This excerpt is provided, in screen-viewable form, for personal use only by members of MIT CogNet. Unauthorized use or dissemination of this information is expressly forbidden. If you have any questions about this material, please contact cognetadmin@cognet.mit.edu

6Guiding Inquiry: The Theoretical Roles ofSystemic Functions

The final arbiter in philosophy is not how we think but what we do.

—Ian Hacking 1983

The theory of selected functions, as we have seen, aspires to explicate theconcept of functions employed in modern biology. The aim is to specifyhistorical conditions that biologists implicitly or explicitly have in mindwhen attributing functions. And according to the theory, the conditionsbiologists have in mind concern the selective success of functional traits;biologists attribute functions when they believe that the persistence ofthe functional type is caused by the selective success with which ancestraltokens performed the specified task. By contrast, the theory of systemicfunctions aspires to understand the concept of functions by consideringthe roles it plays in inquiry into hierarchical systems. The emphasis is onthe theoretical work that function attributions actually accomplish ratherthan the way the concept is implicitly or explicitly understood. Emphasisis on what life scientists actually do and not on what we think they havein mind.

Systemic functions play several roles in inquiry. These include the his-torical explanations to which advocates of selected functions appeal. Thetheory of systemic functions, as explicated in chapter 3, attributes func-tions to traits the effects of which contributed to their own selective suc-cess and thus contributed to evolution or stasis in the population at large.The theory thus can explain everything explicable from within the theoryof selected functions. But systemic functions have other important theo-retical roles. Some of these roles depend upon the extent of our knowl-edge. We can explain, for example, how various lower-level capacities of

158 Chapter 6

known systems give rise to higher-level capacities, as the example of thecirculatory system demonstrates. But when faced with unknown systemsor parts of systems understood only partially, the role of systemic func-tions differs. Guiding, not explaining, is the focus. Systemic functions,when employed in systems we do not yet understand, guide us in ourinquiries. They help us navigate the system by providing a provisionalconceptual map, a provisional chunking of phenomena into tractablecomponents. The ultimate goal, to be sure, is to explain how the systemworks, but that is accomplished only when we discover bottom-up mech-anisms; the immediate goal is to provide entree into the structure of thesystem. So the view that function attributions are explanatory, while trueenough, is incomplete. The aim of this chapter is to consider the broaderrange of inquiry-guiding roles that our attributions of functions occupy,especially when faced with a system understood partially or not at all.

My thesis is twofold. I shall argue that the theory of systemic functions,in its various inquiry-guiding roles, offers a thoroughly naturalistic ap-proach to understanding functions. This can be seen at two levels—thetheory taken as a whole and the attribution of specific systemic functions.Specific attributions are naturalistic because the theory requires evidencefor physical mechanisms that instantiate the attributed functions. Thetheory as a whole is naturalistic because it is an integral part of inquiryinto hierarchical systems and because the ontology of the theory is thor-oughly naturalistic. I shall also argue that, because the theory of systemicfunctions eschews properties that are abstract or noncausal, it cannotaccount for the occurrence of malfunctions as usually conceived. The al-leged normative status of functional properties must be understood differ-ently. To this end, I continue the Humean speculations begun in chapters4 and 5. I suggest that a token trait that has lost the relevant systemiccapacity no longer belongs to the associated functional kind, in whichcase it cannot be classified as malfunctioning. Nevertheless, given ourknowledge of the larger system, we expect such tokens to exercise therelevant capacity. And when our expectations are thwarted by damageor disease, we are inclined to persist in seeing the token in terms of theseexpectations; we persist in placing the damaged and diseased in the func-tional type. That is why we are inclined to (mistakenly) claim of suchtokens that they are malfunctional.

Guiding Inquiry 159

I The Inquiry-Guiding Role of Systemic Functions

We should, I believe, reject the general methodology employed by thetheory of selected functions. As I argued in the previous chapter, weshould reject the attempt to understand functions by examining the con-ditions under which modern biologists apply the term. Biologists are con-cerned to understand specific biological phenomena, but the task ofintegrating their concept of functions into the natural sciences generallyis not a core concern. It is of concern, however, in the philosophy ofscience. The task is to discover whether or not we can understand func-tions in a way that coheres with the methods and postulations of thenatural sciences. We should not ignore the way biologists employ theterm or the concept, but we ought not assume that their usage is the re-sult of careful reflection upon the philosophical question. We ought toapproach the topic of functions differently.

I believe we do better to consider the theoretical roles that functionattributions fulfill—the tasks accomplished by the attribution of func-tions when trying to render intelligible the workings of hierarchical phe-nomena. Instead of asking, Under what conditions do biologists attributefunctions? consider the question, What theoretical work is accomplishedby attributing functions to natural traits? or, What are we, qua inquirers,doing with our function attributions? Answering these questions, I sug-gest, is far more likely to lead to a satisfactory naturalistic account offunctions. On this view, systemic functions are action guiding and, morespecifically, inquiry guiding.

The prevalent view, by contrast, is that function attributions are ex-planatory. The theory of selected functions, in particular, asserts that theattribution of a function is equivalent to a selective explanation of thepersistence of the functional type. I do not deny that function attributionsare sometimes explanatory in this way, but I do not agree that this sortof explanation is the main theoretical role that our attributions of func-tions play. Instead, attributing functions often serves to guide our inquiryinto hierarchical systems we do not yet understand well. Our attributionsaccomplish this in two general ways. First, the attribution of systemicfunctions provides a top-down taxonomy of a system that is highly provi-sional in character. This taxonomy does not explain, but rather provides

160 Chapter 6

a preliminary map with which to parse the system and study its functionalparts. As we uncover increasingly specific systemic components, we reviseor extend our initial taxonomies. And, second, until we discover quitespecific physical mechanisms instantiating the systemic capacities postu-lated at the lower levels of analysis, we must refrain from claiming thatour function attributions are secure. The systemic capacity analysis is top-down at the outset, but it cannot be accepted in full until its variousattributions are tested from the bottom up. It is this that makes somefunction attributions explanatory: We can explain why a system behavesas it does once we have confirmed a taxonomy of functions by pointingto the systemic mechanisms that instantiate the relevant systemic func-tions. This is just to say that our explanations, insofar as they are mecha-nistic, require evidence of the mechanisms to which we appeal. On thisview, then, function attributions are inquiry guiding before they are ex-planatory. They first provide a conceptual wedge into complex phenom-ena and a set of signposts to guide us toward the causally efficaciousmechanisms of the system.

The entering point in any such investigation is the higher-level systemiccapacity we wish to understand. Understanding the implementation ofsome such capacity is the goal of our analysis. It also marks the point ofdeparture for our top-down attempt to taxonomize the system into salientcapacities and components. For systems that are already understood toa significant degree, both the top-down taxonomy and the instantiatingmechanisms are before us, in which case our attributions serve to explainhow the system works, how it instantiates some higher-level capacity.Any system for which we can produce a mostly complete flowchart ofits operations is one for which the inquiry-guiding role of systemic func-tions is greatly diminished while the explanatory role is greatly enhanced.But for systems we do not yet understand or understand only partially,our function attributions serve as a guide for increasingly detailed inquiryinto the system. I will illustrate with two relatively simple examples, thefirst discussed by Bechtel and Richardson (1993) and the second by Enc(1979).

Consider the early investigations of the brain conducted by Franz Jo-seph Gall during the first half of the nineteenth century. Gall rejected thedominant view that the brain is a homogeneous unit resistant to mecha-

Guiding Inquiry 161

nistic analysis. He hypothesized that, to the contrary, the brain is com-posed of several organs or centers, each responsible for a quite specificcharacter trait. His aim was to provide a mechanistic analysis of variousintellectual capacities by postulating a variety of brain organs. He postu-lated distinct cerebral organs for capacities such as speech, understandingtime, navigating the environment, and so on. He also maintained thatthe various organs of the brain develop differently across individuals,growing larger in some and smaller in others, thus accounting for individ-ual differences in the relevant capacities. The larger the neurological or-gan, the greater the correlated capacity. Assuming that the cranium issufficiently malleable to show individual differences in the size and shapeof these brain organs, Gall thought he could determine the degree towhich an individual possessed a given character trait by examining por-tions of the cranium. Hence the practice of craniology.

