European context on AI governance under frameworks like the EU AI Act — this study offers empirical grounding for a critical regulatory argument: algorithmic systems deployed in high-stakes social domains may systematically overstate their own predictive validity. The 21 percent ceiling found by Joel and Eastwick is not just an academic curiosity; it is the kind of finding that should inform how regulators think about mandatory accuracy disclosures for AI systems used in hiring, credit, healthcare triage, and social services.
The EU AI Act, which entered into force in 2024, specifically targets high-risk AI applications in these categories and mandates transparency around performance limitations. As Wired has reported, the broader AI research community has grappled for years with what researchers call the "prediction gap" — the chasm between what practitioners claim algorithms can do and what empirical testing demonstrates. The relationship study is one of the most rigorous demonstrations of that gap to date.
Separating What AI Can Predict From What It Cannot
It is worth being precise about what the study does and does not show. Machine learning models excelled at identifying which variables mattered at all — a valuable function. The algorithms correctly identified that relationship-specific factors outperformed individual traits, which is itself a meaningful finding that can guide future research investment. Where the models failed was in their predictive accuracy for individual couples: knowing that perceived partner responsiveness is a stronger predictor than personality does not mean any specific couple's trajectory can be reliably forecast.
This distinction is important for practitioners. Machine learning can be genuinely powerful for pattern identification, anomaly detection, and population-level statistical inference. Research from Originally reported by Silicon Canals. Summarised and curated by European Purpose.
