Experimental and behavioural economics and, by association, neuroeconomics overwhelmingly involve this kind of experiment: Simple choices for individuals to make, perhaps in the context of a simple social interaction such as a two-person game. Importantly, each subject's access to other subjects is tightly circumscribed. In individual choice experiments, such access is simply forbidden for statistical reasons, and in game experiments it is closely regulated for both game-theoretic and statistical reasons. The hunch seems to be (1) that individual minds are the centrally important information-processing units of economic science, and (2) that simple decision problems and simple games usefully characterize the actual decision problems and actual games that govern the important economic phenomena, such as growth and inequality. For example, McCabe (Reference McCabe2008, this issue) singles out trust and reciprocity as enablers of economic growth, and perhaps they are. We are invited to view experiments about simple trust games between two individuals (and what happens within the individual skulls involved in such games) as crucially relevant to economic growth.
Let me illustrate an alternative hunch. I rather suspect that economic growth is overwhelmingly due to the cultural accumulation of external cognitive artefacts, including social distributions of information processing, which implement external information representations and algorithms. This accumulation increasingly exploits the performance strengths of individual cognition (e.g. hand-eye coordination, fast pattern recognition) and increasingly avoids its weaknesses (e.g. algebra, emotional interference). It increasingly distributes information processing across group members in highly structured ways, aided by external cognitive artefacts, and lessens the importance of individual brains to economically important information-processing tasks. I also suspect this accumulation is not mostly the product of individual learning, but is rather mostly the product of social learning processes that operate across contemporaneous social space and (especially) across generations of social groups confronting related tasks. This story of economic growth is of the increasing marginality of individual brains and the growing centrality of socially distributed cognition. I think this is a plausible alternative hunch about cognitive science's primary relevance to economic growth.
I will call the viewpoint I am criticizing “cognitive individualism”. It powerfully shapes and constrains theory, research questions and empirical inquiry. Importantly for this volume, it unsurprisingly draws researchers and resources ever deeper into the study of individual cognition. In an intellectual milieu guided by cognitive individualism, the emergence of neuroeconomics (and the strong support it gets from many distinguished scholars) was inevitable. To be clear, neuroeconomics is not my target here; it is just one particularly vivid manifestation of my true target.
I suspect that many decision and discovery processes that govern the most important economic phenomena involve a high degree of social and/or computational complexity. This is notoriously true of many dynamic decision problems and their solution algorithms: I suspect that a substantial part of inequality is due to differential success (across dynastic families) at intergenerational innovation and transmission of good heuristics for dealing with life-cycle decisions. Related considerations apply to technological progress and hence growth. I simply don't believe I have this nifty laptop or this swell alphabet merely (or even mostly) because of individual cognitive solutions to simple tasks or games. Rich, external cognitive artefacts like these were accumulated by increment and innovation across years and indeed millennia. I suspect the information-processing units responsible for these increments and innovations were usually groups of people and rarely individual brains (notwithstanding our seeming need to render scientific and technological history in terms of individual heroes). That is, I suspect our accumulation of cognitive artefacts (that transform our tasks and shape distributed cognition) is overwhelmingly the product of social learning processes. And so I suspect that most of the experimental and theoretical work on individual learning, with or without brain scans, has limited relevance to this accumulated cognition, and hence limited relevance to economic growth. Or at least, so goes an alternative hunch.
To put some formal structure on this alternative hunch, I begin by (tentatively) viewing the relationship between neoclassical economics and cognitive science from the perspective of Marr (Reference Marr1982). Marr divides understanding of an information processor into three levels. At the top level, we have a rational model of a processor – a description of what a processor ought to compute, and an argument as to why it should compute that. Marr calls this the “computational theory” or CT of the processor. Addition is a CT of a cash register; a Nash equilibrium strategy is a CT of a game player (given certain well-known assumptions about other players); and so on. The next level of understanding is “algorithm and representation” or A&R. Here we find representations of inputs and outputs, and the algorithms that convert the former into the latter, such that the output approximates the CT's specified output in some satisfactory way – robustly, efficiently, accurately or with some tradeoff amongst these. Usually, a great variety of A&Rs can approximate the specified output of any CT. We might (with some abuse) call the A&R “software”. Finally there is the bottom level of “physical implementation” or PI: What's the actual physical “machine” or “hardware” in which the A&R will “run?” Vacuum tubes, silicon, neurons or Tinker Toys? – Again, many PIs can “run” any given A&R.
