RESEARCH REPORT
Identifying and individuating cognitive systems: a task-based
distributed cognition alternative to agent-based extended
cognition
Jim Davies
1
•
Kourken Michaelian
2
Received: 21 May 2015 / Accepted: 10 March 2016 / Published online: 31 March 2016
Marta Olivetti Belardinelli and Springer-Verlag Berlin Heidelberg 2016
Abstract This article argues for a task-based approach
to identifying and individuating cognitive systems. The
agent-based exte nded cognition approach faces a prob-
lem of cognitive bloat and has difficulty accommodating
both sub-individual cognitive systems (‘‘scaling down’’)
and some supra-individual cognitive systems (‘‘scaling
up’’). The standard distributed cognition approach can
accommodate a wider variety of supra-individual systems
but likewise has difficulties with sub-individual sys tems
and faces the problem of cognitive bloat. We develop a
task-based variant of distributed cognition designed to
scale up and down smoothly while providing a principled
means of avoiding cognitive bloa t. The adv antages of the
task-based approach are illustrated by means of two
parallel case studies: re-representation in the human
visual system and in a biomedical engineering
laboratory.
Keywords Distributed cognition Extended cogni tion
Cognitive systems Visual re-representation Philosophy
of mind
Introduction
This article argues for a task-based approach to identifying
(i.e., singling out) and individuating (tracing the boundaries
of) cognitive systems. The ‘‘The ontology of cognitive
systems’’ section revi ews difficulties faced by traditional
intracranialist or brainbound views as well as parity-based
and complementarity-based variants of extended cognition
with respect to the challenges of ‘‘scaling up’’ (identifying
and individuating supra-individual cognitive systems) and
‘‘scaling down’’ (identifying and individuating sub-indi-
vidual cognitive systems). A distributed approach can in
principle successfully meet the challenges of scaling up
and scaling down, but distributed cognition, like extended
cognition, also faces the challenge of avoiding cognitive
bloat, i.e., providing a criterion for the identification and
individuation of cognitive systems that does not imply an
unacceptable proliferation of ephemeral and explanatorily-
inert systems. In order to meet this challenge, ‘‘The task-
based approach’’ section develops a task-based (as opposed
to agent-based) variant of distributed cognition. According
to the task-based approach, cognitive systems are defined
relative to cognitive processes, which themselves defined
relative to cognitive tasks; the approach makes no essential
reference to agents. In order to show that the task-based
approach meets the challenges of scaling up and scalin g
down, ‘‘Applying the task-based approach’’ section applies
it to a pair of para llel case studies: re-representa tion in the
human visual system (scaling down) and re-representation
in a biomedical engineering laboratory (scaling up). The
Handling editor: John K. Tsotsos, York University, Toronto.
Reviewers: Iris van Rooij, Radboud University Nijmegen, Igor
Aleksander, Imperial College, London.
& Jim Davies
jim@jimdavies.org
Kourken Michaelian
kourken.michaelian@otago.ac.nz
1
Institute of Cognitive Science, Carleton University, Ottawa,
Canada
2
Philosophy Department, University of Otago, Dunedin,
New Zealand
123
Cogn Process (2016) 17:307–319
DOI 10.1007/s10339-016-0759-4
‘‘ Summing up’’ section provides a brief summary of our
argument.
The ontology of cognitive systems
Participants in the debate over extended and distributed
cognition are by now no doubt intimately familiar with the
details of Clark and Chalmer’s thought experiment about
Otto, an Alzheimer’s patient who relies on a notebook in
which he records relevant information in order to compen-
sate for his failing memory. Given its centrality to the debate,
it is nevertheless worth quoting the story once more:
Otto suffers from Alzheimer’s disease, and like many
Alzheimer’s patients, he relies on information in the
environment to help structure his life. Otto carries a
notebook around with him everywhere he goes.
When he learns new information, he writes it down.
When he needs some old information, he looks it up.
For Otto, his notebook plays the role usually played
by a biological memory. Today, Otto hears about the
exhibition at the Museum of Modern Art, and decides
to go see it. He consults the notebook, which says that
the museum is on 53rd Street, so he walks to 53rd
Street and goes into the museum.
