Understanding as Designed Projection

Computational models can predict and generalize while leaving mechanistic questions unanswered. In their recent commentary, Shiffrin et al. argue that this gap gives rise to “illusions of understanding” in sciences, particularly when successful inference is treated as explanation. In this commentary, I focus on one source of such illusions emphasized by the authors: the difficulty of deduction. The authors propose that such illusions arise because deductions, even when correctly formalized, are often difficult to understand in probabilistic, chaotic, or high-dimensional system. Here I reconsider this conclusion by arguing that this limitation is better understood as a consequence of a designed projection under partial observability. Scientific measurement relies on compressing complex phenomena through experimental design, creating an information bottleneck. From an information-theoretic perspective, no analysis can recover structure that was never measured. Drawing on recent work on epiplexity, I further argue that understanding emerges from the extraction of learnable structure from data, shaped by both data collection and analysis choices. In this view, scientific progress is driven by improving and switching projection spaces through new measurements, task designs, and models that make previously inaccessible structure learnable.