From a Marrian perspective, neoclassical economics is CT without any specification of A&R or PI: It is the hunch that the outputs specified by decision and game theory – the neoclassical CTs – are, to an empirically fruitful approximation, the observed behaviour of economic processors under almost all conditions.Footnote 1 Marr thought that CT was centrally important for understanding processors, and some philosophers such as Dennett (Reference Dennett1987) seem sympathetic to that view (with caveats). This is why the title of Gul and Pesendorfer's (Reference Gul, Pesendorfer, Caplin and Schotter2008) spirited defence of neoclassical economics, “The Case for Mindless Economics”, makes me wince. That title just invites distracting name-calling like “behaviourist” and “Skinnerian” for no good reason. With a title like “The Case for Algorithm-, Representation- and Machine-Free Economics”, Gul and Pesendorfer would pretend to know and care about cognitive science in a respectable Marrian sense.
Marr's three-level view puts neuroeconomics in danger of appearing a sideshow about PI (hardware) which, in Marr's view, is frequently a distraction from the really important issues that largely revolve around an A&R's performance relative to CT's specifications. Philosophically, the notion of “multiple realizability” of the mental in the physical also suggests that hardware is a distraction: The important essence of the mental may not be usefully reduced to any particular machine (Fodor Reference Fodor1975; Putnam Reference Putnam1988). McCabe's quick and early attempt to dismiss the software/hardware question attests to the danger. Neuroeconomics answers: We accept Marr's distinctions and views in general, but we suspect that the individual human nervous system PI is in fact an especially strong constraint on the kinds of A&R that will “run” in it. Some neuroscientists go even further, arguing to dispense with CT and dissolve the distinction between A&R and PI in their special case: They say that while it is a cogent distinction in general, it is practically useless in the case of biological nervous systems. This line of thinking has its own philosophical heroes (e.g. Churchland Reference Churchland1981), and usually goes along with some stories about evolution (McCabe hints at these without development).
Not all neuroeconomists want to dispense with CT: For instance, Glimcher (Reference Glimcher2003) seems quite sympathetic with Marr's view that rigorous cognitive science cannot proceed without a relevant CT. Philosophers such as Fodor (Reference Fodor2000) have attacked both the notion that biological hardware is crucially special, and the relevance of adaptationist stories concerning that hardware's structure. I find evolutionary talk inherently seductive, but I try to keep Gould and Lewontin (Reference Gould and Lewontin1979) in mind when considering ex post just-so stories about “adaptations”. Yet in spite of these difficulties, I'll grant neuroeconomists the specialness of the biological hardware. It matters little to my subsequent argument.
Since introducing Marr above, I've said nothing explicit about which processor is the centrally important processor of economic science – what we call an agent at the level of economic CTs. Fudenberg and Levine's (Reference Fudenberg and Levine2006) “dual self” model of impatience and small-stakes risk aversion is just a recent example of theoretical fission of individuals into multiple noncooperative agents, a move endorsed by some philosophers (Ross Reference Ross2005). Fissions of the agency of firms (into managers versus owners) and households (into spouse versus spouse) similarly create new noncooperative games between new smaller agents – an older and well-established move, as Camerer (Reference Camerer, Caplin and Schotter2008) points out. What all of these moves have in common is fission of agency. Might it also be useful to explore some fusions of agency?
Edwin Hutchins is a cognitive anthropologist known to some as the “father of modern cognitive ethnography”, and a contemporary developer of theories of socially distributed cognition. Hutchins enormously influenced philosophers of mind such as Clark (Reference Clark1997). In his book Cognition in the Wild, Hutchins (Reference Hutchins1995) argues that the founding metaphor of cognitive psychology – that individual cognition is like a computer – was always flawed. Hutchins instead suggests that it is socio-cultural symbol manipulation systems (cognition distributed across groups of humans, in interaction with external cognitive artifacts and both natural and artificial environments) that are really like a computer.
Hutchins (Reference Hutchins1995: 50) begins gently and craftily, pointing out that “Marr intended his framework to be applied to the cognitive processes that take place inside an individual, but there is no reason, in principle, to confine it to such a narrow conception of cognition.” To elaborate “Hutchins’ hunch”, individual brains are not the machines that matter the most in social science. The important machines of social science are social groups with special artefacts, communicating with one another both now and across time and even across generational time. In these larger machines, individual brains are not the machines of interest: Instead, individuals (brain and body) are just one type of cog in larger socially distributed cognitive machines. We might think of the usual path of cognitive-cultural history (and economic growth) as the successive accumulation and modification of external cognitive artefacts: external representations of information, external physical devices and external patterns of information flow and aggregation. These are to some extent social A&R, but also additional types of cogs in larger socially distributed machines. Cognitive-cultural change may tend to (1) increasingly exploit strengths of human cogs and avoid their weaknesses, (2) make human cogs interchangeable and redundant where possible, and (3) increase the robustness of the bigger social machines to their errors, weaknesses, and arrivals and departures.