Although Gall’s phrenology came in for a great deal of criticism andderision, his basic strategy was commendable. Along with fellow phrenol-ogist Johan Gaspar Spurzheim, Gall believed they had ample evidencethat our psychological faculties were the products of our brain. Spurz-heim called this the ‘‘First Principle of Phrenology.’’ The brain, they as-sumed, had to have sufficient internal complexity to underwrite the widevariety of psychological capacities observed in human beings. Havingthus rejected the view that the brain was unanalyzable, they consideredthe question, What are the functions of the brain such that human beingshave the intellectual and practical capacities they in fact exhibit? This isto assume, as a provisional guide to further inquiry, that the brain issusceptible to a systemic function analysis. Now, the particular answersoffered—that there is a neurological organ corresponding to each intel-lectual or practical capacity and that we can discover such organs viacraniology—were mistaken, although a remnant of Gall’s organologysurvives in contemporary cognitive psychology in the thesis of modular-ity. But the basic approach—analyzing the system into functional partsrelative to some specified set of higher-level capacities—was clearly ontrack. Thus the inquiry-guiding power of systemic functions.

But this does not exhaust the role of systemic functions in this story.The empirical investigations conducted by phrenologists never ma-tured. Gall began by developing a top-down taxonomy of neurological

162 Chapter 6

functions intended to account for intellectual and practical capacities.Unfortunately, the only confirmation for this taxonomy came in the formof correlations between features of the cranium and capacities of the intel-lect. The systemic function analysis failed to progress further. In particu-lar, it failed to descend to the capacities and mechanisms within the brainitself. The claim that there existed a one-to-one correspondence betweencranial characteristics and capacities of the intellect, even if confirmed byobservation, does not tell us how the neurological system instantiatescapacities of the intellect. Of course, the hypothesis that the brain is com-posed of distinct organs, each responsible for a specific capacity of theintellect, is a nontrivial hypothesis. It stakes a claim about the structureof our brains. But this claim was never put to the test by phrenologists;they failed to pursue the bottom-up portion of the theory of systemicfunctions. Had they done so, they would have come under strong pressureto relinquish their view. Indeed, once it was discovered that the brain isnot organized in the ways described by phrenology, the proposed top-down taxonomy was no longer tenable. This demonstrates more vividlythe inquiry-guiding role of systemic functions: The discovery that thepostulated capacities are not instantiated in any systemic mechanismsprovides powerful reasons for rejecting the proposed taxonomy and be-ginning the search for an alternative.1

Now consider a case in which our knowledge of the relevant system,while incomplete, was not nearly so immature. James Watson wanted tounderstand the synthesis of complex proteins. This process involved therather precise ordering of different amino acids. Watson wanted to knowthe mechanisms responsible for it; he wanted to understand how the pro-cess worked. He reasoned that it could not be accomplished by specialenzymes, since enzymes are themselves complex proteins formed by thevery process in question. He thus hypothesized that the ordering was ac-complished not by further enzymes but rather by some sort of templatethat had the capacity for self-replication. Watson put his suggestion thisway:

It is therefore necessary to throw out the idea of ordering proteins with enzymesand to predict instead the existence of specific surface, the template, that attracts

1. See Bechtel and Richardson (1993), chapter 4, for a fuller discussion.

Guiding Inquiry 163

the amino acids (or their activated derivatives) and lines them up in the correctorder. . . . It is furthermore necessary to assume that the template[s] must alsohave the capacity of serving, either directly or indirectly, as templates for them-selves (self-duplication). (Watson 1970, 179–80, quoted in Enc 1979, 352)

As Enc stresses, Watson’s appeal to a template is clearly intended to referto some mechanism, as yet undiscovered, that instantiates the functionof ordering amino acids. Unlike the situation facing Gall, Watson hadample background knowledge from which to reason. He could say quitespecifically what the template could not be: It could not be an enzyme.He could also say that the template had to be some feature within thecell endowed with the capacity for self-duplication. All of this is to saythat the postulation of a template—a systemic capacity of cells—playeda crucial role in the discovery of messenger RNA. The attribution of asystemic function guided the search for the RNA template. It provided apiece of the taxonomic puzzle that then imposed constraints on the sortsof mechanisms for which to search.2

Enc (1979) discusses other examples—including Harvey’s discovery ofthe function of the heart—that illustrate the inquiry-guiding role of sys-temic functions. He does not put the point in terms of guiding inquiry;in fact, he asserts that function attributions are explanatory. But it is clearenough from the case of messenger RNA that ‘‘explanatory’’ covers abroad range of theoretical contexts, including those that serve to guidethe process of inquiry rather than explain the occurrence of certain events.Moreover, Bechtel and Richardson (1993) develop a wealth of splendidcase studies that illustrate the inquiry-guiding role of systemic functionsin detail. Now, Bechtel and Richardson are not concerned to defend anyparticular theory of functions. They are concerned, rather, to defend theclaim that progress in scientific discovery often involves the applicationof at least two heuristic strategies, namely, decomposition and localiza-tion. But these heuristics overlap the strategy of the theory of systemicfunctions in significant ways. In consequence, the several case studies de-veloped in Bechtel and Richardson serve as evidence for my claim thatsystemic functions play a central role in guiding inquiry into complexhierarchical systems.

2. For fuller discussion of this case see Enc (1979), 352–55.

164 Chapter 6

To see the overlap between Bechtel and Richardson’s heuristics andthe theory of systemic functions, consider first the process of decomposi-tion. Their account of this process is sophisticated and nuanced, but thebasic steps involve the delineation of a system (or what we tentativelytake to be a system), identification of higher-level capacities we wishto understand, and analysis of the system into functional componentsthat together produce the higher-level capacities. Like analysis in termsof systemic functions, the process of decomposition is recursive, pro-ceeding as far as our interests drive us or as far as systemic complexityadmits. The process of decomposition thus applies only to systems thatare hierarchically organized—that is, systems that can be analyzed intodistinct levels of organization. The ultimate goal of this top-downstrategy is discovery of physical mechanisms internal to the system re-sponsible for the functional capacities listed in our taxonomy. This in-volves the second heuristic, localization. It is assumed that the capacitiesof interest are instantiated by mechanisms within the system. They neednot be discrete or otherwise tidy; they may be distributed throughout thesystem. But until we uncover substantial evidence that there are physicalmechanisms the effects of which give rise to the identified capacities,we must temper the confidence with which we accept our functionaltaxonomy.

The strategies of decomposition and localization, or close cousins ofthese heuristics, are part of the theory of systemic functions as developedby Cummins (1983) and as developed in chapter 4 of this discussion.Application of the theory of systemic functions requires identification ofa system, a systemic capacity we wish to understand, and an analysis ofthe system into functional capacities at two or more levels of organiza-tion. This is Cummins’s analytical strategy. Its similarities to the processof decomposition are obvious.3 Cummins further describes what he callsthe instantiation strategy. It involves tracing the links of causation to thephysical mechanisms operating at lower levels of organization. Instantia-tion and localization are close cousins indeed, for both require that the

3. Cummins does not restrict application of the theory to hierarchical systems,at least not explicitly. That restriction is unique to my formulation of the theory,though inspired by Bechtel and Richardson (1993).

Guiding Inquiry 165

top-down strategy be confirmed by the bottom-up identification of caus-ally responsible mechanisms. The importance of bottom-up confirmationis emphasized by Bechtel and Richardson throughout their discussion,including the following:

Showing how systemic functions are, or at least could be, a consequence of sub-tasks [those specified in the decomposition] is an important element in a fullymechanistic explanation. Confirming that the components realize those functions[by way of localization] is also critical. Both are necessary for a sound mechanisticexplanation. (Bechtel and Richardson 1993, 125)

And as we saw in chapter 2, Cummins insists that ‘‘[f]unctional analysisof a capacity C of a system S must eventually terminate in dispositionswhose instantiations [in physical systemic mechanisms] are explicable viaanalysis of S. Failing this, we have no reason to suppose we have analyzedC as it is instantiated in S’’ (Cummins 1983, 31).

