Distributed cognition: the system, not the skull, is the unit of analysis

Cognitive science normally takes the individual agent as its unit of analysis. This third of ten units leaves the skull, since distributed cognition moves that unit. Hutchins analysed a memory task in the cockpit of a commercial airliner. He concluded that the cognitive properties of a socio-technical system can differ radically from those of the individuals who inhabit it. His study of ship navigation argued that cultural activity systems have cognitive properties of their own (Hutchins, 1995).

The claim is really two claims under one name, and their standings differ. The first is empirical and local. What a system can do is not determined by what its members can do, so competence must be described at the level of the arrangement. The second is philosophical: that a reliably coupled external resource is literally part of a cognitive process. Clark and Chalmers (1998) argue for that reading, and Adams and Aizawa (2001) reject it. This unit argues that the design value comes from the first claim and the representational analysis built on it (Zhang and Norman, 1994; Scaife and Rogers, 1996). The second claim remains unresolved.

A system's competence is not the sum of its members

A crew must compute and hold the airspeeds and wing-configuration speeds that apply during an approach. The speed card booklet serves as the system's long-term memory. The speed bugs, the movable markers on the airspeed dial, are not aids to an individual's remembering. They are part of the process by which the cockpit remembers, with the speeds read, written and compared across several representations (Hutchins, 1995). Remove a representation from that arrangement and you change what the system remembers, not merely how clearly it shows it.

Cognition in the Wild makes the argument at book length through the navigation of a naval vessel, where cultural activity systems again show cognitive properties of their own (Hutchins, 1995).

The framework widened until its unit would not stay still

Hollan, Hutchins and Kirsh (2000) turned the finding into a programme for interaction design. They propose distributed cognition as a foundation for HCI. Its focus has always been on whole environments and on how activity is coordinated in them, and the authors claim it is specifically tailored to interactions among people and technologies.

Hutchins (2010) names cognitive ecology as the study of cognitive phenomena in context. He records a movement from units defined by inherent properties of elements to units defined by correlation across elements. An ecosystem is harder to bound than a flight deck, and a framework whose unit will not stay still is harder to test.

The design leverage lies in the arrangement of representations

The most portable part of the framework is a claim about representations rather than about minds. Zhang and Norman (1994) set out distributed representations, in which internal and external representations together represent the abstract structure of a distributed cognitive task. They set out a method of representational analysis that decomposes a hierarchical task into levels, so that each level's properties can be examined independently. Four experiments applied the method to the Tower of Hanoi. If difficulty moves with the representation, then the representation is a design variable rather than a matter of display taste. A screen reader is a representational question here, not an add-on. Zhang and Norman (1994) built the comparison themselves. They made three versions of the Tower of Hanoi that shared one abstract structure and divided its rules differently between person and material. In one, all three rules had to be remembered. In the standard version the discs' sizes enforced one. In a third the cups were filled with coffee, so a smaller cup could not sit on a larger one without spilling, which enforced two. Difficulty fell as more rules moved outside the head.

Scaife and Rogers (1996) give that consequence a mechanism. They characterise representations through computational offloading, the extent to which a representation reduces the cognitive effort a task requires. They analyse it through re-representation, how representations with the same abstract structure differ in the difficulty they impose, and graphical constraining, how graphical elements limit the inferences drawn about what is represented. Neither source needs the extended-mind thesis: both show that the arrangement of representations changes what a task demands.

The framework became an analytic model, not a predictor

Rogers and Ellis (1994) argued that frameworks drawn separately from cognitive, social and organisational disciplines were inadequate for collaborative activity in situ. Even seemingly discrete individual activities draw on all three at once, and organisational constraints act back on individual cognitive processes. They propose distributed cognition as the alternative, with applied studies of computer-mediated work, the framework's bridge into HCI research on collaboration.

Wright, Fields and Harrison (2000) identified the obstacle to using it as a model. Distributed-cognition research, they observe, treats resources for action as central to the interaction between people and technologies. Their resources model answers with a limited number of resource types, abstract information structures that can be represented differently in an interface. It also answers with the concept of an interaction strategy, which describes how different configurations of resources shape what users do. The result is a vocabulary for analysis, not a generator of quantitative predictions. Wright, Fields and Harrison (2000) define six abstract information structures: plans, goals, possibilities, history, action-effect relations and states. Each interaction strategy needs a different configuration of them.