Now let us return to the neuroeconomists’ attempted dismissal of the Marrian suspicion that hardware is a sideshow. From the viewpoint of Hutchins’ hunch, I answer: While I sympathize with your desire to know a source of the relative strengths and weaknesses of one part (the brain) of the human cogs, this knowledge might not be very important for explaining (say) economic growth, or wealth, poverty and inequality. You are looking at just one part (brain, not body) of what may be a relatively inconsequential cog. Put differently, you may be mistaken about what the most important processor is, how flexible it is, and how biologically determined it is. In recounting the evolution of Western navigation artefacts, Hutchins (Reference Hutchins1995: 155) notes that
. . . the existence of such a wide variety of specialized tools and techniques is evidence of a good deal of cultural elaboration directed toward avoiding algebraic reasoning and arithmetic . . . The kinds of cognitive tasks that people face in the wild cannot be inferred from the [CT of navigation] alone. The specific implementations of the task determine the kinds of cognitive processes that the performer will have to organize in order to do the task. The implementations are, in turn, part of a cultural process that tends to collect [external artifacts embodying specialized A&R] that permit tasks to be performed by means of simple cognitive processes.
To Hutchins, social hardware (in the form of both human cogs and external cognitive artifacts) does matter in the short run, but in the long run the hardware of distributed cognition isn't even given: It is culturally evolved to sidestep limits of human cogs and exploit their strengths. Both the software and hardware of distributed cognition are, as Hutchins says, “artificial through and through”. Neural biology isn't social cognitive destiny.
From the viewpoint of Hutchins’ hunch, doing rigorous Marrian cognitive science is overwhelmingly easier in social machines than in brains. I cannot put this better than Hutchins (Reference Hutchins1995: 128–9):
The basic computations of navigation could be characterized at the computational, representational/algorithmic, and implementational levels entirely in terms of observable representations. On this view of cognitive systems, communication among the actors is seen as a process internal to the cognitive system. Computational media, such as diagrams and charts, are seen as representations internal to the system, and the computations carried out upon them are more processes internal to the system. Because the cognitive activity is distributed across a social network, many of these internal processes and internal communications are directly observable. If a cognitive psychologist could get inside a human mind, he or she would want to look at the nature of the representations of knowledge, the nature and kind of communication among processes, and the organization of the information-processing apparatus. We might imagine, in such a fantasy, that at some level of detail underlying processes (the mechanics of synaptic junctions, for instance) would still be obscured. But if we could directly examine the transformations of knowledge representations we might not care about the layers that remain invisible. Any cognitive psychologist would be happy enough to be able to look directly at the content of the cognitive system. With systems of socially distributed cognition we can step inside the cognitive system, and while some underlying processes (inside people's heads) remain obscured, a great deal of the internal organization and operation of the system is directly observable. On this view, it might be possible to go quite far with a cognitive science that is neither mentalistic (remaining agnostic on the issue of representations “in the head”) nor behavioristic (remaining committed to the analysis of information processing and the transformation of representations “inside the cognitive system”) [all emphasis in original].
Neuroscience does not “look directly at the content” of any information representation in any brain as concrete, observable, rich and obvious as a ship's fix and projected bearing on a navigation chart, or a family's current and projected asset position as represented in its own external financial accounts and plans. As near as I can make out, they haven't yet agreed exactly where, or even how, the content of nouns like “ship” are physically represented in a brain (Martin and Caramazza Reference Martin and Caramazza2003). Hutchins is, I think, still correct in spirit after 13 years: What is still a fantastic voyage for neuroscience has long been feasible for social scientists examining distributed cognition systems. We can now (with the right ethnographic skills) literally step into these and directly observe just about every theoretically central entity of rigorous Marrian cognitive science.
For these reasons, I believe it may be fruitful to move away from studying what individuals do with simple decision problems and games, and toward studying distributed cognition in groups that confront complex problems and games. I now view my past work on social learning about hard economic problems (Ballinger, Palumbo and Wilcox Reference Ballinger, Palumbo and Wilcox2003) as my initial steps in that direction. Ballinger and I continue to study the evolution of advice (about hard economic problems) across generations of subjects. Advice is, after all, a relatively simple cultural-cognitive artefact, and one whose evolution we can follow over workable experimental timescales. Schotter and Sopher (Reference Schotter and Sopher2003) share similar interests with us.