That said, there are differences between the two views; I discuss onesuch difference in the next section. But in the context of defending thetheory of systemic functions, the similarities outweigh the differences.The case studies offered by Bechtel and Richardson thus provide powerfulevidence of the inquiry-guiding role of systemic functions. All of theirstudies trace the dynamic processes involved in scientific discovery. Theyare concerned to explicate psychologically realistic strategies with whichwe make scientific discoveries. They focus upon complex and hierarchicalsystems which, at the relevant historical juncture, were understood par-tially or not at all. Their claim is that, among the various strategies webring to inquiry into such systems, decomposition and localization areof great utility. Their argument for this claim consists in an examinationof several historical cases of scientific discovery, including discovery ofthe mechanisms involved in fermentation, respiration, language produc-tion and comprehension, spatial memory, developmental genetics, andmore. In each case, the attribution of systemic functions plays the roleof providing a conceptual roadmap, directing our attention to the likelysorts of mechanisms that instantiate the capacities we wish to understand.And in each case, the confidence with which we attribute such functionsincreases as we confirm, from the bottom up, the existence of the requisitesorts of mechanisms. Hence the naturalistic status of specific systemicfunction attributions. Moreover, the theory of systemic functions, taken

166 Chapter 6

as a whole, is integral to the actual strategies with which we study hierar-chical systems. Hence the naturalistic credentials of the theory.

There are cases, of course, in which the discovery of lower-level physi-cal mechanisms does not provide a complete explanation of the higher-level capacity under study. Churchland and Sejnowski (1992), drawingon Selverston (1988), offer a nice example. Selverston studied thestomatogastric ganglion of the spiny lobster—a twenty-eight–neuroncircuit responsible for the muscles involved in grinding food. Despite hav-ing described the anatomical and electrical properties of each neuron,Selverston could not explain the rhythm of the circuit’s outputs. Thegrinding behavior of the lobster is explained by the fact that the neuronsfire in a given rhythm. What is not clear, however, is how the circuit pro-duces its own rhythm. Pointing to the causal powers of the neurons is notenough, for no one cell (and presumably no cluster of cells) produces therhythm. Churchland and Sejnowski thus suggest that, in addition to thepowers of the mechanisms, we also need to discover the ‘‘interactiveprinciples’’ that govern the system (Churchland and Sejnowski 1992, 5).

Now, I do not think we should dismiss the possibility of such principlesout of hand, but nor do I think that we should accept them on their face.Two points are relevant. First, we should expect that any such interactiveprinciples will be, upon further investigation, explicated in terms of somephysical mechanisms. The lobster’s pattern of grinding may not be expli-cable in terms of any one cluster of neuronal cells, but surely there existsan explanation—one that appeals to some physical properties of the or-ganism or the organism’s environment—that accounts for the pattern.Something in the constitution of the organism or its physical surround-ings accounts for the fact that it grinds in this pattern rather than someother. To this extent, appeals to interactive principles are appeals to igno-rance; they are principles waiting for a fuller mechanistic explanation;indeed, they are appeals to top-down systemic functions waiting a com-plete analysis. Second, we should, I suppose, leave open the possibilitythat mechanistic explanations are limited in scope and that some systemiccapacities cannot be adequately explained by appeal to physical mecha-nisms. This strains credulity, of course, given the stunning success ofmechanistic explanations to date. Nevertheless, naturalism, as I construeit, is a commitment not to a specific ontology but rather to the methods

Guiding Inquiry 167

of inquiry employed by our best natural sciences. So, if nonmechanisticexplanations are clearly successful in some domain, and if such explana-tions appeal to nonmechanistic principles of interaction, then so be it—we should embrace the reality of such principles. And in that event, thetheory of systemic functions would retain its naturalistic credentials byappealing to principles of interaction in addition to physical mechanisms.

*The inquiry-guiding role of systemic functions contrasts sharply with theexplanatory import assigned to selected functions. The central thesis inWright (1973) is that the attribution of an etiological function is equiva-lent to a historical explanation of the persistence of the functional type.As we saw in chapter 5, advocates of selected functions reject Wright’sbroad notion of selection and restrict their explanations to natural-selective explanations within evolutionary theory. And in light of this re-striction, the theory of selected functions is essentially explanatory andhistorical. By contrast, the theory of systemic functions allows that theattribution of functions is sometimes explanatory on historical grounds,other times explanatory on nonhistorical grounds (Harvey on the func-tion of the heart), and, most notably, sometimes not explanatory at all.As Enc (1979) and Bechtel and Richardson (1993) demonstrate, inquiryinto hierarchical systems that are not yet understood involves the attribu-tion of systemic functions, the role of which is nonexplanatory. Or, atminimum, explaining is not the immediate role of such attributions. Theirimmediate role is to provide a tentative taxonomy of systemic capacities,that we might formulate hypotheses concerning the sorts of mechanismsinstantiating those capacities. We want to locate and describe the causalmechanisms that form the system, that we may predict and explain howthe system does what it does. The appeal to systemic functions helps usexplain, to be sure, but only after the system is understood; before that,the appeal to systemic functions guides but does not yet explain.

Some may wish to challenge the distinction I have drawn between ex-plaining and guiding inquiry. After all, attributing systemic functions insystems we do not yet fully understand plays an explanatory role. It tellsus how the system might be working; it provides a how-possibly explana-tion of the specified systemic capacity. I agree that the attribution of sys-temic functions can be construed in this way. But the distinction between

168 Chapter 6

explaining and guiding inquiry is important nonetheless. It is true enoughthat, in attributing systemic functions to systems poorly understood, weventure a kind of how-possibly explanation.4 The relevant question, how-ever, is, What role do these sorts of function attributions play in ourinquiry? My answer is that that they guide inquiry into the unknowncomponents of the system. They provide a provisional mapping of thesystem that we then proceed to test. That is the payoff of such attribu-tions; that is why we make them. And these attributions do not becomegenuinely explanatory—they do not acquire the role of providing an ac-tual account of the way the system works—until we have discovered evi-dence that the attributed functions are instantiated by mechanisms in thesystem. Systemic functions in systems poorly understood are best con-strued not as explanations but rather as potential explanations, the imme-diate goal of which is to guide us into the system’s structures.

The requirement that we locate physical mechanisms responsible forinstantiating the systemic capacities itemized in our taxonomy ensuresthat the attribution of functions is well confirmed. When functions areattributed to systems well understood, such as the circulatory system, wehave extensive knowledge of the physical mechanisms that instantiate thelarger systemic capacities. Our function attributions, in this case, oughtto engender confidence. When attributed to systems poorly understood,our function attributions are tentative and hypothetical. Their existentialimport depends upon the degree to which available evidence confirmstheir existence in the system. In the early stages of inquiry into a givensystem, our attributions are highly tentative and our confidence mustbe tempered. We must await bottom-up confirmation before assertingthat the functions in our proposed taxonomy in fact exist. Our func-tion attributions, in this case, are hypotheses with evident heuristic

4. There are two importantly different kinds of how-possibly explanation—twokinds of ignorance that they are intended to alleviate. One kind is given whenwe are ignorant of the specific processes involved but not of the general structureof the larger system. Kingsolver and Koehl’s (1985) explanation of the evolutionof insect wings, discussed in chapter 3, is illustrative. The structure of wings andindeed the entire insect was known from fossils. But another kind of how-possiblyexplanation is given when we are ignorant of the overall structure of the phenom-ena we are trying to understand. Gall on the brain and Watson on RNA areillustrative. And here the guiding role of systemic functions is most pronounced.

Guiding Inquiry 169

value, but hypotheses to be tested all the same. Such attributions acquiretheir naturalistic credentials on the grounds that they require empiricalconfirmation.

II The Limits of Decomposition and Localization?

Despite my appeal to Bechtel and Richardson (1993), there is one aspectof their view I wish to question. Toward the end of their discussion theyclaim that there are certain systems to which decomposition and localiza-tion do not apply. Some systems exercise higher-level capacities by virtueof highly integrated interactions among lower-level capacities. A high de-gree of interaction means that the various lower-level components cannotproduce their salient causal effects independent of the effects of most orall other components. Significant integration among systemic compo-nents excludes significant independence of operations. The problem, ac-cording to Bechtel and Richardson, is that, as integration increases andindependence decreases, the system becomes increasingly less decompos-able. Decomposition assumes some minimal degree of causal indepen-dence among some or most components; we cannot assign specificsystemic functions to any component unless we can discover what itsunique causal contribution to the system is. But if the system is highlyintegrated, we are barred from any such discovery because we cannotisolate the efficacy of any one component. So, decomposition fails. Andif the system cannot be decomposed, localization cannot even begin. De-composition and localization, therefore, are limited to systems in whichthe degree of integration remains below a certain threshold. If, as I havesuggested, the theory of systemic functions substantially overlaps theseheuristics, then the applicability of systemic functions is similarly limited.It is this claim I wish to question.

