The theory of distributed cognition (Hutchins 1995a) has been used to analyze and evaluate the flow of representations in real-world cooperative work settings. In this critique, we outline the basic motivation and background of distributed cognition. We will explain the theory and examine one case study (Ackermann and Halverson 1998, 1999, 2000) that attempts to apply the theory in a real-world work setting. Finally, we ask whether the theory of distributed cognition is really useful for answering design questions.
Early work in distributed cognition was motivated by the basic insight that cognition is a socially (also, materially and temporally) distributed phenomenon, one that is essentially situated in real practices (Hutchins 1995a). The theory does not posit some new kind of cognitive process. Rather, it represents the claim that cognitive processes generally are best understood as situated in and distributed across concrete socio-technical contexts.
The traditional sort of theory in cognitive science emphasizes an internalism that marginalizes (some would argue ignores) the role of external representation and problem solving in cooperative contexts. Traditional approaches to description and design in human-computer interaction have similarly focused on users internal models of the technologies with which they interact.
Alternative theoretical perspectives have since offered criticisms of the tradition, emphasizing instead the role of concrete social and technological contexts. An ecological approach emphasizes the role of agent-environment interaction. The role of socio-cultural context in mediating intentional action is emphasized by an approach called activity theory. And situated action theorists stress how attention to the concrete details of intelligent action reveals the shortcomings of abstract theoretical construals of human-machine interaction.
The theory of distributed cognition offers this same sort of critique of the tradition but does so in a way that does not reject traditional computationalist wisdom. In modern complex cooperative work environments, we find humans and technologies together maintaining and manipulating representational states, carrying out processes that solve problems. The theory of distributed cognition is motivated by the idea that such systems instantiate genuinely cognitive processes and that the cognitive properties of these types of socially, materially and temporally distributed systems differ from those of the individuals that act in them.
Most recently, the theory od distributed cognition has been proposed as a "new foundation" for human-computer interaction (Hollan, et. al. 2000).
The theory of distributed cognition was originally developed by Edwin Hutchins to provide a more balanced theoretical treatment of problem solving in real work situations, and to supply a new framework for cognitive science generally. The seminal study (Hutchins 1995a) comprises an investigation of a ship's navigation. This is a task that, in modern western society, is carried out by teams of sailors working with various types of artifact. Hutchins describes how individuals use tools to generate and maintain representational states, which are then propagated through the system to finally produce a "fix" of the ship's location.
In another study (1995b) Hutchins investigates an airline cockpit as a distributed cognitive system. An airline cockpit "remembers its speeds" through complex interactions between the pilots and various representational devices, tools and media. Hutchins shows that the cognitive properties of the distributed system depend on the physical properties of the representational media in which it is implemented. Further, "a complete theory of individual human memory would not be sufficient to understand that which we wish to understand because so much of the memory function takes place outside the individual".
The basic insight of the theory of distributed cognition is that cognitive phenomena generally are best understood as distributed processes. These processes are cognitive in the traditional sense. That is, they are computational processes. The theory does not do away with the notion of individual cognition, though early motivation and ongoing emphasis has been on studying instances of socio-culturally distributed cognition. Here, theoretical focus is on how cognition is distributed across people and artifacts, and on how it depends on both internal and external representations.
The traditional idea that cognition is computation is preserved in the theory of distributed cognition. However, computation is conceived broadly as "the propagation of representational state across representational media" (Hutchins 1995a). In a socially distributed system, people interact with artifacts to create and coordinate representations. Thus, representational state can be "propagated" across distinct media and across different representational systems.
The theory of distributed cognition employs a variable unit of analysis. Traditional cognitive theory takes the individual person as the proper unit of analysis. On this traditional view, cognitive processes are internal processes. The social and cultural context is thus often left out of the analysis. In contrast, the theory of distributed cognition makes of the individual and its socio-cultural context a larger cognitive system, one that is to be analyzed in a broadly traditional way, as a computational system.
Enlarging the unit of analysis in this way has the benefit that representations internal to the system are now "external" representations with respect to the individual agents that use and make use of them. Thus, cognitive processes are, from this theoretical standpoint, fully observable. Certainly, the internal representations of individual agents remain important. But, as Hutchins is keen to point out, many of these are "still all cultural in the that they are the residua of a process enacted by a community of practice rather than idiosyncratic inventions of their individual users" (1995a, p. 130).
