Georg Theiner --- How to Argue for Group Cognition: A Guide for Naturalists
New Waves in Philosophy of Mind
bryce
Thanks for sending along a fun and interesting paper. Unsurprisingly, there's a lot here that I agree with, and a lot here that I'm skeptical of. I'm still struggling to get a handle on the precise conditions under which group cognition is supposed to emerge. So, let me start off slowly with a question about your use Woolley et al. (2009). You claim that "social sensitivity" is doing the work of a collective intelligence factor. I would have thought that this is precisely what every smart individualist since (at least) Mill would have expected, so I'm wondering if you can tell my why the individualist reading of this result is wrong. If you don't have individuals who are working well together, you should expect aggregate performance to taper off; if you have individuals who work well together, pay attention to one another's expertise, and do so in a sensitive way, the team dynamics should be more fluid and successful. But that seems—at first blush—to be a fact about the way that individual sates and processes work in a particular social context, not a claim about anything particularly "groupish". So, can you say a bit more about why this is an interesting fact relative to your claims about group cognition?
Thanks a bunch, and more soon!
Bryce
Georg Theiner
Hi Bryce,
Thanks a lot for taking a look at my draft. I really appreciate your comments. Ok, so in response to your “deflationary” individualist response, I should clarify my claim about the c-factor, and two related assumptions that I probably should state more explicitly in connection with the ant raft case:
(1) (1) Social sensitivity is a part of the total realization of the c-factor.
(2) (2) The “wide functionalist” premise that the group-cognitive contribution of a given type of individual (mental and non-mental) state, and the type of social interactions among individuals, is determined at least in part by the role that they play vis-à-vis intelligent group-level behavior. (Example: the raft-building contributions of the claws, mandible, and adhesive pads at the legs of individual ants.)
(3) (3) The realization of a psychological property P which is a natural kind-ish predicate at the group level (Example: the c-factor) need not be – and in many cases, will not be – a natural kind-ish predicate at the level of individuals.
I don’t think it’s always intuitive to discern the role described in (2). Consider the conflicting evidence about the role of trust for the formation of transactive memory systems (TMS). In some studies, trust seems to foster collaborative inhibition, while in others it seems to facilitate collaborative facilitation. This is partly, I think, because some earlier studies did not sufficiently distinguish between emotion-based and cognition-based trust (the latter being more conducive to collaborative facilitation).
There is a difference between the role of social sensitivity vis-à-vis collective intelligence a la Woolley, and the role of trust vis-à-vis transactive memory a la Wegner, though, and that has to do with the fact that the c-factor is a much more domain-general psychological predicate then TMS. In most studies of TMS, trust is not analyzed/measured as a constitutive component, but rather as a modulating factor. In contrast, it seems more plausible to think of social sensitivity, and the way in which affects group interactions, as part of the realization of the c-factor. But the difference between constituents and contexts is fuzzy, of course; and the fuzzier it is, the more reason we have to think that (3) is true. And that, I think, is because the more interesting cases of group cognition will score relatively high on the scale of Wimsatt-emergence. (Hence the decompositions won’t look as natural kind-ish. Compare: the way in which propositional attitudes are realized in a neural network.)
Looking forward to continuing our discussion,
-Georg
Bryce Huebner
Bryce Huebner
Georg Theiner
Hi Bryce,
Good idea, let’s get our intuitions re: Wimsatt-emergence on the table first. In the paper with Tim O’Connor (2010), we used Wimsatt-emergence (together with the other two suggested dimensions) to score a few illuminating cases. I think I changed my mind on some of them, so here we go:
Let’s begin with a well-oiled TMS that’s developed for remembering how to perform specific tasks collaboratively, as a group. I focus a lot on the comparison between interacting and nominal groups (i.e. pool non-redundant info) in my paper, and under which conditions you can get collaborative inhibition vs. facilitation, because that seems to be a way of operationalizing the Wimsatt-criteria. TMSs in this sense clearly fail condition 1, because of the differentiation of expertise. They fails condition 2, in the sense that (a) group performance plummets dramatically if you knock out the relevant expert(s), (b) process losses typically increase with group size (but this could be a linear effect). They fail (c) in the sense that groups that are trained together and then remember as groups perform better than groups trained individually and then (re-)assembled as a group. But in the long run, those groups develop a kind of modular organization (near-decomposability), and hence meet condition (d). As Bechtel & Richardson already pointed out, condition (d) is the critical one. Those would be cases in which eg collaborative retrieval strategies are so highly interactive that they significantly change what people remember in dyads or teams vs. in isolation. (John Sutton and his group have analyzed the dialogues that occur in the course collaborative remembering in this respect – interesting stuff).