Bechtel and Richardson describe two connectionist models. These, theysuggest, exemplify the sorts of integration that render decomposition inef-fectual. The first model is a simple two-layered network that, after fiftyepochs (training runs), acquires the capacity to recognize and classifycups, buckets, hats, and shoes. More modestly, the system acquires a ca-pacity that can be interpreted in semantic terms as involving the capacityfor recognition. The second model is a five-layered network, developed

170 Chapter 6

by Hinton (1986), that acquires the capacity to recognize and classify themembers of two family trees after 1500 epochs. The details of both casesare interesting but not essential to understanding the central claim. Theresult of training is that certain patterns of activation between variousprocessing units (nodes) are correlated with certain features of the ob-served objects or relations. After training, when certain features are pre-sented to the network, the associated activation pattern tends to fire. Itthus is plausible to interpret these activation patterns as having semanticcontent, as referring to or being about the correlated feature of the objectpresented.

The problem, however, is that the recognitional capacity of the net-work, though composed of patterns of activation distributed acrossseveral nodes, is not decomposable into meaningful subtasks. The lower-level interactions are far too integrated to be appropriately decomposed.Bechtel and Richardson conclude as follows:

It is not important that we take a stand on the ultimate viability of connectionismas a framework for cognitive theorizing in order to make our main point: connec-tionism represents a break with traditional mechanism, pointing toward a differ-ent category of models and employing an alternative strategy for developing them.This alternative emphasizes systems whose dynamic behavior corresponds to theactivity we want to explain, but in which the components of the system do notperform recognizable subtasks of the overall task. . . . The overall architectureof the system—and especially the way the components are connected—is whatexplains cognitive capacities, and not the specific tasks performed by the compo-nents. We have abandoned decomposition and localization. (Bechtel and Richard-son 1993, 222–23)

If we have abandoned these heuristics, we have likewise abandoned ananalysis of the system into systemic functional components. A few pagesearlier they summarize their claim in similar terms:

Network models do account for the cognitive performance, but often they do sowithout providing an explanation of component operations that is intelligible interms of the overall task being performed. The network is a cognitive system; thecomponents are not. The result is that we do not explain how the overall systemachieves it [sic] performance by decomposing the overall task into subtasks, orby localizing cognitive subtasks. (214)

I agree with most of the premises in these arguments, but not all, and Idisagree with the conclusions. As I read it, the above objection is, at bot-tom, a worry about the limits of our knowledge and perhaps our under-

Guiding Inquiry 171

standing of our own models. The worry is about the extent of our presentignorance. Worrisome as our ignorance may be, it does not, in my view,warrant the pessimistic claim that decomposition and localization cannotbe applied. At least three considerations are relevant.

(1) I agree with Bechtel and Richardson that the ‘‘overall architectureof the system—and especially the way the components are connected—is what explains cognitive capacities, not the specific tasks performed bythe components.’’ I agree, that is, that the connectionist model accountsfor the higher-level capacity (recognition) in terms of the structure of thenetwork—in terms of patterns of activation—and not in terms of specifictasks performed by specific lower-level devices. But we should not agreethat, in consequence, a systemic function analysis cannot be given. Forwe sometimes explain the instantiation of a higher-level capacity by ana-lyzing that capacity into the organized structures that constitute the sys-tem at its lower levels. The suggestion is that we conceptualizeconnectionist models, at least at the outset, as offering analyses of cogni-tive capacities in terms of structural—not interactive—systemic func-tions. Just as Haugeland (1978) analyzes the capacity of a fiber-opticcable to transmit images into the structural features of the cable itself—the size, location, and number of individual fibers—so too we can analyzethe capacity of a connectionist network to exercise some recognitionalcapacity in terms of various structural features of the network—patternsof activation and weighted connections. This seems to me an acceptable,even if simplistic, application of the theory of systemic functions.

On the account of decomposition offered by Bechtel and Richardson,higher-level systemic capacities are properly analyzed not into lower-levelstructures, but rather into lower-level tasks, and this seems to me an overlyrestrictive notion of decomposition. After all, the ultimate goal of decom-position is the localization of lower-level mechanisms responsible for thehigher-level capacity. And this is accomplished often enough by analyzingthe system into lower-level structures. We certainly accomplish this in ouranalysis of Haugeland’s cable. And, as Bechtel and Richardson concede,connectionist models do indeed explain cognitive capacities by appeal tothe ‘‘overall architecture of the system’’ rather than the specific tasks per-formed by systemic components. Why, then, insist that decompositionmust appeal to lower-level tasks and not to lower-level structures?

172 Chapter 6

Bechtel and Richardson hold that a system can be decomposed onlywhen there is some degree of causal independence among lower-levelcomponents, and that seems right. The question, then, is whether or notlower-level structures exhibit the relevant sort of causal independence;the question is whether or not we can, by experimental means, isolatethe causal contribution of a given structure to the more general systemiccapacity. It seems to me that we can in the case of Haugeland’s fiber-optic cable. We can measure quite precisely the quantity and wavelengthof light conducted by individual fibers. Likewise for the connectionistmodels described above. We can identify quite specific patterns of activa-tion that fire when the system is presented with appropriate input. Indeed,it is the correlation between specific activation patterns and the presenceof certain objects that underwrites our claim that the patterns have somesort of semantic value. We ought to insist, therefore, on a notion of de-composition that appeals to the organized causal effects of lower-levelstructures.

(2) There may be a quite different worry driving Bechtel and Richard-son’s claim concerning the limits of decomposition and localization. Onthe analysis given so far, we analyze the semantic capacity of a connec-tionist system into lower-level structural components, specifically, intolower-level patterns of activation. More to the point, we analyze a singlecognitive capacity into a single pattern of activation. This, however, mayjar our sensibilities. We may find it unsatisfying, indeed puzzling, to claimthat a given semantic capacity is explained simply by identifying the pat-tern of activation involved. Such an analysis moves immediately froma higher-level cognitive capacity (recognition) to a lower-level structure(pattern) that is difficult to interpret semantically. There is a sudden shiftfrom the cognitive to the noncognitive and perhaps this is the real diffi-culty in trying to decompose connectionist systems. Indeed, at times Bech-tel and Richardson seem to assert that connectionist systems are notdecomposable at all (full stop), but other times they suggest that thesesystems are not decomposable into component capacities that are ‘‘intelli-gible in terms of the overall task being performed.’’ The result, they say,‘‘is that we do not explain how the overall system achieves it [sic] perfor-mance by decomposing the overall task into subtasks, or by localizingcognitive subtasks’’ (1993, 214, my emphasis). This seems to suggest that

Guiding Inquiry 173

the real worry is the abrupt move from a higher-level cognitive capacityto lower-level capacities that are barely cognitive or completely noncogni-tive. It is as if we can see the structures involved in the production of thehigher-level capacity but have no intuitive understanding of how or whythis capacity emerges out of these structures. We have the feeling thatsomething important—namely, our understanding—has been left out. Atthe close of their chapter on highly integrated systems, and referring spe-cifically to connectionist networks, they say: ‘‘We may not be able tofollow the processes through the multitude of connections in a more com-plex system, or to see how they give rise to the behavior of the system.We may fail in the attempt to understand such systems in an intuitiveway’’ (229, my emphasis).

If this is the root of their complaint, then there is no especial objectionto the applicability of the theory of systemic functions or of decomposi-tion and localization, for the sense we have that something is left out isnot unique to connectionist models of cognition. The very same problemafflicts traditional computational models as well—though it tends to ariselater in the analysis of a system. Cummins (1983), for example, worriesover this (and a related) problem.5 Traditional computational models an-alyze complex cognitive capacities into less-complex cognitive capacitiesinteracting at some lower level of organization. This strategic step is iter-ated for each level at which some form of cognitive activity occurs. Fi-nally, however, our analysis of low-grade cognition takes us to the levelof capacities and components that are entirely mechanical. It is at thispoint that the above worry arises. Our analysis of some lower-level cogni-tive capacity in terms of purely mechanical components may give thesense that something is left out. We may have the feeling that we do notunderstand how or why these specific mechanisms, which by hypothesisare devoid of cognitive capacities, give rise to cognitive capacities at thenext level up.