Contested question: does coupling constitute cognition, or only support it?

Clark and Chalmers (1998) supplied the philosophical parallel through which the framework is often read. They argue that when an external resource is reliably coupled to an agent, cognitive processes can extend beyond the skin. The notebook of their Otto case stands in for biological memory. Inga recalls from memory that the Museum of Modern Art is on 53rd Street. Otto has Alzheimer's disease. He carries a notebook everywhere, writes new information into it, and looks it up when he needs it. The notebook plays for Otto the role that memory plays for Inga, and that parallel is what carries the argument. Adams and Aizawa (2001) hold the opposing position. Their claim is not that people fail to use external resources. It is that no case has been made for the current existence of transcranial cognition, even though it may be a logical and a nomological possibility.

The disagreement is narrower than it sounds. Both sides accept that the notebook, the chart and the speed bugs are used and shape what gets done. The question is whether a reliably coupled resource is part of the cognitive process or a cause of the work that process does. Showing that outcomes depend on the resource does not settle it. Michaelian and Sutton (2013) add a second reading. The philosophical debate has often remained distant from relevant empirical research, and the field has moved from a first-wave, parity-based version of the thesis to a second-wave, complementarity-based one.

The two literatures are worth keeping apart. Hutchins takes the socio-technical system as the unit of analysis. The extended-mind claim takes an individual and asks how far that individual's cognition reaches. Adams and Aizawa attack the second, and their objection does not touch the first. What would settle it is a criterion fixed in advance, separating a resource that constitutes a cognitive process from one that merely enables it. Convergent evidence would also be needed, that a system-level analysis predicts something an individual-level analysis does not. Until then Adams and Aizawa hold the strongest single argument against reading the framework as a theory of mind, while the empirical half stands on its own.

The evidence on both sides of the philosophical dispute is conceptual. Adams and Aizawa (2001) argue from the structure of the case rather than from data, and Clark and Chalmers (1998) answer in the same currency. The system-level claim rests on ethnography of two settings, and the design-facing claim on four laboratory experiments with one puzzle task. While the dispute is open, a designer should design as though the representation were part of the process. The Tower of Hanoi result holds whether or not the extended-mind thesis does. No project should claim a cognitive extension it cannot demonstrate. The strongest objection is that the unit's design evidence belongs to the level it sets aside. Zhang and Norman (1994) tested individual participants on representational variants of a puzzle in a laboratory, which is not a socio-technical finding.

Where the argument is heading

Two lines of current work are named and active. The first carries the framework into settings where the artefact computes. Zhao and Han (2026) treat the model, the prompt and the human as components of one coupled cognitive system. They describe a five-phase cycle: intent envisioning, representation externalisation, generative reasoning, outcome assessment and cognitive update. Three gaps recur: of capability, of instruction and of intentionality. On their account, whether interaction enhances or atrophies cognition depends on whether the full cycle completes. That makes it a property of the system rather than of output quality.

The second line tests the framework where the setting has come apart. Ilomaki and Stevanovic (2024) report a conversation-analytic study of a video-mediated teleconsultation. The participants must direct one another's attention and action towards technological artefacts to solve an audio-connection problem, because video mediation limits physical access to a distant participant's ecology. They fit their interactional practices to the medium available, and offer their findings as specifications for existing theories of problem solving. Between those lines sits the question Hutchins (2010) leaves open. It is whether a framework that follows correlation across an ecosystem can still say in advance what will break, or whether it only describes settings after the fact.

