Model-based algorithms shape automatic evaluative processing

Computational theories of reinforcement learning suggest that two families of algorithm—“model-based” vs. “model-free”—tightly map onto the classic distinction between automatic and deliberate systems of control: Deliberate evaluative responses are thought to reflect model-based algorithms, which are accurate but computationally expensive, whereas automatic evaluative responses are thought to reflect model-free algorithms, which are error-prone but computationally cheap. This framework has animated research on psychological phenomena ranging from habit formation to social learning, and from moral decision-making to cognitive development. Here we propose that model-based and model-free algorithms may not be as aligned with deliberate and automatic evaluative processing as prevailing theories suggest. Across two preregistered behavioral experiments involving adult human participants (total n = 1,740), we show that model-based algorithms shape not only deliberate but also automatic evaluations. Experiment 1 numerically replicates past findings suggesting that deliberate (but not automatic) evaluative responses are uniquely shaped by model-based algorithms but, critically, also reveals confounds that render interpretation of this evidence equivocal. Experiment 2 eliminates these confounds and reveals robust model-based contributions to automatic evaluative processing. Together, these results suggest that dominant frame-works may drastically underestimate both the ubiquity of model-based algorithms and the computational sophistication of automatic evaluative processing.