Of course the gap between the cognitive and the noncognitive is greaterin connectionist models, since we move quickly from the fully cognitiveto the noncognitive or, at a minimum, to a level of nodes or sets of nodesthat are difficult to track in semantic terms. Perhaps this explains why

5. See Cummins (1983), chapter 2.

174 Chapter 6

some theorists find connectionist models ultimately unsatisfying—theyappear to leave out more than traditional computational models. But thecrucial point is that decomposition and localization apply without prob-lem to traditional computational models of cognition, as does the theoryof systemic functions. So the difficulty involved in bridging the cognitive/noncognitive gap provides no particular reason for thinking that theseheuristics or this theory cannot apply to connectionist models as well. Ifwe have the sense that rather little fruit is to be harvested from analyzingconnectionist systems in terms of systemic functions, we should suspectthat this is a deficiency in the connectionist network under consideration.Or perhaps we should suspect that our inquiry into the capacities andmechanisms at lower levels of neurological organization is too undevel-oped for us to assess the functional taxonomy that our connectionistmodels provide.

(3) Of course Bechtel and Richardson may be right that there are somehierarchical systems to which the theory of systemic functions does notapply. Perhaps there are systems constituted not out of layers of structuraland interactive mechanisms, but units of some other type. Nevertheless,I am not convinced that connectionist models are an instance of thisunfamiliar type. Inquiry into the neurological basis of our cognitivecapacities is still maturing and it is hard to predict what advances tech-nology will bring. We know too little about the neurological mechanismsthat underwrite, or may underwrite, our most accurate connectionistmodels. On my view, the analysis of a cognitive capacity into a patternof activation is merely the first step. It remains for us to investigate fur-ther the neurological mechanisms involved in the production of suchpatterns. And as I have been urging in this chapter, the attribution ofsystemic functions often occurs in trying to understand a system poorlyor partially understood. Such attributions guide more than they explain.We should understand connectionist models in this way. Our knowledgeof the implementation of connectionist models depends on progress inneuroscience, and while great strides have been made in recent years,much about the brain remains unknown. To the extent we can attributesystemic functions to connectionist components—even if only at thelevel of activation patterns—we do so as a guide to further systemic in-quiries, including inquiries into the functional capacities of the brain. So,

Guiding Inquiry 175

I agree that we should be cautious in the attributions we make, but Ido not agree that we know enough to conclude that the heuristics ofdecomposition and localization are as limited as Bechtel and Richardsonsuggest.

Finally, even if Bechtel and Richardson are right about the limits ofdecomposition and localization, systemic functions nevertheless are ofgreat utility in inquiry. And it is the depth and breadth of their inquiry-guiding roles that gives systemic functions both their scientific value andtheir sound naturalistic credentials.

III Systemic Malfunctions and Expectations

The attribution of systemic functions, then, is sometimes explanatory,while other times it serves to guide our inquiry into unknown phenom-ena. But this does not exhaust the theoretical roles of systemic functions.In attributing systemic functions, and in searching for the mechanismsthat instantiate these functions, we engage in a process of classifying sys-tems and their parts. We classify components according to the work theyperform. Enc (1979), after noting the classificatory power of functionattributions, argues that functional types impose limitations on the kindsof physical mechanisms in which such functions can be implemented.This is an important thesis—especially in philosophy of mind6—but Iwish to draw a distinct lesson concerning the classificatory power of sys-temic functions. My suggestion is that the classificatory practice involvedin applying the theory produces in us certain expectations concerningsystemic types and tokens of those types. Indeed, these expectations in-form our predictions and our explanations concerning the behavior oftoken traits. They thus facilitate our inquiries. But it is also these expecta-tions that explain why we are inclined to say of certain token traits—those unable to perform the associated functional task—that they aremalfunctioning. Or so I wish to suggest.

6. See Enc (1979), 349ff. Enc’s argument appeals to features of natural kinds.In order for his claim to apply to traits such as hearts and eyes, however, we needan account of natural kinds that appeals not to microstructural properties but tosomething else, for otherwise it is hard to see how it applies to lineages, popula-tions, species, or the like.

176 Chapter 6

My aim here is to explain why we are inclined to classify some tokensas malfunctional. The aim is to account for the presence and efficacy ofan inclination exercised in the course of inquiry. I am not interested intrying to explain how or why natural traits in fact malfunction, becauseon my view natural traits cannot malfunction. Of course, sometimes traitsare defective; more precisely, sometimes we are inclined to conceptualizea token trait as defective. We are so inclined to the extent the token failsto perform in ways we expect. But, of course, the failure to perform inways that satisfy our inclinations is not enough to show that the tokenis malfunctioning. And, as we have seen, systemic functions are identicalto systemic capacities, in which case the loss of the relevant capacity en-tails loss of the systemic function. So traits that fail to perform in waysthat match our expectations—expectations concerning the workings ofthe systemic type; more on this presently—are not malfunctional at all.They are merely nonfunctional. This is a consequence of the theory ofsystemic functions that cannot be dodged.

The basic point here is quite general: Any attempt to explicate functionsin terms of some type of success—success in contributing to higher-levelsystemic capacities, for example—cannot account for the occurrence ofmalfunctions. This point applies to the theory of selected functions insofaras it defines functional types in terms of success in selection. The problemis that, if functional types are defined by reference to some form of success,then incapacitated tokens that lack the success property do not belong tothe relevant functional types. Now, the aim of chapter 7 is to develop thisargument against the historical approach generally; I defend the thesis thatselected and strong etiological malfunctions are impossible. For presentpurposes, however, I assume that no theory of functions, including thetheory of systemic functions, can account for the occurrence of malfunc-tions by defining functional types in terms of some sort of success. Incapaci-tated tokens thereby lack the associated systemic function. What remains,then, is to explain why we are so inclined to see nonfunctional tokens asmalfunctional, why we insist on placing tokens that have lost the requisitephysical capacity into their former functional categories.

I speculate that we are inclined to say of a nonfunctioning token thatit is malfunctioning—that it should be performing such-and-such taskeven when it cannot—because we have acquired the expectation that

Guiding Inquiry 177

components situated in systems of this type perform the stated task. Ourexpectations are caused by our knowledge of the type of system, orknowledge of analogous systems, and the workings of their parts. Ourexpectations are caused, in short, by prior applications of the theory ofsystemic functions. More specifically, our knowledge of complex, hierar-chical systems causes us to expect the appearance of various componenttypes, as well as a range of regular systemic effects produced by tokensof those types—that is, effects that contribute to the exercise of somehigher-level capacity. We thus come to conceptualize components of thesystem in terms of systemic capacities that qualify as systemic functions.And when we happen upon a component token that fails to produce thesystemic effects we have come to expect, we are inclined to (mistakenly)place the token into the relevant systemic functional type.

