TY - JOUR
AU1 - Schroeter,, Laura
AB - 1. Introduction Ruth Garrett Millikan’s new book, Beyond Concepts,1 is a tour de force that integrates and refines the distinctive naturalistic accounts of mental representation, language and information that she has developed over the past 30-odd years. She has been a remarkably systematic theorist over this time, continuing to flesh out the naturalistic approach to the mind she first elaborated in her landmark book Language Thought and Other Biological Categories (1984). Over the years, her detailed arguments, changes in view and shifts in focus can make it hard to discern the forest for the trees. This book is the remedy. It can be read as a sort of prolegomenon to her other work, showing how her teleosemantics, her critique of traditional views of concept identity, her account of linguistic conventions and her approach to natural information all fit into a satisfying and mutually reinforcing whole. The primary virtue of the book is its compactness and accessibility. The book avoids engaging in detailed arguments in favour of laying out the overall picture. It keeps the broad sweep in view, often ending sections with pithy formulations of the gist of specific positions, and offers helpful references to earlier work where her views are elaborated and defended in more detail. But for all that, it’s not an easy book to digest, due to the sheer amount of philosophical terrain it covers, the neologisms marking new ways of thinking about familiar topics and the interlocking nature of the topics it deals with. Millikan’s stated aim in the book is to answer Kant’s question, ‘how is knowledge possible?’ from a thoroughly naturalistic perspective. More specifically, Millikan seeks to characterize the cognitive and worldly mechanisms involved in gathering and storing information about one’s environment through perception, testimony and induction. According to Millikan’s teleofunctional semantic framework, when these belief-forming mechanisms function in the ways that explain their survival and proliferation, they yield non-accidentally true beliefs about the world. Thus, empirical knowledge is a matter of one’s cognitive-cum-worldly mechanisms functioning ‘normally’ – that is in accordance with their own ‘proper’ (naturally selected-for) functions. The book is divided into two parts. The first part focuses on Millikan’s account of the cognitive and worldly mechanisms that ground empirical knowledge. The second part outlines a general account of signs and information transfer, which provides an integrated framework for characterizing natural signs (tree rings, storm clouds), simple intentional signs (warning calls, bee dances) and conventional signs (natural language, maps). These two parts are interdependent in complex ways. On the one hand, cognition involves both looking for natural signs in the environment and producing internal signs that are interpreted by cognitive subsystems. So cognition depends on signs. On the other hand, to be a natural sign, according to Millikan, just is to have the potential to play a certain role in cognition. So signs depend on cognition. This interdependence makes the sequential presentation of the view somewhat tricky, and it means that a second reading is helpful in trying to bring the whole picture together. 2. Unicepts and cognition The starting point for Millikan is a metaphysical picture of how the world is organized into ‘clumps’ of co-occurring properties. For instance, all cats tend to share a cluster of properties (body plan, diet, behaviour, developmental trajectory etc.) that distinguish them from canines, chrysanthemums or corkscrews. Each of these further kinds has its own largely disjoint cluster of characteristic properties – yielding distinct ‘clumps’ in property space. Such clumps count as ‘real kinds’ when some mechanism, such as homeostasis or reproduction, explains why those properties tend to stably occur together. Typically such stabilizing mechanisms do not ensure perfect reproduction of all properties in the cluster. So there won’t be strict necessary and sufficient conditions for falling within the same real kind (e.g. some cats are three-legged, some are vegetarian). Still, the basic metaphysics of disjoint clumps of non-accidentally correlated properties is what makes learning and induction possible. Whenever you encounter a curled up feline-shaped thing, there’s a very high probability that it’s also four-legged and carnivorous. This basic metaphysical story of discrete clumps in property space extends to other stable features of the world, such as individuals or affordances. An individual cat, Felix, is characterized by a cluster of more specific properties (colour, size, shape, space-time trajectory) that distinguishes it from other individuals of its kind. And affordance kinds, like food or mate or how to catch mice, are constituted by property clusters relevant to satisfying an animal’s goals. A crucial task for cognition, Millikan argues, is to detect and store information about the metaphysically real kinds that can ground induction. In Chapters 2–6, Millikan presents her account of the cognitive mechanisms that do this job: their structure, proper function, semantic interpretation and the mechanisms that test their reliability in tracking real kinds. The general shape of Millikan’s account will be familiar to readers of her earlier work. The main innovation is her introduction of ‘unicepts’ and ‘unitrackers’ in place of concepts. Unicepts are cognitive mechanisms that bind and store information as pertaining to a single feature of the world. Here is Millikan’s gloss for unicepts: Unicept: [a] neural node that helps in storing factual or procedural