The sources this unit rests on

• Hutchins, E. (1995). How a cockpit remembers its speeds. Cognitive Science 19(3). doi:10.1207/s15516709cog1903_1 Analyses speed computation in a commercial airliner cockpit and finds that the cognitive properties of the distributed system can differ radically from those of its members, with the remembering achieved across the crew, the booklet and the speed bugs.
• Hutchins, E. (1995). Cognition in the Wild. MIT Press. doi:10.7551/mitpress/1881.001.0001 The book-length case for distributed cognition, built on the navigation of a naval vessel. It argues that cultural activity systems have cognitive properties of their own, so the team is a cognitive system in its own right.
• Hollan, J., Hutchins, E. and Kirsh, D. (2000). Distributed cognition: toward a new foundation for human-computer interaction research. ACM Transactions on Computer-Human Interaction 7(2). doi:10.1145/353485.353487 Proposes distributed cognition as a foundation for HCI, arguing that the theory focuses on whole environments and is tailored to interactions among people and technologies. It shows the framework was offered for design, not analysis alone.
• Hutchins, E. (2010). Cognitive ecology. Topics in Cognitive Science 2(4). doi:10.1111/j.1756-8765.2010.01089.x Names cognitive ecology as the study of cognitive phenomena in context, pointing to the web of mutual dependence among the elements of a cognitive ecosystem and tracing its ancestry to Gibson, Bateson and cultural-historical activity theory. It records a shift to units defined by correlation across elements, a widening the unit treats as a cost.
• Rogers, Y. and Ellis, J. (1994). Distributed cognition: an alternative framework for analysing and explaining collaborative working. Journal of Information Technology 9(2). doi:10.1057/jit.1994.12 Argues that even seemingly discrete individual activities draw on individual, social and organisational factors at once, and that single-discipline frameworks cannot analyse collaborative work in situ; it proposes distributed cognition as the alternative, with applied studies of computer-mediated work, the framework's bridge into HCI research on collaboration.
• Zhang, J. and Norman, D. A. (1994). Representations in distributed cognitive tasks. Cognitive Science 18(1). doi:10.1207/s15516709cog1801_3 Sets out distributed representations, in which internal and external representations together represent a task's abstract structure, plus a method that decomposes a hierarchical task into levels, applied in four experiments on the Tower of Hanoi. It is the unit's strongest design-facing evidence.
• Scaife, M. and Rogers, Y. (1996). External cognition: how do graphical representations work? International Journal of Human-Computer Studies 45(2). doi:10.1006/ijhc.1996.0048 Proposes external cognition and characterises representations through computational offloading, re-representation and graphical constraining. It supplies the unit's mechanism for why changing the representation changes the task.
• Wright, P. C., Fields, R. E. and Harrison, M. D. (2000). Analysing human-computer interaction as distributed cognition: the resources model. Human-Computer Interaction 15(1). doi:10.1207/s15327051hci1501_01 Observes that distributed cognition treats resources for action as central while leaving them undefined enough to analyse an interaction, and answers with a resources model of a few information structures plus an interaction strategy. The framework becomes an analytic model, not a predictive one.
• Clark, A. and Chalmers, D. (1998). The extended mind. Analysis 58(1). doi:10.1093/analys/58.1.7 Argues that a reliably coupled external resource can be part of a cognitive process, so cognition extends beyond the skin, using the case of Otto's notebook. It is the philosophical parallel the unit keeps apart from Hutchins's system-level claim.
• Adams, F. and Aizawa, K. (2001). The bounds of cognition. Philosophical Psychology 14(1). doi:10.1080/09515080120033571 Defends contingent intracranialism: transcranial cognition may be a logical and a nomological possibility, but no case has been made for its current existence. It is the unit's strongest single objection, and it targets the extension of an individual's cognition.
• Michaelian, K. and Sutton, J. (2013). Distributed cognition and memory research: history and current directions. Review of Philosophy and Psychology 4(1). doi:10.1007/s13164-013-0131-x Reviews distributed-cognition research on memory and reports that the philosophical debate has often remained distant from relevant empirical research; it describes a move from first-wave, parity-based to second-wave, complementarity-based accounts. It is the unit's check on how much testable work has followed.
• Ilomaki, S. and Stevanovic, M. (2024). Distributed cognition in fractured ecologies: collaborative problem-solving in video-mediated interaction. Social Interaction 7(1). doi:10.7146/si.v7i1.132207 Reports a conversation-analytic study of a video-mediated teleconsultation in which participants solve an audio-connection problem by directing one another's attention and action towards technological artefacts, because video mediation limits access to a distant participant's ecology. It tests the framework in a fractured ecology.
• Zhao, H. and Han, Z. (2026). Understanding the mechanism of human-AI interaction: a distributed cognition perspective. Journal of Documentation 82(4). doi:10.1108/jd-01-2026-0003 Treats the model, the prompt and the human as one coupled cognitive system with a five-phase cycle and three recurrent gaps, of capability, instruction and intentionality. It finds that enhancement or atrophy depends on whether the cycle completes, so the framework now faces an artefact that computes.