Using the notion of a variable unit of analysis, we can decompose larger cognitive systems into subsystems that can be similarly decomposed. The unit of analysis is defined computationally (i.e. functionally). Nothing in the theory of distributed cognition prevents us from taking the individual as the unit of analysis. We must be prepared, however, to analyze individual cognition as a process that is distributed among the important functional components of the brain.
There are several methods for collecting empirical data that may be employed in a distributed cognition analysis. These include close study of video or audio recordings of work situations and computer (neural network) simulations. For research in the area of socially distributed systems, an especially important method is a new kind of cognitive enthographic study (Hollan et. al. 2000, pp.179-180). "The ethnography of distributed cognitive systems retains an interest in individual minds, but adds to that a focus on the material and social means of the construction of action and meaning. It situates meaning in negotiated social practices, and attends to the meanings of silence and the absence of action in context as well as to words and actions." There is a real sense, then, in which collecting data for a distributed cognition analysis presupposes becoming an expert of sorts in the domain of inquiry. There really is no other way to ensure that such subtleties of social and cultural context can be captured by the analysis.
Ackerman and Halverson (1998, 1999, 2000) apply the theory of distributed cognition to help identify memory systems within an organization. In their research, they make a case study of a telephone help line that is set up to respond to human-resource inquiries. The theory of distributed cognition is useful for such an investigation because its variable unit of analysis allows researchers to accommodate in a single theory different types and scales of organizational memory (2000, p. 59).
The telephone help line group (HLG) answers employee questions concerning things like benefits and matters of personnel policy. From the standpoint of the theory of distributed cognition, the HLG computes solutions to incoming queries. An incoming call is received as input. The caller's question is somehow represented and this representation is then propagated through the system, eventually being transformed into a solution. The computational work done in this system can be repetitive. As a result, an individual operator's own memory is sometimes sufficient to handle an inquiry. There is redundancy, too, in the kinds of representation and processing that generally occurs when handling any call. Help line operators make regular use of various types of artifacts: monitors, software systems, a telephone system, and public and private paper resources (e.g. manuals and Post-It notes). As the author's point out (1998, p. 41) the unit of analysis may vary from call to call, depending on the resources employed in computing a particular solution. An individual operator, for example, may answer most inquiries using only certain resources and seek information or advice from other agents in the group only in special circumstances.
One of the simplest computations carried out by the HLG has to do with verifying an individual's employment status. This "employment verification" task requires looking up information in a database in order to verify a given individual as an employee. As simple as this task may seem, it turns out upon analysis to involve distinct memories and distinct processes that coordinate to propagate representational state from input to output, from inquiry to answer.
Employee records are stored in a database called EMPLOY. And this database is accessed via a central terminal, which is located away from an operator's desk. In a typical employee verification request, the telephone system automatically routes an incoming call to an available operator. As the operator listens to the caller's verbal request she audibly repeats some of the relevant information herself while generating an electronic record of the call. This is typed directly into a software system called CAT. At the same time, the operator requests needed information from the caller, typing this information as well into the CAT system. The operator next uses pen and paper to write information just typed into the CAT system onto a piece of scrap paper and then physically moves to the central terminal where she can access the EMPLOY database. She types the information written on the scrap paper into the terminal and retrieves an employee record. Some of this information is then written down onto the scrap paper and the operator returns to her desk to complete the call.
According to a distributed cognition analysis, the HLG system computes an employment verification, transforming input (incoming query) into output (employment status verification) by propagating concrete representational state across various media. The analysis in the present case study reveals several distinct artifacts and processes (rather than a single, centralized one) serving as memory in the computation. An incoming query is accessed directly from the telephone system that routes calls to the next available operator. Verbal rehearsal and other forms of individual and perhaps internal short-term memory are coordinated until a representation in the CAT system is created. The operator's own short-term memory is again employed in creating and coordinating representational state between the CAT system and a piece of scrap paper, between the paper and the EMPLOY database, and back again to the caller and the telephone system. The analysis reveals a computational path "consisting largely of transferring information from memory to memory until the right pieces of information can be coordinated to answer the caller's question" (2000, pp. 61-2). In this process, then, representational state is propagated across both internal and external media, from one memory to the next.
