15 - Scepticism about Big Data’s Predictive Power about Human Behaviour: Making a Case for Theory and Simplicity

Summary

A core claim of big-data-algorithm enthusiasts – producers, champions, consumers – is that big-data algorithms are able to deliver insightful and accurate predictions about human behaviour. This chapter challenges this claim. I make three contributions: First, I perform a conceptual analysis and argue that big-data analytics is by design a-theoretical and does not provide process-based explanations of human behaviour, making it unfit to support insight and deliberation, which is transparent to both legal experts and non-experts. Second, I review empirical evidence from dozens of data sets, which suggests that the predictive accuracy of mathematically sophisticated algorithms is not consistently higher than that of simple rules (rules that tap on available domain knowledge or observed human decision-making); rather, big-data algorithms are less accurate across a range of problems, including predicting election results and criminal profiling (this work presented here refer to understanding and predicting human behaviour in legal and regulatory contexts). Third, I synthesize the above points in order to conclude that simple, process-based, domain-grounded theories of human behaviour should be put forth as benchmarks, which big-data algorithms, if they are to be considered as tools for personalization, should match in terms of transparency and accuracy.