The distinction between instances that fail 1-3 and those which fail all 4 is crucial, and generalizes across many different types of group cognition. For instance, in their normal mode of operation, Hutchins-style navigation teams meet condition 4, and thus are really not all that emergent sensu Wimsatt. Philosophers have focused on those, because they can be described as a kind of group-level production system with a modular organization, and that seems to fit their preconceptions of how individual minds are organized. But I think Hutchins’ analysis of situations where things go awry, standard operating procedures fail, new strategies, behavior, and perhaps also meanings have to be negotiated on the fly, and learning is described as an adaptive reorganization of the complex system as a whole are more interesting cases of emergent group cognition, because condition 4 fails. As a general rule, I would say that intelligent group behavior whose coordination depends on high-bandwidth, interactive communications are likely to score highest on all 4 dimensions of Wimsatt-emergence. Back in 2010, we used a longitudinal study of choreographic cognition, i.e., the emergence of creativity in groups of choreographer & dance practitioners as our example. Keith Sawyer has a great book on improvisational dialogues; also, much of the literature on creativity in organizations clearly falls into this category.
Now, let’s go back to your example of collective decision-making. Groups can adopt different aggregation procedures to form collective decisions (cf. List & Pettit 2011), and the properties of those aggregation schemes themselves can be classified according to Wimsatt-emergence. For instance, one can take the three conditions of universality, anonymity, and systematicity as a kind of “aggregative” default. But as List has shown with his Impossibility Theorem, no group can satisfy all three of them at one and the same time. This puts an upper limit on the collective rationality of groups. So you can view the deviations from those as a kind of “collective epistemic action” by which groups take to circumvent its bounded rationality. For instance, adopting a distributed premise-centered procedure fails at least conditions 1 and 3 of Wimsatt (cf. Theiner, in press).
About “wisdom of crowd” cases, I keep going back and forth. My intuition still is that those do not score all that high on the Wimsatt-scale, because it’s important to maintain a great amount of independence among individual information-gatherers or decision-makers. Social signaling is essential, of course, but that is only a very narrow bandwidth of communication. Do such collectives fail condition 1? Well, yes, if you count the diversity of information to which they ought to have access for the effect to kick in, but not necessarily in their cognitive processing. They fail 2 to some degree, because it matters how many individuals you have if you don’t satisfy the two optimality conditions laid out by Bettencourt (2009); but again, this may be linear. But the secret of their information-theoretic success seems to be precisely that they meet conditions (3) and (4). This would seem to imply that the relevant sense in which the wisdom of crowds is emergent is a genuinely information-theoretic sense, and not organization-dependence sensu Wimsatt. They may even be inversely related. I’m esp. curious what you think about “wisdom of crowds” cases.
And yes, we should absolutely get together real soon! Maybe we can even organize a workshop or something in the future...
Best,
-Georg
Bryce Huebner
I think that I am just getting more curmudgeonly as I get older, and the more I think about collective mentality, the harder I think it is to establish. I've actually been thinking a lot about each of the cases that you raise, and I think we pull in different directions on some of them. First off, though, le me say that I think there are liely to be cases where we end up with distributed systems that are systems, but which are not cognitive systems. One of the things at I try to do in my book on collective mentality is to pull these claims apart in a way that makes sense of the intuitions that pull oth ways on group cases. While I'm happy to be groupy at e level of systems, I think that individualist arguments typically win out where cognition is concerned. But, that's a long argument, andi can't get into the details here.
I just finished a paper if my own on TMSs, and i end up more skeptical than you. You are right that som grops end up with a crypto-modular organization, but I worry that the relevant mechanisms are not integrated in a way that allows for distributed processing. The cases where you get interactive strategies for recall are better, but I think you need a mechanistic account of these systems that can make it clear where the inputs, interfaces, and processing are going on. Maybe your paper on TMSs gives this, but I haven't had a chance to read it yet. I would say something similar about Hutchins style cases (and about Kirsh on high traffic cafes) but I do like your suggestion about how the navigation crew adapts to breakdowns. I think we are perfectly in agreement there, though I'm going to want to see the mechanistic story in each case that makes it clear how the information processing cascades work in each case.
I think I'm on record as arguing against this sort of model for collective decisions a la. List & Pettit (2011). I have a review of this book in Ethics where I argue that it is a mistake to think that these decision making patterns justify a claim about collective mentality. I think that inference relies on a thin functionalism about mental states lie beliefs and decisions that I just don't buy--but I'm presenting that stuff at a conference where List will be next week, and we'll see if he canges my mind. Finally, I spend quite a bit of time on “wisdom of crowd” cases in my book. I'm skeptical because of the "independence among individual information-gatherers or decision-makers" that you note. I think we just agree on this sort of case.
Also: It would be totally cool to organize a workshop on collective mentality...