Clearly, Otto walked to 53rd Street because he
wanted to go to the museum and he believed the
museum was on 53rd Street. And just as Inga [an
agent with a normal memory] had her belief even
before she consulted her memory, it seems reasonable
to say that Otto believed the museum was on 53rd
Street even before consulting his notebook. For in
relevant respects the cases are entirely analogous: the
notebook plays for Otto the same role that memory
plays for Inga. The information in the notebook
functions just like the information constituting an
ordinary non-occurrent belief; it just happens that this
information lies beyond the skin. (Clark and Chal-
mers 1998, pp. 12–13)
The moral of the story, if Clark and Chalmers are right, is
that, contra the traditional intracranialist view, cognition is
not ‘‘brainbound’’ [as Clark has subsequently put it (Clark
2008)]: when Otto looks up the museum’s address in the
notebook, the situation is best described not as an
individual cognitive agent retrieving information from an
external store, bu t rather as a single, extended cognitive
system retrieving information from an internal store, in a
manner that is functionally indistinguishable from that in
which a human agent with a normal memory, such as Inga,
retrieves information from her internal (biological) mem-
ory (Clowes 2013). In other words: cognition ‘‘ain’t (all) in
the head’’ (Clark and Chalmers 1998, p. 8): the process of
remembering loops out into the world, running from Otto’s
brain through the notebook and back again.
1
The fact that the memory process is viewed as starting
with the agent and ‘‘extending’’ to include the contribution of
the external resource is not inciden tal to the extended cog-
nition approach. Consider a series of three cases, this time
involving mental rotation rather than memory retrieval, that
Clark and Chalmers (1998) use to make the same point.
• In case 1, a person mentally rotates a two-dimensional
figure displayed on a computer monitor in the standard
(onboard) way.
• In case 2, the person has the computer rotate the
figure on the monitor itself.
• In case 3, he makes use of a neural implant to
(mentally?) rotate the figure.
The thought we are invited to have is that the process
unfolding in case 1 is uncontroversially cognitive; most of
us would be willing to count the process in case 3 as
cognitive; but if we are willing to count the latter process
as cognitive, then we should also be willing to count the
process in case 2 as cognitive, given that the only relevant
difference between cases 2 and 3 is the location of the non-
biological resource. The point to note here is that, just as in
the Otto example, in all three cases cognition is viewed as
starting from the human agent. While this approach to
identifying and individuating extended cognitive sys-
tems—starting with an agent who is himself uncontrover-
sially cognitive and then asking whether an external
resource counts as engaged in cognitive processing in
virtue of its relation to the agent in question—is natural
enough, we will argue that it is ultimately unsatisfactory.
Agent-based approaches: intracranialism
and (first- and second-wave) extended cognition
The agent-based approach fits comfortably with Clark and
Chalmers’ explicitly functionalist reasoning: it is in virtue
of the fact that an external resource (such as a notebook)
1
Clark and Chalmers argue that both cognition and mind are
extended; the Otto case was originally used to argue that mind, in
particular, is extended, but it can be used for either purpose. Our focus
here is on the hypothesis of extended cognition; hence we will not be
concerned with the claim that Otto’s beliefs are located in his
notebook but rather with the claim that Otto and his notebook
constitute a single cognitive system. See Huebner (2014) for a defense
of distributed mentality. Our focus is also distinct from that of
theorists who have argued for extended consciousness (Manzotti
2011; Honderich 2014), and we take no stand on the question of
extended or distributed consciousness. A fuller treatment would,
however, eventually have to deal with this question, especially in
view of recent debates on cognitive phenomenology, many partici-
pants in which have argued that cognition has a distinctive conscious
character (see, e.g., Strawson 2011; Smithies 2013.)
308 Cogn Process (2016) 17:307–319
123
plays the same functional role as the relevant internal
resource that it is legitimate to view the relevant cognitive
process as being realized in part by the external resource.
This functionalist reasoning is encapsulated in the parity
principle (as it has come to be known):
If, as we confront some task, a part of the world
functions as a process which, were it done in the
head, we would have no hesitation in recognizing as
part of the cognitive process, then that part of the
world is (so we claim) part of the cognitive process.
(Clark and Chalmers 1998, p. 8).
Given functionalism, and given that internal and external
resources are in fact sometimes functionally isomorphic,
we have a powerful reason to accept the extended cognition
hypothesis.
2
It shoul d be emphasized, however, that, in order to
derive extended cognition from functionalism, we do need
to assume that external resources do in fact sometimes
play functional roles that are typically carried out by
internal resources, and this assumption can be challenged.