knowledge through its connection to other unicepts or with behavior controllers. (225) On this picture, unicepts are particular concrete mechanisms within an individual’s cognitive system. Your unicept ‘cats’, for instance, binds together all aspects of your understanding of cats: what they look like, what sounds they make, how they react to being washed, your fondness for cats, the fact that your mom is allergic to cats and so on. In addition to binding and storing all of this information as pertaining to the same topic, your unicept figures as a constituent of beliefs, perceptions, intentions and procedural know-how. Your belief that cats are carnivores and your ability to wrangle cats in the bath, for instance, both include your cat unicept as a proper part. (Millikan models standing beliefs as stable relations between unicepts, and procedural know-how as relations between unicepts and action dispositions).2 Each unicept is paired one-to-one with its own ‘unitracker’ – a sort of informational gatekeeper: Unitracker: [a] neural network whose function is to recognize information arriving at the sensory surfaces that concerns one particular thing and present it for use or storage by its proprietary unicept. (225) In the case of your unicept ‘cats’, your unitracker will be a complex of mechanisms that ground your ability to recognize a cat silhouette on a dark night, or a cat’s caterwaul, or cat memes on the internet, or a textbook’s lessons about cat evolution.3 In effect, the unitracker determines your dispositions to form beliefs on the basis of incoming information. Given epistemic holism, your unitracker must include virtually all of your standing beliefs and know-how about cats; but it probably excludes action-guiding attitudes, like your fondness of cats or your intention to adopt a cat. Unitrackers are constantly evolving, as new recognitional capacities are added and old ones revised or deleted. This unicept/unitracker framework represents an important simplification of the account of concepts in Millikan 2000. There, Millikan defined ‘empirical concepts’ as abilities to re-identify a substance in thought. Abilities, in turn, were explained in terms of proper, selected-for functions. This abstract account made it hard to see what exactly concepts are. Do concepts remain stable as one acquires new (selected-for) methods for recognizing a substance? And can different individuals share the same concept? On the new picture, unicepts are stable, recurring constituents of an individual’s thoughts while unitrackers are the evolving mechanisms for recognizing the referents of those constituents. Both are concrete cognitive mechanisms, which cannot be shared between individuals. Unicepts like some theorists call ‘mental files’ – cognitive binding mechanisms that are individuated independently of the information they bind.4 Millikan avoids the ‘file’ terminology because she thinks it misleadingly suggests binding mechanisms are like containers for their associated information: so your ‘Elizabeth II’ file would have to contain your ‘is a queen’ file, and vice versa. And she avoids the term ‘concept’, because it invokes a range of Fregean assumptions. It’s commonly assumed (i) that concepts can be shared by different thinkers, (ii) that concepts can’t refer directly to individuals, and (iii) that concepts have determinate individuation conditions that define an equivalence class (47–49). None of these claims holds true for unicepts. Because unicepts are internal binding devices, moreover, their intentional content can shift over time – unlike the content of a traditional concept.5 Chapter 3 sketches the cognitive mechanisms that allow unitrackers and unicepts to play different roles in cognition. First, Millikan posits primitive tracking mechanisms that are not tied to a unicept, and hence do not store information for later use. She calls these same-trackers (4.1–4.5). Same-trackers are mechanisms of sensory perception, which locate environmental features like colour, shape, distance and object-trajectories in an animal’s immediate environment. Same-trackers generate perceptual states with analogue, self-locating space-time structure that some theorists attribute to non-conceptual perceptual content (e.g. Peacocke 1992). Next, Millikan introduces three different types of unicept – affording, substantive and attributive – which are distinguished by the roles they play in judgement and the stable knowledge structures needed to play those roles (4.6–4.8). Typically, affording unicepts represent affordances – features of one’s environment that are relevant to one’s goals, such as food or slippery stuff. But what’s really distinctive of affording unicepts is that they can figure by themselves in occurrent judgements about features in one’s immediate environment: for example, ‘food again!’ or ‘slippery stuff again!’ or ‘Obama again!’.6 To play this role, affording unicepts must collect and store practical know-how: abilities to recognize and deal with particular features of the world on the basis of sensory inputs. Thus, the unitracker of an affording unicept must include egocentric same-tracking abilities. In contrast, substantive and attributive unicepts figure in multiple-unicept judgements with subject/predicate structure, which allows for negation: for example, ‘Cats are carnivorous’, or ‘Cats are not food’. Judgements with this compositional structure represent complete states of affairs and are thus suitable for long-term storage – unlike one-unicept affording judgements like ‘food again!’