Further, the process of employee verification itself can be recognized as simultaneously embedded within different larger systems. Adopting a higher-level unit of analysis the CAT system, for example, acts as a "group memory" of all calls that have been processed. Records can be recalled to supply useful information on how an atypical request might best be handled. It is interesting to note, for example, that in the present case study the operator does not bother to provide many details of the employee verification call when creating the CAT record. This type of call is typical and a CAT record is therefore not likely to be retrieved for those details, so there is little practical use in recording them.
However, some CAT records are recalled precisely in order to help reconstruct the solution to an unfamiliar problem. The author's of the study use the concepts of decontexualization and recontextualization to help describe this type of memory implementation and access. CAT records are what the author's refer to as "boundary objects". These are representational artifacts that have the dual responsibility of serving the needs of those who create and those who use them. It therefore matters how incoming representational states are decontextualized when they are coordinated with new representations created in the CAT system. For these coordinate representations to be useful, they must somehow anticipate the recontextualization that will be made of them in future processes. These concepts are helpful for understanding how memory artifacts influence memory processes. They are also helpful in providing a clearer picture of the material and temporal distribution of organizational memory in the present case study.
A distributed cognition analysis of memory in cooperative organizations reveals several aspects of organizational memory that seem useful from the standpoint of HCI design. One helpful thing about a distributed cognition analysis is the way it reveals constraints that are implied by the embodied nature of the representational media that are inevitably employed in carrying out a given task. In a larger organization with a busier call center, for example, a central terminal where help line operators must access employee verification records (a typical sort of request) could clearly become a kind of bottleneck for productivity.
One of the clearest advantages of a distributed cognition analysis with respect to design issues in organizations and cooperative work environments generally is surely that the theory can accommodate the rich variety of representational media and systems that in fact implement a group or organization's cognitive processes. Because the theory construes cognition traditionally as a computational process, and because the theory is not committed to any fixed unit of analysis, the same theoretical perspective can be used to decompose an entire organization into the smaller, functional groups that make it up.
Some (Nardi 2002; Rogers 1997, 2004) have argued against the general utility of a distributed cognition analysis as a guide for design. The most important complaint seems to be with enthographic data collection, a methodology to which this theoretical perspective is admittedly committed (Hutchins 1994, Hollan, et. al. 2000).
For one thing, there may still be some question whether there is any coherent theory that can mediate (or synthesize) the traditionally opposed conceptual frameworks at play in the cognitive, computationalist tradition and the situated/embodied-embedded critique. Perhaps no theory can fully achieve this goal. In any case, it isn't so clear exactly what makes the theory of distributed cognition our best candidate theory in this regard.
But practical concerns are more troubling. The commitment to enthography means that there is a substantial investment required to actually apply the theory to any specific design issue. "It is not a methodology that one can readily pick off the shelf and apply to a design problem." (Rogers 1997). And while some will argue that distributed cognition's "data driven" approach is a bonus for design (Halverson 2002), others argue that it is precisely this aspect of the theory that makes it so unwieldy (Nardi 2002). Nardi, for example, argues that apart from some particular cases, distributed cognition's low-level analyses "will not enhance engineering practice for building CSCW applications".
What may be right is that designers do need some kind of common theoretical framework in order to make sensible design assessments and decisions. It might be that distributed cognition analyses do not provide this type of framework. But they are not incapable of generating useful abstractions, such as the notions of decontextualization and recontextualization of boundary objects. Perhaps, in time, distributed cognition analyses will generate concepts and ideas about human-computer interaction that will better approximate a useful theory for design. The benefit, of course, would be that this sort of theoretical framework would have an empirical basis.
Few research groups and no journals are currently dedicated to the topic of distributed cognition. But work on the theory and its application appears from time to time in interactive HCI labs worldwide and in journals concerned with cooperative, collaborative work and learning environments. The theory's influence in HCI and Interaction Design is (partially!) represented by the links listed here.