Among the most influential arguments against extended
cognition and in favor of intracranialism are those pro-
vided by Adams and Aizawa (2001, 2008). One of their
key arguments (we will not consider their discussion of
the ‘‘mark of the cognitive’’) points to apparent functional
differences between internal and external resources (cf.
Rupert 2004, 2013). Consider, for example, Donald’s
(1993) discussion of exograms or external memory traces
(so-called by analogy with engrams, i.e., internal memory
traces). While Donald explicitly refers to exograms as
‘‘exact functional analogues’’ of engrams, he himself
points to many functionally relevant differences between
exograms and engrams; e.g., in contrast to engrams,
which have a labile and distributed character, exograms
are designed for stable storage of discrete items of
information (Sutton 2010; Michaelian 2012, 2014). The
specific disanalogies to which Adams and Aizawa point
do not show that internal and external resources are never
functionally isomorphic, but they do suggest that the
parity principle may be unable to support a form of
extended cognition which sees the existence of extended
cognitive systems as anything more than an exceptional
occurrence.
Intracranialists take functional disanalogies between
internal and exte rnal resources to constitute a reason for
rejecting extended cognition, but it is possible for an
extended approach to take such disanalogi es on board.
Indeed, it is (in part) in recognition of the importance of
such disanalogies that extended cognition theorists have
tended to shift away from parity-based to complementarity-
based argumen ts for the extended cognition hypothesis, in
what Sutton has referred to as a move from first-wave to
second-wave extended cognition (Sutton 2010) [though he
argues that the germ of the complementarity-based
approach can be distinguished as early as Clar ke (1998b)].
3
On the second-wave approach,
external states and processes need not mimic or
replicate the formats, dynamics, or functions of inner
states and processes. Rather, different components of
the overall...system can play quite different roles and
have different properties while coupling in collective
and complemen tary contributions to flexible thinking
and acting. (Sutton 2010, p. 194)
The second-wave approach brings extended cognition
much closer to distributed cognition (des cribed below),
but extended and distributed cognition continue to differ
in that, while the latter tends to focus on distributed
socio-technical systems which may involve multiple
agents and artifacts and which do not necessarily have a
clear center, the former continues to focus on systems
which are centered on a single agent (see Hutchins 2011).
As Clark recently put it, ‘‘[i]ndividual cognizing...is
organism-centered even if it is not organism-bound’’
(Clark 2007, p. 176) (cf. Giere 2011, 2012). Thus second-
wave extended cognition is continuous with first-wave
extended cognition in the sense that it advocates an agent-
based approach to the ontology of cognitive systems: we
identify a cognitive system by focusing on an uncontro-
versially cognitive agent and individuate the system by
asking which external resources interact with the agent in
the right way.
The challenge of cognitive bloat
A distinct line of objection to extended cognition con-
cerns ‘‘cognitive bloat’’ (Clark 2001; Rupert 2004; Row-
lands 2009; Palermos 2014). The worry here is that Clark
and Chalmers’ original criteria for cognitive extension,
the so-called trust-and-glue conditions—according to
which the agent (1) must have reliable access to the
information contain ed in the resource, (2) must have
reliable access to the resource itself, (3) must tend to
endorse information contained in the resource upon
retrieval, and (possibly) (4) must have previously
endorsed the information (Clark and Chalmers 1998)—are
too easily satisfied, with the consequence that extended
cognition entails an ‘‘unacceptable proliferation’’ of
2
But see Sprevak (2009) for an attempt to use the relationship
between functionalism and the extended cognition hypothesis to
ground a reductio of the latter.
3
For an overview of the transition from first-wave to second-wave
extended cognition theorizing, see Kirchhoff (2012).
Cogn Process (2016) 17:307–319 309
123
‘‘extremely short-lived’’ cognitive systems (Rupert
2004, p. 396).
4
The problem of cognitive bloat, however,
may not be as severe as many participants in the debate
(on both sides) have assumed. Second-wave extended
cognition theorists, in fact, have tended to emphasize the
fact that extended systems may in fact often be extremely
short-lived—in other words, that we may just need to
learn to live with cognitive bloat.
