Article Title
Predicting Precedent: A Psycholinguistic Artificial Intelligence in the Supreme Court
Abstract
Since the proliferation of analytic methodologies and ‘big data’ in the 1980s, there have been multiple studies claiming to offer consistent predictions for Supreme Court behavior. Political scientists focus on analyzing the ideology of judges, with prediction accuracy as high as 70%. Institutionalists, such as Kaufmann (2019), seek to make predictions on verdicts based on a thorough, qualitative analysis of rules and structures, with predictive accuracy as high as 75%. We argue that a psycholinguistic model utilizing machine learning (SCOTUS_AI) can best predict Court outcomes. Extracting sentiment features from parsed briefs through the Linguistic Inquiry and Word Count (LIWC), our results indicate SCOTUS_AI (AUC = .8087; Top K=.9144) outcompetes traditional analysis in both class-controlled accuracy and range of possible, specific outcomes. Moreover, unlike traditional models, SCOTUS_AI can also predict the procedural outcome of the case as one-hot encoded by remand (AUC=.76). Our findings support a psycholinguistic paradigm of case analysis, suggesting that the framing of arguments is a relatively strong predictor of case results. Finally, we cast predictions for the Supreme Court docket, demonstrating that SCOTUS_AI can be practically deployed in the field for individual cases.