Consider, for example, an infant born with only one eye, where theother orbit contains a gelatinous blob that lacks the capacity to processlight. We would be inclined to say of the blob that it is a malfunctioningeye or, at minimum, a malformed eye. After all, we know that humaninfants typically are born with an eye on either side of the face and weknow a good bit about the developmental processes involved. On thebasis of this knowledge, we expect the formation of two eyes endowedwith the capacity to process light. When things do not go as we expect—when, for example, we happen upon the infant’s blob—we are takenaback. We have the feeling that this token ought not be of this form; wehave the feeling that this token has failed in some respect—that it hasfailed to become what it is supposed to be or that it is failing to do whatit is supposed to do. But we are misled by these inclinations and, perhaps,by certain residual beliefs that, upon reflection, we would reject. The feel-ing we have that this token ought to be formed differently is an expressionof our expectations concerning token components within this type of sys-tem—expectations, for example, concerning the systemic capacities ofglobular capsules crucial to the general capacities of the visual system.Precisely these sorts of expectations, I submit, are at the heart of ourinclination to mistakenly see nonfunctional tokens as malfunctional.7

7. ‘‘Expectation’’ should not be construed narrowly. Some expectations dependupon knowledge of a given system acquired via repeated interactions with thesystem. But others may depend more on knowledge hardwired into our brains

178 Chapter 6

Crucially, then, when we assert of some natural trait that it is supposedto perform some task, we are not attributing to that object a norm ofperformance. We are simply expressing our expectations of the token inlight of our knowledge of the larger system. We thus are quite mistakento describe the infant’s blob as malfunctioning. We may also be wrongat a more basic level. We may be wrong even to classify the infant’s blobas an eye. That is, membership in the category of eyes plausibly requirespossession of the relevant systemic functions (as opposed to mere histori-cal relatedness) and hence possession of the requisite physical capacity.The infant’s blob, however, has no such capacity. So, it may be a mistaketo describe the blob as a malfunctioning eye for the more elementaryreason that it may be a mistake to place it in the category of eyes. Weshould not conclude, however, that there are no norms at play here. Thereare no normative properties ascribable to token traits, to be sure. Hence,there is no justified claim to the effect that this or that token occupiesan abstract, noncausal role. Nevertheless, our expectations underwritevarious epistemic norms. The basic standard is given by our knowledgeof the larger system and the expectations this knowledge generates. Weuse this standard to formulate predictions and explanations of systemicbehavior.

Consider, again, the speculations offered in chapters 4 and 5. Let usassume that natural phenomena are hierarchically organized and that, ashierarchical systems increase in complexity, the tendency to persist andperpetuate their kind increases. Add the further assumption that we arepredisposed to want to understand how hierarchical systems work, in-cluding systems that are complex and persistent, and that we are predis-posed to approach such systems by applying the theory of systemicfunctions. We are disposed, that is, to conceptualize systems in terms of

and less on experience of the system. Some expectations—those concerning facesand eyes, for example—probably include a deeply visceral component. A mal-formed eye grabs our attention and holds it, thanks to an innate capacity to payattention to anything resembling a human face. This is not true of our expecta-tions concerning natural systemic components generally. Surely some such expec-tations rest squarely on the results of inquiry into the relevant system. And, ofcourse, we harbor expectations of the components of artifactual systems. Thoseexpectations may be derived from experience of the system, but their main sourcesinclude various intentions and conventions.

Guiding Inquiry 179

higher-level capacities and the lower-level capacities in which they areimplemented. With these assumptions in place, it is plausible that inquiryinto complex hierarchical systems involves the categorizing of systemiccomponents by reference to systemic capacities, by reference to contribu-tions to higher-level capacities. And it is precisely here, in the process ofclassifying on the basis of systemic capacities, that we come to harborexpectations concerning the future form and effects of systemic compo-nents. These expectations form the basis for further inquiry into theseand analogous systems, and the basis for our attempts to control andmanipulate such systems for practical ends.

These expectations should be only as strong as our knowledge of thesystem warrants. In systems that we understand very little, we are pre-vented from seeing token components in light of informed expectations.We thus have little inclination to attribute malfunctions. In systems un-derstood in substantial ways, the inclination to attribute malfunctions isproportionately greater. Our expectations are most potent for systemsthe workings of which we know well. The infant’s visual system is a clearcase in point. Our knowledge of human anatomy, of the mechanisms ofinheritance, and the mechanisms of development all incline us to expectfairly specific systemic forms and capacities. But, of course, various eventscan result in relative formlessness or relative inefficacy, and here the feel-ing that something has gone wrong is the strongest. That is not surprising,given the extent of our knowledge of the visual system and our impressivesensitivity in perceiving faces.

As in chapter 5, I am in no position to defend the psychological conjec-tures to which I appeal; that is a job for psychologists. For purposes ofthis discussion, however, the essential point is the plausibility of theHumean strategy generally and the revision in our concept of functionsthat it entails. As suggested, there are two powerful reasons to foregothe attribution of malfunctions and consider instead the hypothesis thatincapacitated traits are nonfunctional rather than malfunctional. The firstreason rests on the results of chapter 5. The alleged naturalistic standingof so-called ‘‘proper’’ functions is difficult to substantiate. The secondreasons rests on the results of chapter 7. As I argue there, selected mal-functions are not possible. If we take both reasons to heart, we mustrevise our concept of functions, at least to some extent. Some theorists—Millikan (1989), Neander (1991), Preston (1998), etc.—insist that our

180 Chapter 6

concept of functions is a concept of properties that are inherently norma-tive. My view, by contrast, is that, although there are good historicalexplanations for why our concept includes a normative component, thereare no good philosophical grounds for retaining this component. On thecontrary, to the extent we are naturalistically inclined, there are simplebut sound philosophical arguments for rejecting any such component.Functions are systemic capacities, nothing more. Of course, there aremany different kinds of systems and many different kinds of systemiccapacities. The capacities of components within highly complex systemsexhibit substantial regularity and tend to recur over time; the capacitiesof components within relatively simple systems may not. But what marksthe difference between the complex and simple is not a difference in thekind of functional property at play, but simply a difference in the kindof system involved and, in some cases at least, a corresponding differencein the ways we are affected by the system.

The Humean strategy is also warranted on the grounds that it makessense from within the theory of systemic functions as a whole. As wehave seen, the expectations that underwrite our inclination to (mistak-enly) attribute malfunctions arise when we employ the theory of systemicfunctions. These expectations are the product of conceptualizing systemiccomponents in terms of how they work within the larger system; theyare the product of identifying from the top down and confirming fromthe bottom up the existence of systemic functional types within a system.In the ideal case, such expectations are the fruits of our theoretical labors.They also guide us in those labors. They enable us to anticipate, predict,control, and render intelligible the behavior of tokens of these types. Insome cases, however, these expectations mistakenly lead us to see non-functional tokens as possessed of properties they do not have. The mis-take comes in thinking that there are norms that attach to natural traits,when in fact the operative norms are nothing more than the expectationswith which we approach various natural systems. This way of under-standing our inclination to attribute malfunctions is recommended, then,by placing the practice of producing such attributions within applicationsof the theory of systemic functions.

Finally, this approach to the topic of malfunctions taps into a key in-sight of the view defended by Bigelow and Pargetter (1987)—the selective

Guiding Inquiry 181

propensity view.8 This view has been criticized in various ways, but thebasic insight that function attributions are in some way forward-lookingstill stands. The theory of systemic functions, as developed in this book,captures the forward-looking feature of functional properties and doesso in ways that make it superior to Bigelow and Pargetter’s own account.Two considerations are relevant.

First, recall that, on their view, traits possess functions insofar as theyhave the propensity to be selected for the performance of certain tasks.Functions are dispositional properties of traits, properties that would beselected for were the organism’s natural habitat to remain the same. Iaccept part of this view. I accept that functions are capacities belongingto systemic components and that these capacities need not be exercisedin order to qualify as functions. The capacities must in fact belong to thesystemic component; indeed, we must have evidence of physical mecha-nisms capable of instantiating the capacity within the system. But condi-tions external to the system may be such that the capacity is neverexercised, that the underlying mechanism is never activated, in which casethe functional capacity is idle but present nonetheless. But there is alsopart of this view I reject. Bigelow and Pargetter hold that functions arecapacities that will be selected for under relevant conditions; this is torestrict the realm of functions to the potential for success in selection.This I do not accept. On my view, functions are the capacities of compo-nents that give rise to the exercise of higher-level systemic capacities.Whether or not these capacities have the potential to be favored by selec-tion is quite beside the point—except in cases where the higher-level ca-pacity we wish to explain (a population’s capacity to evolve via selection,for example) is instantiated in mechanisms of selection. The internaleconomy of organisms can continue its various tasks—circulation, respi-ration, digestion, etc.—even while being selected against. Systemic func-tions persist and sometimes even flourish in the face of negative selectiveforces. And if the world were such that selection never occurred but therenevertheless existed hierarchical systems with structural and interactivecomponents operating at lower levels of organization, the componentsof these systems would possess a range of systemic functions. To fixateon the functions that emerge (or would emerge) in the process of selection

8. Explicated briefly in chapter 2, section III.

182 Chapter 6

is to fixate on one among several general kinds of systemic functions (seechapter 3 for further discussion).