. To play their respective roles in these ‘factic’ judgements, substantive unicepts (which play the subject role) and attributive unicepts (which play the predicate role) require specific types of knowledge structures.7 It’s worth emphasizing that the distinction between affording, substantive and attributive unicepts does not demarcate mutually exclusive kinds of unicept. It’s possible for the very same unicept to play different roles in judgement: for example, your unicept ‘cats’ could figure in an affording judgement (‘a cat again!’), in a substance role (‘cats are carnivores’) or in an attributive role (‘Felix is a cat’). Nor do the different kinds of unicept correspond to any ontological distinction in what a unicept represents: a unicept targeting an affordance property like food, for instance, can play both the substantive and attributive roles in factic judgements (‘food is nourishing’, ‘cats are not food’). The only difference between the three types of unicept is that they call on different knowledge structures: understanding the referent via direct recognition, understanding the referent as a thing with attributes or understanding it as an attribute of things. However, these knowledge structures are all mutually compatible. Chapter 5, ‘How Unicepts Get Their Referents’, is the linchpin for the first part of the book. Here, Millikan explains how unicepts come to refer to stable features of the world and how they are tested for representational adequacy. This was the most difficult chapter of the book for me. On the surface, it simply articulates different mechanisms that help fix the reference of unicepts. According to Millikan’s teleosemantics, a unicept refers to that feature (if any) that its unitracker was naturally selected to track.8 So her account of reference-fixing hinges on an account of how unitrackers become tuned to tracking information about specific property clusters. But this chapter is also crucial to answering Millikan’s initial Kantian question, ‘How is empirical knowledge possible?’. On Millikan’s account, when a unitracker does what it was designed to do – by reidentifying information about the same stable feature of its environment – it will generate non-accidentally true beliefs about that feature. So Millikan’s explanation of how unitrackers acquire the function of tracking specific features doubles as an answer to how a unitracker generates empirical knowledge about them. Millikan proposes three types of cognitive mechanisms that can play this crucial role of tuning unitrackers to track something real (e.g. a ‘real kind’ – a clump of non-accidentally correlated properties). The first is priming. Priming involves a mechanism that’s designed to launch new unicepts on the basis of specific sensory cues. Innate face-recognition mechanisms, for instance, prime for new unicepts that can track individual human beings on the basis of their visual appearance. Priming mechanisms can be either innate or acquired: for example, biological theory can prime the formation of a new species unicept, on the basis of encounters with a single animal that exhibits novel biological features (75). The key point is that priming mechanisms themselves can have a prior function of enabling the tracking of things falling into a general category: their function is to generate unitrackers that use a general template to identify attribute clusters that tend to remain stable from one encounter to the next. The second type of targeting mechanism involves practical reward. Animals need to learn reliable ways to satisfy their goals – where the goals may range from an innate desire to suckle to an acquired desire to solve Rubik’s cube. Reward mechanisms help reinforce such learning, selecting for unicepts that target affordances – empirical features relevant to satisfying an animal’s goals. Through trial and error, the infant can learn to bind together different ways of recognizing milk-giving affordances (‘mama again!’), and the puzzle-solver can learn to recognize patterns that allow a solution strategy (‘now for the finishing manoeuvre!’). Millikan’s third targeting mechanism involves testing a unitracker for coherence among its different recognition criteria. The idea is that cognitive mechanisms select for coherence among these criteria by testing for consistency in the judgements they generate. When the same factual judgement, ‘x is ‘φ, is generated on the basis of different recognitional criteria, this convergence in judgement is evidence that the different criteria united by one’s unitrackers, ‘x’ and ‘φ’, are all successfully tracking a single objective feature of the world. For instance, the very same judgement, ‘cats are carnivores’, might be generated on the basis of perception of Felix eating tuna, via an indirect natural sign like a partially eaten bird surrounded by cat prints, or via the testimony of a biology textbook. Each of these ways of generating the judgement corresponds to a different information source exploited by your unitrackers ‘cat’ and ‘carnivore’. If all of your information sources support the same judgement, that’s evidence that your unitrackers are well-tuned to tracking a stable feature of the world via different sources of information. But suppose your unitrackers generate a conflicting judgement, ‘cats are herbivores’, on the basis of visually identifying a cat eating grass. According to your current template for animals, being a carnivore excludes being a herbivore. So you now have two inconsistent judgements: ‘cats are carnivores’ and ‘cats are herbivores’. This inconsistency is an indication that something in your unicept system needs revision: perhaps that wasn’t really a cat you saw, perhaps that cat wasn’t really eating grass, or perhaps being a carnivore isn’t a projectable attribute of cats after all. Each of these verdicts involves revising your unitrackers or your