Though it is rare relative to all work, some experimental economists have been contrasting the behaviour of individuals and groups. So far, though, these groups are overwhelmingly either unstructured groups or “symmetric committees” in which all group members have equal access to information, cognitive artefacts and each other, and all have equal sway in decision making (e.g. majority or unanimity rule). This is not distributed cognition in Hutchins’ interesting sense (though such groups may be an interesting control treatment when studying distributed cognition, and such groups may have, on occasion, endogenously created their own truly distributed cognition). Most distributed cognition in real groups has asymmetric and hierarchical structural features that divide tasks into subtasks, restrict data availability and information flows, and vest final decision making in only a subset of group members.Footnote 2
Asymmetry and hierarchy are consequential features of distributed cognition. When viewed from the perspective of its place in distributed cognition, we discover that some shopworn normative failings of individual cognition may be social virtues in correctly structured groups. Hutchins (Reference Hutchins1995: 240) shows how “. . . some ways of organizing people around thinking tasks will lead to an exacerbation of the maladaptive aspects of [confirmation bias], whereas other forms of organization will actually make an adaptive virtue on the group level of what appears to be an individual vice”. Hutchins’ simulation studies of networks of individual interpreters suggest that it is just those social networks with lots of symmetric linkage of individuals (e.g. symmetric face-to-face committees) that would exacerbate confirmation bias. We might want to look closely at some real economic groups (e.g. extended families, parts of firms and/or public bureaucracies) to identify some typical structures of distributed economic cognition: Those might be good candidate structures for study. Symmetric committees are probably quite rare in distributed economic cognition – and with good reason, given Hutchins’ arguments.
Told that I could select only one tool from a neighbouring discipline, it would be ethnography, not medical imaging. A simple and false choice, I hear you cry: It excludes a large middle of more complex alternatives you actually have, and also ignores a rich division of cognitive labour across researchers and disciplines. Fair enough. But if that cry was yours, then have some sympathy for individuals in typical experiments: They too are confronted with simple and false choices that exclude many middles (and extremes) they'd actually face in the field; and they too are sealed off from advisors, artefacts and other distributed cognitive resources they'd have in the field. The situated character of individual cognition in the field is one reason Harrison and List (Reference Harrison and List2004) advocate more field experiments. Understanding situated individual cognition is almost definitionally crucial to understanding distributed cognition systems. But Hutchins’ hunch that the theoretical entities of classical cognitive psychology are better applied to distributed cognition than individual cognition is quite a bit beyond that.
I'm well aware that the ideas set out above will raise many hackles. To spare others the effort, I'll hurl the epithet “group selectionist” at myself, since I seem to advocate the fusion of individual persons into a larger agent. I'll also admit that I'm personally not all that interested in dealing with “crucial foundational problems” about groups. For instance, subjects seem to effortlessly overcome the pull of many bad equilibria when they can talk, observe and gossip the way they actually do in the real world (Duffy and Feltovich Reference Duffy and Feltovich2006). Yes, it is an interesting question why that is so: I hope for good answers and look forward to reading them with deep interest. No, the serious study of distributed economic cognition does not need to, and should not, wait for good answers to that question. So don't bug me with the predictable question: “But what about the (insert favorite game-theoretic puzzle) problem you are obviously ignoring in this group?” One person's – even one paper's – foundational problem is another's distracting detail.
More seriously and sympathetically: Economic theorists have already added important insights to a full understanding of distributed cognition, and I hope this continues. By now it is banal that economic institutions are important external cognitive artefacts that mediate specific kinds of distributed cognition at many scales: Philosophers of mind such as Clark (Reference Clark1997) are now getting this. At the smallest scales, within groups, smooth distributed cognition might very well be aided and abetted by specifically economic institutional gizmos that help solve problems of coordination, shirking, trust and so forth. Yes, of course – good thinking. But distributed cognition is not merely the solution of those economic problems: Solving them is no guarantee that distributed cognitive systems will increment and innovate their own hardware and software so as to grow an economy. More generally, distributed cognition systems, and economic mechanisms, are almost necessarily interdependent. I view neoclassical economic theory as a necessary part of a complete and rigorous cognitive science but, unlike Gul and Pesendorfer (Reference Gul, Pesendorfer, Caplin and Schotter2008), I think neoclassical economic theory just doesn't address certain important economic issues – genuinely cognitive ones below the level of computational theory. And I'm beginning to doubt that the most important kinds of economic cognition occur solely in machines that only reveal their secrets to medical devices. That is a criticism of our general obsession with individual cognition, and neuroeconomics is just one manifestation of that.