But learning to live with cognitive bloat will not
make the problem go away entirely. It may not be a
problem to posit extremely short-lived cognitive sys-
tems, but it is a problem to posit too many systems. In
order for an approach to the ontology of cognitive sys-
tems to have any explanatory tracti on, i t cannot imply
that an extended cognitive syst em c omes into existence
any time the agent makes use of an external resource in
order to accomplish a given cognitive task; we require a
way of distinguishing between cases in which an agent
merely makes use of an external resource and cases in
which the agent and the resource constitute a joi nt
cognitive system [cf. Adams and Aizawa (2008)onthe
coupling-constitution fallacy]. Second-wave extended
cognition theorists thus still owe us a means of identi-
fying and individuating cognitive systems. It is not
obvious how the second-wave approach might provide
such a means, however, especially as it tends to
emphasize that coupling between agents and external
resources is a complex, multidimensional affair (Heers-
mink 2012, 2015), with the result that it becomes diffi-
cult to see how the extended c ognition theorist might
‘‘drawtheline’’insuchawayastopreventthetheory
from implying that an extended cognitive system pops
into existence every time an agent relies to some extent
on an externa l resource (and pops out of existence aga in
as soon as he ceases to rely on it). Distributed cognition,
which is likewise prepared to acknowledge short-lived
distributed systems composed of temporarily interacting
elements, is, as we will see, in the same boat when it
comes to cognitive bloat.
Two additional challenges
In addition to the problem of cognitive bloat, approaches to
the identification and individuation of cognitive systems
face the interrelated challenges of scaling up and scaling
down.
Scaling up
‘‘Scaling up’’ refers to ability of an approach to acknowl-
edge that an assemblage of distinct entities, including
individual agents, might, under the right conditions, itself
count as a cognitive system. For straightforward reasons,
intracranialism does not acknowledge that such a set of
entities might qualify as a cognitive system: if cognition is
brainbound by definition, it crosses neither the boundary
between agent and artifact nor the boundary between agent
and agent. Extended cogni tion was designed in part to
overcome this limitation of intracranialism, but standard
forms of extended cognition have difficulty scaling up in
certain cases.
Extended cognition can of course acknowledge both
agent-artifact and agent-agent systems. But due to its
agent-based character, it flounders in cases where the sys-
tem is not centered on a single human (or other biological)
agent and in cases where no human is involved. While the
possibility of purely artefactual cognitive systems may be
consistent with the letter of extended cognition, extended
cognition theorists have in practice focused largely on
systems with humans in the loop. And while extended
cognition may have no difficulty identifying a cognitive
system in cases where the system is not centered o n a
single human agent, it will have difficulty individuating the
relevant system. For the extended cognition theorist, cog-
nitive systems are identified by starting from a human
agent; they are individuated by determining which external
resources (possibly including other human agents) are
hooked up to the relevant agent in the right way: functional
isomorphism (pl us satisfaction of the trust-and-glue crite-
ria), in the case of first-wave extended cognition; suffi-
ciently tight coupling, in the case of second-wave extended
cognition (see ‘‘The task-based approach’’ section below).
In the discussion to follow, we cannot hope to cover all
forms of scaling up. Starti ng with Hutchins (1995a), dis-
tributed cognition researchers have provided detailed
investigations of case s involving complex networks of
agents and artifacts (see Sutton 2006; Dror and Harnad
2008; Michaelian and Sutton 2013). Here, we will focus on
a relatively simple case involving a single agent and
multiple artifacts but in which the relevant artifacts appear
to be doing the lion’s share of the work involved in per-
forming the cognitive task (i.e., a case in which the putative
system is not centered on the human agent), as such a case
provides an appropriate illustration of the advantages of the
task-based approach.
Scaling down
Intracranialism, extended cognition, and many standard
versions of distributed cognition fail to meet the challenge
4
In fact, it is unclear to what extent it is compatible with the spirit of
first-wave extended cognition to impose criteria in addition to
functional isomorphism; in retrospect, the ‘‘trust-and-glue’’ conditions
to some extent look like an ad hoc attempt to prevent cognitive bloat
(Palermos 2011).
310 Cogn Process (2016) 17:307–319
123
that we refer to as ‘‘scaling down’’: moving below the level
of the agent to acknowledge that, in a given case, it may
not be the agent as a whole that constitutes the relevant
cognitive system but rather some subsystem within the
agent. Cognitive neuroscientists and cognitive scientists
more generally are often interested in distinguishing which
parts of an individual’s brain are implicated in a given
cognitive process. Many tasks are associated with specific
brain regions. These associated regions are inferred
through a variety of methods, including lesion studies and
neuroimaging. In this context, it often makes sense to view
a certain brain area or structure as itself performing a
cognitive task, with respect to which other brain areas
count as its environment. The limbic system, for example,
appears to be the seat of emotion, the frontal lobes the seat
of planning, and the hippocampus is centrally involved in
the consolidation process responsible for turning labile
short-term memories into stable long-term memories.