The second consideration is that the forward-looking feature offunctions is more compelling on my view than on the view of Bige-low and Pargetter. On their view, functions are forward-looking justbecause they are properties that will be selected for if the environmentremains stable in salient ways. The forward-looking aspect of func-tions is a feature—a subjunctive property—of the functional disposi-tions themselves. The dispositions look forward in the sense that, if thehabitat of the organism does not change significantly, the dispositionswill be selectively successful in the future. But this, I believe, missesthe key insight, for it seems highly likely that there are dispositions thatmeet Bigelow and Pargetter’s subjunctive conditions, and hence qualifyas functions, that nevertheless are not forward-looking in their senseof the term. To see this, consider that, on their view, there is an indefi-nite number of functional properties that exist unmanifested. There aredispositions belonging to organisms that, because of some change inhabitat, have not been selected for. And it seems safe to assume thatthere also are dispositions that never will be selected for because of ir-reversible changes in habitat. We may not know which dispositions theseare, but it is plausible that numerous such capacities exist. And it ishard to see in what sense these unmanifested dispositions are forward-looking. If the probability of their ever being selected for is practicallyzero, then there is nothing of the future that they will ever see, inwhich case they are not forward-looking in Bigelow and Pargetter’ssense.

To capture the insight that functions are forward-looking, we shouldfocus not on the propensity for selection but rather on the practice ofattributing functions. The central thought is that functions are attributedwith an eye toward the future behavior of the functional trait. They areattributed, more specifically, on the basis of our informed expectationsderived from our study of the relevant type of system. They are attributedinsofar as application of the theory of systemic functions has generatedin us the appropriate expectations. So, the forward-looking aspect offunctions is not a feature of the systemic capacities, as Bigelow and Par-getter suggest, but rather an epistemic feature of the attribution of sys-temic functions. It is we—not the systemic capacities—who look forward

Guiding Inquiry 183

in the course of attributing systemic functions. We look to the futureinsofar as we expect future tokens of the systemic functional type to pos-sess and to exercise the relevant systemic capacity. And, while there isnothing intrinsically forward-looking about functions, the forward-looking aspect of function attributions is of great importance nonethe-less. It is important in scientific inquiry. It is part and parcel of theinductive inferences we make in the course of such attributions. It is alsowhy we are so inclined to (mistakenly) say of tokens that disappoint ourexpectations that they are malfunctioning.

Systemic functions thus play a variety of roles in the course of inquiry.When attributed to the components of poorly understood systems, ourattributions are tentative and hypothetical with little or no existentialcommitment. When attributed to systems that we understand well, ourattributions provide an account of the way the system works. This ac-count forms the basis for our explanations and predictions regarding suchsystems. Our attributions also express our informed expectations regard-ing the functional types and these expectations explain two further phe-nomena related to functions. First, they explain what we are assertingwhen we attribute malfunctions: We are asserting merely that the inca-pacitated token has failed to behave in accordance with our informedexpectations. Second, they explain the forward-looking feature of func-tion attributions: We are expressing the sorts of behavior we anticipatefrom or predict of future tokens of the functional types. In all of theseroles, our attributions are constrained by the bottom-up search for mech-anisms that instantiate the attributed systemic capacities. Hence the natu-ralistic credentials of systemic functions.

IV Malfunctions and Statistical Norms

Before closing my discussion of the naturalistic status of systemic func-tions, I want to point out that the theory of systemic functions is notvulnerable to a surprisingly common complaint. The complaint is thatthe theory appeals in an implausible way to statistical norms in order toaccount for the apparently normative status of functions. This worry ismisplaced in two ways. First, if I am right that there are no malfunctionsamong natural traits but only unsatisfied expectations concerning tokens

184 Chapter 6

of systemic functional types, then the theory of systemic functions is un-der no obligation to account for malfunctions. It may be required thatwe account for our inclination to attribute malfunctions, but we are notobligated to account for that which does not exist. Second, even if myrevisionist view of malfunctions is mistaken, the appeal to statisticalnorms is not required by the theory of systemic functions. Systemic func-tions are capacities that contribute to the exercise of some higher-levelsystemic capacity. There is no requirement that functional tokens contrib-ute successfully more often than they fail; the only requirement is thattheir contributions account in some systemic way for the exercise of therelevant higher-level capacity. I close this chapter with a brief consider-ation of this second point.

The worry about statistical norms is pressed most forcefully by Milli-kan (1989) and Neander (1991). Millikan points to examples intendedto show that statistical normalcy is irrelevant to the possession of func-tions. ‘‘It is quite possible,’’ she says, ‘‘that the typical token of a matingdisplay fails to attract a mate and that the typical distraction displayfails to distract a predator’’ (Millikan 1989, 295). It is assumed thatthe function of, say, a token mating display is to attract a mate. This isthe function of such displays even if the production of displays failsto attract mates more often than it succeeds. The conclusion is thatstatistical normalcy is not constitutive of functions. This is supposedto tell against the theory of systemic functions and in favor of the the-ory of selected functions, since selected functions are effects that wereselectively efficacious and selective success does not require a positivesuccess rate. It requires only that the trait’s effects are successful oftenenough. The odds of any human sperm cell fertilizing an ovum are low,but because fertilization happens often enough and because its contribu-tion to reproductive success is so great, sperm cells seem to have ob-vious selected functions but no systemic functions. Or so Millikanclaims.

I do not find this convincing. Before giving my main objection, I pauseto consider details of the case. We need not agree with Millikan that thefunction of mating displays is to attract mates in so simplistic a manner.For, it is reasonable to claim that the function of such displays is not toattract mates but rather to increase the probability of attracting mates.

Guiding Inquiry 185

Compare organisms that produce the display to those that do not; suchvariation must occur between organisms, or must have occurred at sometime, for otherwise selection could not sort between them. We will likelydiscover that those organisms producing the display tend to attract matesand reproduce more frequently than those that do not. What does thisshow about the function of such displays? We might take it to show thatmating displays have the function that Millikan attributes, namely, thefunction of attracting mates. But we need not take it that way. It mightshow merely that mating displays have the function of increasing theprobability of attracting mates. This leaves room for widespread failuresamong token displays. It also entails a claim about relative correlations:The production of mating displays is better correlated with reproductivesuccess than the failure to produce such displays. I see no obvious objec-tion to this sort of appeal to statistical norms.

There is a further response to Millikan’s charge. Millikan focuses upontoken mating displays but gives no argument for individuating behavioraltraits so finely. We can individuate behavior as Millikan does, but weneed not. Consider, for example, a case in which the females of somespecies go into season once a year. During this time they produce a spe-cific pheromone that causes the elevation of certain hormones within sex-ually mature males, resulting in the otherwise absent displays of dramaticcolors, or melodious songs, or offerings of food, or whatever. In this case,how should we count the mating displays of the males? One option is tochunk behavior by season, so that each mature male produces a singledisplay that lasts the entire length of the mating season. After all, if thehormonal rage persists for the entire period and then ceases until the nextseason begins, that provides some grounds for conceptualizing the male’smating behaviors as part of a single, sustained display. And if we acceptthis scheme of individuating displays, then the correlation between dis-play and mate attraction is high indeed. Another option is to count thedisplays of a male as a single display so long as they are directed towarda single proximate female. And, of course, other individuating schemesare available. The point is not to defend any one scheme; that should bedone on the basis of details of the organisms involved. The point ratheris that we should not blithely accept the fine-grained scheme that Millikantakes for granted.

186 Chapter 6

This is true even in the case of sperm cells. I see no obvious reason toaccept that individual cells are the relevant functional units from the pointof view of mating or from the point of view of fertilization. Single ejacu-lates are plausible functional units relative to the capacity to mate. Theindividual cells that comprise an ejaculate may well have a function at alower level of analysis, relative to the capacity of the ejaculate as a whole.But that is a different matter. The same point holds with respect to fertil-ization. The relevant functional unit is the aggregate, not the individualcells. Consider an analogy. When we inhale, alveoli in our lungs absorboxygen. They typically do not absorb all of the oxygen inhaled; some isexhaled, unused. We thus are not inclined to say of each oxygen moleculethat it has a systemic function involved in respiration. We do not individ-uate that finely. Why think otherwise about sperm cells and fertilization?It is more plausible that the function of sperm relative to fertilization,like the function of oxygen relative to respiration, is properly attributedto the aggregate. It is the aggregate that gets the job done. Of course,getting the job done involves systemic capacities of the various lower-level components. So, perhaps the function of any one sperm cell is tocontribute to the mobility of the aggregate, in which case the majority ofcells regularly fulfill their function. At any rate, and absent a compellingargument to the contrary, there are no good grounds for thinking thateach sperm cell has the function of fertilizing an ovum. We should notsuppose that sperm are functional in the way Millikan supposes.

