Poster at SPP 2026

Abstract

The computational theory of mind defines cognition as information processing occurring within individual brains, establishing a clear boundary between internal mental processes and the external environment. A growing collectivist movement challenges this framework, arguing that human cognition must be understood as a property of social collectives rather than isolated minds. While this movement encompasses diverse theoretical positions, we observe that they can be organized into two distinct arguments. The emergentist argument, exemplified by Hutchins’s distributed cognition framework, holds that social collectives can become genuinely cognitive in ways irreducible to individual minds. The constitutivist argument goes further, claiming that individual cognition is fundamentally constituted by social processes—that separating individual cognition from social interaction is like trying to understand breathing while ignoring air. Both arguments face serious philosophical objections. Critics charge that attributing genuine mental states to collectives commits a category error – what does it mean for a collective to possess cognitive representations reducible to no individual’s mental states? The constitutivist position faces the coupling-constitution fallacy—inferring that because social processes causally influence cognition, they must literally constitute it. Without maintaining this distinction, we lose precise mechanistic explanations of cognitive phenomena. We propose that these debates can be productively reframed. The question is not whether collectives literally possess minds, but whether we can study the psychology of groups at the group level using the same tools that have proven successful for studying individual cognition. This reframing reveals something important – groups exhibit behaviors that simply do not exist at the individual level. An individual cannot exhibit “degree of collaboration” or “division of cognitive labor”—these are inherently collective variables describing how information and effort are distributed across minds. Just as psychology studies individual-level behaviors like reaction time that emerge from neural processes, we can study group-level behaviors like coordination that emerge from individual cognitive processes. And just as computational cognitive science models mechanisms underlying individual behavior, we can model mechanisms underlying collective behavior—applying the same resource-rational framework and experimental rigor. Memory provides an ideal domain to illustrate this approach. Transactive memory systems—where groups collectively encode and retrieve knowledge through a division of cognitive labor—have long featured in collectivist arguments. Yet memory also has well-established computational models at the individual level, with capacity limitations that are precisely characterized. This makes memory a domain where we can rigorously quantify when and how collaboration helps, and model the collective’s “decision” to distribute information using the same resource-rational principles applied to individual memory allocation. We illustrate this with an experimental study of collaborative visual working memory. Participants memorized grids of 4, 16, or 36 images both alone and with a partner. Each trial consisted of a 10-second encoding phase followed by a retrieval test on a randomly cued location of the grid. In dyadic trials, participants could see which images their partner was currently studying via a visual indicator, enabling tacit coordination without explicit verbal strategy. Participants were rewarded based on the first response submitted by either dyad member, incentivizing true collaboration. We analyzed this task at the collective level – the dyad faces a memory problem, and we ask how it allocates combined cognitive resources. We developed a computational model adapted from signal detection theory of working memory which assumes each individual must trade-off between remembering a few items really well and remembering many items less well. When individuals collaborate optimally by dividing the grid, each can allocate more memory resource to encode each item. The model predicts that collaboration benefit—the performance gain from dividing versus independent encoding—peaks at intermediate grid sizes. Small grids offer little benefit because individual capacity suffices; very large grids offer diminishing returns because even distributed resources remain insufficient. We operationalized degree of collaboration as how much time dyad members spent studying the same location, which should decrease relative to solo trials, if dyads divide the grid. Results confirmed our predictions. Participants spontaneously reduced spatial overlap for medium and large grid. No such coordination emerged for small grids where individual performance was already at ceiling. Notably, participants collaborated even for very large grids where our model predicted minimal benefit, suggesting an interesting gap between the normative computational model and the behavior of actual human dyads that awaits further exploration.This research program demonstrates that collective cognition is amenable to the same scientific approach as individual cognition – define group-level behaviors, build computational models predicting when they emerge, and test predictions experimentally. This points to an exiting future direction of studying interacting groups as their own emergent cognitive systems.