5
These generalizations about brain areas are, of course,
drastic oversimplifications, but it does appear that certain
brain areas are much more implicated in particular cogni-
tive tasks than others.
Intracranialism and extended cognition, again, are
agent-based and thus lack a principled means of singling
out sub-individual cognitive systems. Both tend to view the
individual cognizer as a unified whole, a sort of black box,
whose cognition may or may not extend to objects nor-
mally considered to be part of his environment. But the
goal of cognitive neuroscience is precisely to open up the
black box and figure out its workings—what its compo-
nents are and how they interact in cognitive processing.
The intracranialist view can do no better than to say that
planning, for example, is something done by a cognitive
agent, operating on his own. The extended cognition view
can do no better than to say that planning is something
done by a cognitive agent and which sometimes does and
sometimes does not extend into external resources. Hence
both approaches fail to scale down. Finally, distributed
cognition, with its non-agent-based character, can in prin-
ciple acknowledge sub-individual systems. As Hutchins
recently put it, ‘‘the interesting question...is not ‘is cogni-
tion distributed or is it not?’ or even ‘is cognition some-
times distributed and sometimes not distributed?’ Rather,
the interesting questions concern the elements of the cog-
nitive system, the relations among the elements, and how
cognitive processes arise from interactions among those
elements’’ (Hutchins 2014). As Theiner (forthcoming)
points out, this means that the perspective of distributed
cognition can be adopted with resp ect to systems at mul-
tiple scales; at a microscale, the perspective of distributed
cognition might shade into that of embodied cogni tion,
which views cognitive processes as distributed across not
only the brain but also the body (Shapiro 2011). In practice,
however, the focus of distributed cognition research tends
to be firmly on supra-individual cognitive systems. Exist-
ing distributed approaches, moreover, fail to provide a
principled means of defining sub-individual systems. The
task-based approach developed here is distinctive in
focusing equally on sub-individual and supra-individual
systems and in attempting to provide a principled means of
defining both kinds of system.
Distributed cognition as a non-agent-based
approach
As noted above, moving to second-wave extended cogni-
tion already brings us closer to distributed cognition, and it
is ultimately a form of distributed cognition that we want to
defend here.
Distributed cognition is like extended cognition in that it
argues that certain cognitive tasks [involved, e.g., in
mathematics (Lave 1988) or reading (Donald 1993)] are
accomplished in part by structures external to the agent.
Cognition, viewed as the set of processes that accompl ish
cognitive tasks, is often distributed across richly interacting
systems of minds and non-biological resources. This much,
extended cognition can agree with. Distributed cognition is
unlike extended cognition in that it goes further in the same
direction, rejecting the agent-based approach to identifying
and individuating cognitive systems. Hutchins, for exam-
ple, famously analyzes ship navigation as being accom-
plished by a distributed system involving multiple agents
and instruments, no one of which is responsible for the
overall direction of the system (Hutchins 1995a). Similarly,
he argues that the distributed system constituted by an
aircraft’s cockpit—including human agents, computers,
display screens, and sensors—itself is responsible for
piloting a plane (Hutchins 1995b).
In virtue of its non-agent-based approach, distributed
cognition is better positioned than extended cognition to
meet the challenge of scaling up. As noted above, however,
standard versions of distributed cognition do have difficulty
scaling down. Moreover, the threat of cognitive bloat
affects distributed cognition just as much as extended
cognition: the distributed cognition theorist must provide a
principled approach to the bounds of cognition, a means of
identifying and individuating (possibly short-lived) cogni-
tive systems which does not imply an unacceptable prolif-
eration of such systems. The following section of the paper
is devoted to articulating a version of distributed cognition
5
Similarly, artificial intelligence researchers will sometimes perform
‘‘ablation experiments’’ on their complicated programs. These are
analogous to lesioning experiments in animals in that they remove a
part of the program and observe the resulting behavior, allowing the
researcher to draw conclusions regarding the functions of specific
parts of the program.
Cogn Process (2016) 17:307–319 311
123