The UCSD Distributed Cognition & HCI Laboratory
The laboratory at UCSD, supervised by James Hollan and Ed Hutchins, combines enthnography and experiment to improve the design of digital artifacts for real-world work environments.
The Interactive Cognition Lab at UCSD
Another laboratory at UCSD, run by David Kirsch and Aaron Cicourel, studies the fundamentals of designing interactive environments. Their research includes topics in distributed and collaborative cognition.
The Interact Lab at Sussex
The Interact Lab is now directed by Geraldine Fitzpatrick and co-directed by founder Yvonne Rogers. The focus of research here is on developing novel user experiences by exploiting the possibilities for interaction between people, technologies, and representations.
COTCOS: Cooperative Technologies for Complex Work Settings
COTCOS is an interdisciplinary research program that lasted from 1996 to 2000, dedicated to investigating cooperative technologies in complex work settings.
TECFA Education & Technologies
TECFA at the University of Geneva pursues work on educational technology. One of their research projects focuses on distributed learning systems.
The DME at Bath
This research group investigates the causes of failure in systems of distributed cognition.
The ACM Special Interest Group SIGGROUP
Computer Supported Cooperative Work (CSCW), The Journal of Collaborative Computing
The Journal of Interactive Learning Research (JILR)
Some other, more specific online references to distributed cognition and its application to related areas in HCI and beyond.
A nice summary of information on distributed cognition.
HCII 2003, T25 (Gerhard Fischer)
HCI Themes for the Future
Fischer, Gerhard (2003)
Distributed Cognition: A Conceptual Framework for Design-for-All
Waloszeck, Gerd (2003)
Dissolving Boundaries with Distributed Cognition and xApps
Distributed Cognition (a PPT Presentation by Steven Harris)
This is a PowerPoint presentation of my own on the role of the theory of distributed cognition in HCI design
There is reason to be optimistic about the role distributed cognition has to play in the future of HCI design. I believe the theory is entirely right-headed in what it attempts to do. Traditional approaches in the cognitive sciences have too often sought to understand how people do what they do without a proper appreciation for the tools they actually use in carrying out these tasks. I believe that the emphasis distributed cognition theory places on concrete agent-artifact interaction and its context is the right sort of emphasis for informing understanding and design.
Ackerman, M.S., and C.A. Halverson (1998) 'Considering an organization's memory'. Proceedings of the 1998 ACM Conference on Computer supported Cooperative Work, pp.39-48.
Ackerman, M.S., and C.A. Halverson (1999) 'Organizational Memory: Processes, Boundary Objects, and Trajectories'. IEEE January 05 - 08, 1999 Maui, Hawaii
Ackerman, M.S., and C.A. Halverson (2000) 'Re-examining organizational memory'. Communications of the ACM, 43(1), 58-64.
Halverson, C. A., (2002) 'Activity theory and distributed cognition: Or what does CSCW need to DO with theories?' Computer Supported Cooperative Work, 11:243-267.
Hollan, J., Hutchins, E., Kirsh, D. (2000) ‘Distributed cognition: toward a new foundation for human-computer interaction research’. ACM transactions on computer-human interaction, 7(2), 174-196.
Hutchins, E. (1995a) Cognition in the wild. MIT Press: Cambridge, MA.
Hutchins, E. (1995b) 'How a cockpit remembers its speeds'. Cognitive Science 19, 265-288.
Hutchins, E., and Klausen, T. (2000) 'Distributed cognition in an airline cockpit. Cognition and communication at work'. In Cognition and communication at work, Y. Engström and D. Middleton, Eds. Cambridge University Press, New York, NY, 15-34.
Nardi, B. A., (2002) 'Coda and response to Christine Halverson'. Computer Supported Cooperative Work, 11:269-275.
Rogers, Y. (1997) A brief introduction to Distributed Cognition.
Rogers, Y. (2004) New theoretical approaches for HCI. To appear in ARIST: Annual Review of Information Science and Technology, No. 38.
Rogers, Y. and Ellis, J. (1994) 'Distributed Cognition: an alternative framework for analyzing and explaining collaborative working'. Journal of Information Technology, 9 (2), 119-128.