Explainable AI as a Bridge Between Machine Learning and the Social Sciences

In our recent paper, An Explainable Artificial Intelligence Handbook for Psychologists: Methods, Opportunities, and Challenges (Psychological Methods, 2025), my co-authors and I set out to provide psychologists with a practical guide to explainable artificial intelligence (XAI). While we wrote this with psychologists in mind (being my background), I believe the lessons extend much further—to anyone in the social sciences grappling with modelling subjective, highly collinear, human data using machine learning models.

Machine learning is increasingly used across disciplines to analyze large and messy datasets. Economists use machine learning to predict labor market outcomes, political scientists to understand voting patterns, sociologists to model social networks, and education researchers to forecast student achievement. These models are powerful, but they are also often “black boxes”. They produce accurate predictions, but leave us guessing about why those predictions were made.

XAI offers a way forward. Methods such as SHAP, LIME, partial dependence plots, and accumulated local effects (among others) allow researchers to open up these black boxes and examine which variables matter, how they interact, and whether models are embedding biases. Importantly, XAI can help us move beyond pure prediction to generating insights that connect machine learning back to theory or policy.

But our paper also highlights the challenges. XAI methods were not designed to uncover the “true” causal structure of social phenomena, and if misapplied, they can be dangerously misleading. Many applying machine learning haven’t thought beyond using XAI to understand the model they have built, to improve it. While as social scientists, we are using these models to try and understand the phenomena behind the data—let’s not forget how different these aims are. These methods are typically designed with the former in mind, making them suboptimal for the goals of social scientists. 

Emerging issues using XAI for understanding, such as multicollinearity, spurious variable importance caused by imputations (which we will unpack further in a future paper currently in preparation), or over-interpretation of predictors, are not unique to psychology—they are concerns for any field using XAI to understand the patterns that exist in observational data.

The broader message is this: explainable AI is not just a technical add-on. It has the potential to be a methodological bridge between predictive algorithms and the interpretive traditions of the social sciences. If used thoughtfully, it can help us harness the power of machine learning to understand phenomena, to question theory, and think inventively—generating new hypotheses or research questions. In an applied world, it can increase transparency and identify bias.

My hope is that this work encourages not just psychologists, but all researchers using machine learning to understand human behavior, at both the individual and collective levels, to engage with XAI—both its possibilities and its pitfalls. To interrogate the insights gleaned in relation to theory, not least common sense, and contrast them with the findings from laboratory experiments, randomised controlled trials, and natural/field studies, to provide a holistic investigation of the complexities of human behaviour.

An overview of different XAI methods that are useful to social scientists.