templates of projectable characteristics of the referents. The moral Millikan draws is that cognitive mechanisms that select for consistency in judgement tend to favour the survival of unitrackers whose sources of information all manage to track a single real feature of the world. Over time, these mechanisms function to target unitrackers more tightly on specific clumps of non-accidental property correlations – which is highly beneficial to the organism as a whole. In effect, then, cognitive mechanisms select for internal consistency in judgement because this helps to optimize the representational adequacy of unicepts. Millikan rightly stresses the importance of consistency for her framework. By itself, priming cannot explain how existing unitrackers and templates can be revised through empirical feedback; and reward only applies to the acquisition and fine-tuning of unitrackers for affordances. But there must be some mechanism that explains how it’s possible to originally acquire unitrackers for new kinds of thing – and to revise one’s unitrackers for old things. Testing for consistency in judgement is thus the key to Millikan’s naturalistic epistemology – it explains how organisms can learn to recognize new objective features of the world. The process, she suggests, is like hypothesis testing in science. In bundling together information on the basis of a variety of criteria, a new unitracker is tantamount to a hypothesis that a single non-accidental property cluster in the world is tracked by all of those criteria. Testing the resulting judgements for consistency provides confirmation or disconfirmation of that hypothesis.9 Testing for consistency in judgement is, on Millikan’s account, first and foremost a matter of testing for the coherence (and hence the representational adequacy) of the specific unicepts involved in a judgement. Millikan’s naturalistic epistemology is thus very different from Quine’s (83). Although they both take coherence in the representational system as the ultimate test for its justification, Quine’s notion of coherence is radically holistic – it concerns the adequacy of a complete belief system considered as a whole. Millikan’s notion of coherence, in contrast, is targeted at testing specific unitrackers involved in specific judgements. So the question of what to revise in the face of negative feedback is more tractable: something has gone wrong with the way these specific unitrackers generate these specific inconsistent judgements. In sum, Chapter 5 plays a pivotal role in Millikan’s overall theory of reference determination, not just for unicepts, but as we will see for every other type of sign. And this chapter lays the foundations for Millikan’s naturalistic epistemology.10 An important corollary of Millikan’s account is that there are a number of different ways that unitrackers may fail to perform the tracking functions for which they were selected. Chapter 6 is devoted to explaining the different types of misrepresentation that are possible as the result of such failures. A false belief requires a univocal, non-empty unicept: although the belief-forming mechanism succeeds in representing a determinate state of affairs, it fails to track the actual facts. However, unicepts may fail in this prior representational task by being empty, redundant or equivocal. An empty unicept fails to track anything (e.g. unicepts of Santa or Satan). Redundant unicepts treat a single thing as two (e.g. unicepts of Hesperus and Phosphorus). And equivocal unicepts treat two distinct things as if they were the same. The last case is the most interesting and challenging. Millikan highlights two kinds of equivocation. The first involves distinct levels of tracking function. New-born goslings are primed to treat the first moving thing they encounter as their mother – but unlucky goslings can latch onto something else altogether, a pair of boots, treating the boots as their mother. In that case, Millikan says, the unlucky gosling has an equivocal unicept that targets both those boots and mother, treating them as the same. A different type of equivocation involves merging information from two external sources: for example, you may have a single unicept that conflates Tweedledum and Tweedledee. Again, Millikan says your unicept targets both individuals (91). One might object to this interpretation. If the function of unicepts is to identify a single feature, then shouldn’t we say unicepts are empty if they don’t succeed in this function? What’s the difference, then, between equivocal unicepts and empty unicepts? What fixes the precise content of a unicept in hard cases? And how should we evaluate beliefs that deploy these unicepts? Millikan is unfazed by this sort of challenge. All empirical distinctions, she thinks, are fuzzy and a matter of degree. It’s a matter of degree how well a unitracker’s recognitional criteria converge on a single clump in property space, and it’s a matter of degree how well a region in property space is divided into clearly isolated property clumps. In addition, there are different layers of proper functions governing a unitracker, each of which could potentially track different targets if internal and external mechanisms fail to operate as they were selected to do. So how well a unicept succeeds in keeping track of a single feature of the world will be a matter of degree, with plenty of vagueness at the margins. 