Georg Theiner
Hi Bryce,
First off, congrats on the new book, and I’m looking forward to reading it. Count me in if you plan on having an online conference on your draft or something similar. As for getting older, I think I’m experiencing the exact opposite reaction – the older I get, the more “instrumentalist” I get about attributions of (non-phenomenal aspects of) mentality in general.
Compared to extended cognitive systems like Otto+notebook, whose ontological integrity, robustness, and persistence may be said to be lacking at times, there can be no doubt that social systems form a genuinely new level of organization and “locus of control,” capable of exhibiting system-level behavior and properties. The Wimsatt-criteria can be used to classify both psychological and non-psychological properties of such systems, and I also believe that Bechtel/Craver-type mechanistic explanations can be used to describe the operation of such systems at multiple levels of organization. A lot of interesting insights can be gained from this perspective alone, regardless of whether the group-level properties under consideration meet any putative “mark of the cognitive” or whether we scrap the notion of cognition, and rather talk about information-processing, adaptive problem-solving, enactive sense-making or whatever.
Plus, similar to the debate over extended cognition, one can presumably rank claims regarding group cognition in terms of their respective strength: group cognitive systems (weakest) – group-cognitive processes (stronger) – group cognitive states (even stronger, esp. depending on the type of state). Perhaps distributed cognitive systems do not always support the strongest construal, except in the general sense that we attribute some type of psychological capacity to the group as a whole. This is by no means trivial, because – as Bechtel and others are fond of pointing out – there are no mechanisms per se, but only mechanisms for certain phenomena of interest. And those group-level attributions of psychological capacities are certainly one of those phenomena of interest. And relative to such an analysis, we can equally “look down, around, and up” as we can do at the level of individuals (or brains). That’s essentially the construal I’m advocating for TMS, and for capacities such as memory, learning etc. at higher organizational levels as well.
As for joint cognition, such as collective belief/desire psychology, collective agency, or collective knowledge, those attributions may require higher forms of social integration to be plausible. (You can’t sue wiki communities for a reason). You say that those attributions typically rely on an insufficiently “thin” variant of functionalism to be plausible. By “thin,” do you mean (a) coarse-grained, (b) disembodied, or (c) based on controversial intuitions (like good-old fashioned commonsensical, analytical functionalism? Or maybe all three? You are probably right that the philosophical literature that has focused on those attributions often tends to be a bit aprioristic in its methodology, and perhaps this has caused the impression (seemingly in Rupert 2011) that there can’t be any empirically grounded “naturalistic” arguments for those kinds of states. I tend to think that this bias is more a result of the philosophical methodology employed by those authors, rather than based on the subject matter at hand; and that there can be equally naturalist studies of those phenomena (e.g. studies of collective knowledge in socialized philosophy of science). On the other hand, there are, of course, different views of what grounds the attribution of folk-psychological attitudes in individual minds, and I’m not convinced that there is such a deep asymmetry between the two levels. (BTW, how would your now-curmudgeonly self interpret your earlier experimental studies?) Let me know what List has to say about this.
Re: workshop: I can think of a number of other people who'd also be interested in this, as I’m sure you can, too…
Cheers,
-Georg
Orestis Palermos
Georg Theiner
Hi Orestis,
Thanks a lot for your comments – I’ll be sure to check out your paper; is it available online?
As
for your question: Yes, I agree that one of the standard reductionist moves is
to argue that group cognition as it is conceived here would be epiphenomenal,
assuming the “causal completeness” of individual-level regularities and the supervenience
of group cognition. That is, for every
aspect of an individual’s behavior, there is a sufficient cause which refers
only to properties of that individual and her social interactions, but does not
take into account any psychological or non-psychological (e.g., behavioral)
properties of the entire group to which that individual belongs. We discussed this objection in Theiner &
O’Connor 2010, Section 4.1. (due to space constraints, I decided to focus on other aspects here which I didn't get to discuss in that paper).
Very briefly, we argued that whatever strategy works for the non-reductive physicalist to avoid the “causal exclusion” problem for individual mental states, should in principle work here too. Here are some of the usual moves:
(1) A “Dual explananda” strategy: group cognition is posited to explain group-level behavior, individual cognition is posited to explain individual behavior. No causal competition between wholes and parts. Problem: once you factor in social interactions, which are part of the total realization of group cognitive states, the problem recurs.
(2) “Causal inheritance” solutions: Higher-level properties “inherit” the causal powers of their realizers: they exercise their causal influence through the causal powers of the lower-level properties.
a. For example, the Jackson & Pettit (1990) “program model” of causal relevance falls into this category. List and Pettit recently seem to favor that strategy.
b. Appeal to counterfactual dependence to ground causal relevance. Argue that one and the same event-token can be causally relevant on more than one level, in virtue of falling under multiple event-types which are not nomologically coextensive. From here, we briefly developed a reply based on the formalization of “preemption” in Loewer (2002).