In the end, though, none of these considerations bears on the theoryof systemic functions, for statistical norms are quite irrelevant to the attri-bution of systemic functions. What is required for possession of systemicfunctions is success in contributing to higher-level systemic capacities;but there is no requirement that the incidence of success outnumber theincidence of failure. To see this, consider the relevant system and systemiccapacity in the case of mating displays. I take it that the systems involvedare potential pairs of opposite-sexed organisms. I also take it that thehigher-level systemic capacity we wish to explain is the act of mating orperhaps the production of offspring. A systemic functional analysis ofthis system will identify traits of the male and female involved in theexercise of this higher-level capacity, including behavioral traits involvedin securing a mate. For concreteness, let us suppose that one such behav-

Guiding Inquiry 187

ior is the singing of a distinctive song. Our question, then, is this: In orderfor the singing of this song to qualify for a systemic function within thiskind of system, is it necessary that token singings succeed in attractingmates more often than they fail? The answer is No. All that is requiredis that, in the context of the specified system, the singing of this songcontributes to the exercise of the specified systemic capacity. And a typeof behavior can contribute to a higher-level capacity even though its to-kens fail more often than they succeed. It may take a dozen recitals tofinally arouse the interest of a female, but arousing her attention is pre-cisely the systemic contribution of such singing relative to the larger sys-temic capacity of mating. The singing of this song will appear in oursystemic functional analysis of the system; it will appear in our taxonomyof functions because it contributes in a significant way to the act of matingand the production of offspring. Systemic efficacy is the key to acquiringsystemic functions. Such efficacy does not require the statistical norms towhich Millikan objects.

Neander (1991) argues that statistical norms are irrelevant to functionsfor the simple reason that, if they were required, then an epidemic thatdestroyed all or most token capacities would result in the obliteration ofthe functional property. If a virus infected our eyeballs and rendered allof us blind, we would be forced to conclude that eyes are no longer forseeing. And that, according to Neander, is nonsense. Eyes would be forseeing even if every member of the population were so afflicted. An epi-demic would result in widespread malfunction, to be sure, but that is tosay that the functional standing of our eyes would not be lost. Neanderconcludes that an ahistorical theory of functions, like the one endorsedhere, cannot be correct.

Now, it must be conceded that Neander’s point has obvious intuitivepull. That, however, is hardly decisive. To see this, notice first that muchdepends on the details of the story. Suppose, for example, that the virusattacks and devours the neurological connections involved in vision with-out devouring the eyeballs themselves. In that case, the larger systemwithin which eyeballs once functioned is now destroyed. Our eyeballswould be ‘‘systemic components’’ in name only, without a system inwhich to operate. Like Aristotle’s severed hand, our eyeballs would havelost their functional status because they no longer contribute to the

188 Chapter 6

exercise of some higher-level capacity. Suppose further that the diseasealso attacks the genotype that codes for the visual system, with the effectthat our offspring develop ‘‘eyeballs’’ but fail to develop the neurologicalapparatus required for vision. In this case, phenotypic and genotypiccomponents of the visual system have been destroyed and hencefor-ward human eyeballs would have no systemic function. And yet, onNeander’s version of selected functions, these ‘‘eyeballs’’ would retaintheir functional status, at least for a few generations. For selected func-tions are constituted by selective history; current capacity is irrelevant.The eyeballs of our offspring, despite the absence of any visual system,nevertheless would be descended from selectively successful ancestral eye-balls. So, advocates of selected functions are stuck attributing functionsto eyeballs in organisms that lack crucial pieces of the genetic and neuro-logical equipment involved in vision.9 This is akin to insisting that Aris-totle’s severed hand would retain the function of enabling the organismto grasp and manipulate objects, even after the complete annihilation ofall human appendages. No doubt some of us, upon being afflicted withthis dreadful disease of the eyes, would point toward our faces and la-ment, ‘‘These things are supposed to enable us to see!’’ But that, at best,expresses an unfulfilled expectation concerning the status of our aban-doned ‘‘eyeballs.’’

Alternatively, suppose the disease does not destroy the larger systemof neurological connections nor the relevant genotype. Suppose it merelyinterferes in some way with the capacity of our eyeballs to process light.

9. Neander can appeal to the proposal of Griffiths (1993) that purports to showthat vestigial traits (like the human appendix) do not retain their functional status.Griffiths proposes that functional status is lost so long as enough time has passedto allow for the cleansing effects of regressive evolution. (See chapter 7 for fullerdiscussion of this suggestion.) But this will not save Neander’s view from embar-rassment. By hypothesis, the genotypes that code for the larger visual system havebeen destroyed by the virus, while the genotypes that code for eyeballs remainintact. Now, regressive evolution will lead to the loss or atrophy of eyeballs onlyif there is variation among the alleles that code for eyeballs. Such evolution willrequire two or more generations. So even if we endorse Griffiths’s proposal, wenevertheless must grant that our eyes (those afflicted with the virus) and thoseof our immediate offspring continue to possess the selected function of enablingus to see, despite the fact that the larger visual system has been destroyed. This,I claim, puts pressure enough on the theory of selected functions.

Guiding Inquiry 189

In that case, the diseased tokens have lost their capacity to contribute tothe larger systemic capacity of seeing. On the theory of systemic func-tions, therefore, these tokens are rendered nonfunctional. They do notmalfunction for they no longer belong to the relevant functional type.They are now devoid of their former functional status. This is true evenif the disabling effects of the disease are temporary. So long as the tokenlacks the capacity to fulfill the specified task, it is not a member of thefunctional type.

Here, however, is where Neander’s view exerts its force, for we mayfeel that our diseased eyes belong to their former functional type despitethe destruction of the underlying capacity. We may feel inclined to catego-rize these tokens as merely malfunctional as opposed to nonfunctional.This is even more intuitive if we suppose that the affliction is temporary,for then it appears that I am forced to the unintuitive view that the rele-vant tokens had a function, then lost it, then regained it.

Again, I concede the prima facie power of the intuition. Our folk bio-logical or physiological intuitions incline us to think that diseased eyesare malfunctional and not merely nonfunctional. But precisely here iswhere I believe that our concept of functions must be revised. As I arguedin chapter 5, the explanation of malfunctions offered by advocates ofselected functions commits them to a dilemma: Either they fail to accountfor malfunctions or they endorse an ontology at odds with our naturalis-tic orientation. This should make us question the intuitions that seem tofavor Neander’s position. I also suggest that certain of our psychologicalcapacities incline us to see some natural traits as ‘‘properly’’ functionalor genuinely ‘‘teleological’’ and others as merely useful. In particular, ourinclination to see some tokens as malfunctional can be explained by ap-peal to our informed expectations of systemic functional categories. ThisHumean strategy provides speculative grounds on which to explain awaythe intuitions to which Neander and others appeal.

V Conclusion

But there is another reason for rejecting Neander’s intuitions: The theoryof selected functions, despite its claims to the contrary, cannot accountfor the possibility of malfunctions. The theory endorsed by Neander and

190 Chapter 6

others cannot substantiate the central intuitions to which Neander andothers appeal. This is a point about the internal resources of the theoryof selected functions—there are no grounds within the theory to warrantthe attribution of malfunctions. I defend this claim in the next chapter.The general lesson is that the intuitions on which Neander and othersrely—intuitions concerning the norms of nature—have no standing. Weought to reject the lingering sense we have that natural traits have func-tions that are ‘‘proper’’ or genuinely ‘‘teleological.’’ We do better tochange the way we understand the nature of functions: Nonengineered,natural traits do not and cannot malfunction. As naturalists, we shouldnot pretend otherwise.

This excerpt is provided, in screen-viewable form, for personal use only by members of MIT CogNet. Unauthorized use or dissemination of this information is expressly forbidden. If you have any questions about this material, please contact cognetadmin@cognet.mit.edu