3. Signs and information transfer Let’s now turn to the general theory of signs and information that Millikan presents in the second half of the book. Her project here is to explain what signs are, how they are organized into systems, and the role they play in helping cognizers glean information about the natural world and communicate that information to others. A helpful way to get a feel of Millikan’s overall picture is to contrast it with Paul Grice’s account of natural and non-natural meaning (Grice 1957). Millikan’s notion of a natural informational sign (an ‘infosign’) plays roughly the same role as Grice’s notion of natural meaning. To a first approximation, an infosign is a carrier of veridical evidence about some distinct state of affairs. Those tracks in the snow, for instance, are an infosign of a dog having taken this path earlier; and that smoke on the horizon is an infosign of a distant forest fire. The information-carrying relation between sign and signified requires a non-accidental correlation between two types of states: for example, forest fires cause smoke. This correlation can be stronger or weaker, depending on how probable it that a forest fire when there’s smoke. But the content of an infosign does not vary with the strength of this correlation: smoke can be an infosign of fire even if the conditional probability that smoke is accompanied by fire is very low. It’s worth stressing that infosigns, like Gricean natural meaning, cannot be false: that smoke on the horizon is an infosign of a forest fire only if whatever explains the correlation between smoke and fire is actually operating in this particular case. An infosign thus requires the actual manifestation of the mechanism that grounds the non-accidental correlation linking two types of states of affairs. What determines an infosign’s representational content? In Chapter 11, Millikan argues that correlations alone can’t settle this question. First, such correlations come in varying strengths and there is no non-arbitrary cut-off point for the conditional probability required for conveying informational content. Second, and more importantly, whether there is a correlation at all depends on assumptions about a relevant comparison class. Maybe in the past smoke on the horizon was reliably correlated with forest fires, but now the forests have all been logged and smoke is only correlated with dust storms. In that case, is there a correlation between smoke and fire? It depends on whether the relevant comparison class is past or present conditions. Mere correlations are too indeterminate to settle what information an infosign carries – what it signifies. Instead, Millikan suggests that infosigns are like affordances – they’re food for cognition (145). On this approach, an event counts as an infosign only insofar as an animal can use it to glean information about something else. This means that whether one thing is an infosign of another is determined only relative to the cognitive capacities and interests of a particular animal. That smoke on the horizon is a sign of a forest fire for a hawk only if (i) a hawk is capable of noticing the smoke/fire correlation and (ii) it is capable of using information about smoke as a way to track information about fire. Millikan argues that this approach helps provide non-arbitrary answers to the two challenges facing a pure correlation account of the information carried by signs. First, there is no single probability cut-off point that fixes the content of infosigns. The correlation between sign and signified must only be strong enough for natural selection to favour the animal’s use of the sign as evidence of the signified. If a hawk’s interest in locating fires is important enough to it, natural selection may favour its learning to use smoke as an indicator of fire even when the correlation is very weak. Second, relativizing the sign/signified relation to a particular animal helps solve the reference class problem. Smoke could be an infosign of dust storms for a hawk if, in its current environment with its current goals and sensory capacities, the hawk could acquire the ability to use smoke as an indicator of dust storms. On Millikan’s approach, then, the infosign/signified relation is always relativized to: (i) an animal and the correlations that animal can become sensitive to and (ii) a reference class that the animal could learn to distinguish. This relational approach means the cognitive capacities of animals are the key to making something an infosign. There are infosigns only if natural selection could reinforce an animal’s use of the sign to track the signified. Millikan’s notion of an intentional sign is analogous to Grice’s non-natural meaning (Ch. 12). Roughly, an intentional sign (or ‘sent sign’) is a state whose proper function is to communicate veridical information about some distal state of affairs from a sender to a receiver. For instance, bee dances, beaver warning splashes and linguistic utterances are all intentional signs that are produced by senders and interpreted by receivers. Senders, signs and receivers together form a communicative system that is favoured and preserved by natural selection insofar as the signs are sufficiently reliable in conveying facts from sender to receiver. When the system operates in the way it was naturally selected for, the intentional sign will function as an infosign for the receiver.11 For instance, this bee’s dance will function as a natural infosign of the location of nectar for the other members of the hive, and that beaver splash will function to inform other beavers of the location of a predator. But of course no system always functions in the way it was designed: illness may lead a bee to produce a misshapen dance, for instance, or the beaver may splash in reaction to the sudden movement of a harmless deer. In such cases, the sign malfunctions: it fails to carry the kind of information it is designed to convey. Thus, unlike natural infosigns, intentional signs can be false.12 Like Grice’s