I don’t think I need to be committed to any strong versions of downward causation. An unproblematic version would be e.g. the formation of “socially manifested” individual-cognitive properties, like the group emotions discussed by Seger and Mackie, in which one’s membership and participation in certain group-level activities causes me to form a mental state M which reflects those activities. As long as we can identify such a mediating individual-psychological state M, the group-level activities would not directly exercise any causal influence on my behavior. Stronger versions of downward causation may be cases like emotional contagion, social mimicry, social entrainment (similar to the cases described by LeBon in his book on the “crowd mind”) all the way up to purely dynamicist models of leadership formation etc. (I’m sure you’re discussing those in your paper), because one could argue that there aren’t any mediating individual-psychological states (though it’s not always clear). This will depend on how we interpret “downward causation” in DST-models more generally. Many of the cases of group cognition that I have in mind are not like that, because they are dependent on, and partly constituted by fairly complex mental representations that are had by individual minds (like the members of a TMS).
On the issue of TMS: I changed my mind on this a bit. In Theiner (2010), I did argue that TMSs violate condition (4), but now (cf. forthcoming) I don’t think that’s always the case. In the short run, esp. when people still need to learn about one another’s knowledge, ideas, and skills, communication is paramount. But in the long run, once there is a well-defined division of cognitive labor, individual will have developed very detailed, accurate, and stable representations of who knows what, and act (and interact) accordingly (Unless there is a breakdown or failure or membership turnover.) In that case, acts of communication will become less and less important to predict the outcome at the group level: we would essentially just have to look at what each individual knows (both in terms of individual and transactive memories, interrole knowledge etc.). I take this to be an indication of a progressive modularization of TMSs (which I take to be defined by a failure of 1-3, but not 4).
I think the fact that we can look at a TMS on different timescales has sometimes caused a bit of confusion about the role of communication in TMS. For example, the study by Moreland and Myaskovsky (2000) was explicitly designed to show that the performance benefits of group training are due to TMS, rather than communication (!). Their argument was that groups whose members were trained apart, with no chance to communicate with one another, performed at the same level as “communicating” groups (who were trained together) after individually receiving explicit information (handouts) about the distribution of expertise in the group. But I think that there are several problems with this study. First, giving people handouts that summarize the distribution of knowledge is essentially a functional substitute of communication, rather than saying that there is no communication at all; and the experimental method is biased against other functions of communications in TMS (see Propp 2003). However, I can see how TMS can develop, in the long run, a certain amount of modularity, and the communication bandwidth between members in the performance of routine group tasks is fairly minimal; in that case, I would say that their operation as a TMS would be near-decomposable.
Does this make sense?
Best,
-Georg
Orestis Palermos
Georg Theiner
Hi Orestis,
Very interesting research projects you have going on there. If I understand the description of your dissertation correctly, your reconciliation project seems to run like this:
(1) Viewed through the lens of socially distributed and extended cognition, the social relationships which are the topoi of traditional social epistemology are essentially “cognitive” in nature.
(2) Epistemic group agents are really higher-order individuals (“supra-indidividuals”) themselves.
(3) Hence we can preserve some of the main intuitions of mainstream individualistic epistemology (e.g., virtue reliabilism), by showing that they can be applied to supra-individual agents.
This makes s lot of sense to me. I pursue a similar line in a forthcoming paper (“Onwards and upwards with the extended mind”) in which I argue that we can generalize Kirsh and Maglio’s (1994) notion of “epistemic action” from individual to collective epistemic action, and use that as a way of synthesizing a variety of research on how groups circumvent cognitive limitations that arise from social aggregation. (You can find a draft online).
Re: the other move you mention, concerning downward causation: Yes, I agree that group-psychological properties are usually multiply realizable (that was the third notion of emergence we discussed in the earlier paper), and if there are lawful regularities about groups in terms of their cognitive/epistemic properties, they will often not be preserved at the individual level. However, as you point out, all by itself, this may not be sufficient to block the threat from epiphenomenalism, unless you have an extra argument showing that the relationship between higher-level functional properties and their (social, in this case) realizers is not a zero-sum game. And then you’re back to the downward-causation problem (“Descartes’ revenge,” as Kim would say.)
For what it’s worth, I tend to think that many social properties will turn out to be multiply realizable in this sense, whether or not they qualify as being cognitive in kind. Non-reductive individualism (NRI) in general seems, to me at least, a promising way to overcome the old-fashioned dualism between methodological individualism and collectivism in the social sciences. I like Keith Sawyer’s (2003a, 2003b), papers about NRI, and his 2005-book on social emergence in this respect, although he does not make an explicit move towards group cognition in those papers.
Best,
-Georg