non-natural meanings, Millikan’s intentional signs are essentially tied to communication. But unlike Grice, Millikan does not invoke a speaker’s communicative intentions to explain how the content of her signs is fixed. Instead, the content of an intentional sign is determined by the processes of natural selection that operated on the evolution of a whole communicative system (sender, sign and receiver). The proper function of a sign sent between two animals is to convey an adaptive state from sender to receiver: the function of descriptive language, for instance, is to supply natural information about some distal state of affairs to hearers (173). Thus the content of a linguistic signs seems to depend on the content of the internal states – in this case uniceptually articulated beliefs – involved in properly sending and receiving the sign.13 The fact that intentional signs are designed to function as infosigns means that all signs – both intentional and informational – have the same basic structure. This observation has far-reaching consequences for Millikan’s account of the semantics and pragmatics of natural language. It also means that, on Millikan’s account, the cognitive processes involved in interpreting natural and intentional signs is the same. Let’s start with the structure of signs. The first point Millikan emphasizes is that, at the most basic level, both signs and signified are complete states of affairs. If a tin can bears the label ‘spinach’, the state of affairs signified is that this particular can contains spinach. But what is the sign that signifies this state of affairs? Clearly the word ‘spinach’ alone is not doing all the work. By itself, the word just gets us to think of spinach, without specifying the contents of this can. In order to understand the proposition being communicated, you need to include both the can and how the label was related to it in the sign system. So the intentional sign in this case is the complete state of affairs: that ‘spinach’ was written on this can. In this communicative system, both the tin can and the word function as a signs; but while the word ‘spinach’ represents spinach, the can represents itself – it is a ‘self-sign’. Beaver splashes and bee dances also involve self-signs: the time and location in which the actions are performed function as self-signs determining the time and location of the signified states of affairs (65). Similarly, an affording judgement of the form, ‘slippery again!’ involves self-signing: the time and place of the judgement represent the time and place of the slipperiness. Thus an intentional sign system can include more elements than those actually produced by a sender: features of the context in which a sent sign occurs may need to be included as proper parts of a complete sign if we are to explain how the sign system functions in communication. Millikan uses the notion of self-signs to develop an account of the function of indexicals, demonstratives and anaphora in natural language: these linguistic devices recruit elements of the context in which they are used to stand for themselves (Ch. 9).14 Millikan’s second point about the structure of signs is that all signs are part of combinatorial systems (Ch. 10). That is, all signs are articulated into proper parts that can be recombined in systematic ways to signify a range of different states of affairs. Self-signs illustrate how this can be true even for apparently monadic signs like beaver splashes: the complete sign is a splash occurring at a particular time and place, and the state of affairs signified is a predator being near that time and place. In this simple sign system, distinct states of affairs (predators lurking at various locations and various times) can be represented by varying the time and place of splashes. More complex systems (like bee dances, photographs, maps, factic beliefs or sentences) have many more basic elements, which expand the range of states of affairs that can be signified exponentially. Each sign system will have its own combinatorial structure, which determines how the different elements can be recombined to form new patterns. Millikan’s account of intentional signs and infosigns grounds a distinctive approach to the semantics and pragmatics of natural language. A natural language is just a very complex system of intentional signs. The semantic content of a linguistic sign is determined by the content that that sign has been naturally selected to communicate, given the combinatorial structure of the sign system (the language). But a sent sign can also be a natural infosign of things it wasn’t selected to communicate: The location of a true beaver danger splash intentionally signifies the location of imminent danger to beavers, but it also non-intentionally infosigns the location of beavers, of water, and usually of nearby felled trees and a lodge. … Any normally true token of ‘It’s raining’ intentionally infosigns that it is raining where it’s said. It also signs nonintentionally that it will soon be wet outside, that the humidity is high, that the speaker knows English, that the speaker believes it is raining, and, in the right context, it may infosign that this afternoon’s little league game will be canceled, that the grass will finally recover, or that it will now be cooler outside and likely slippery on the steps. (162) Thus, Millikan’s distinction between intentional signs and infosigns allows her to distinguish an intentional sign’s semantic content from what it merely pragmatically conveys to a particular receiver. Moreover, Millikan rejects the idea that speakers’ communicative intentions play any role in determining linguistic truth conditions. So her teleosemantics by itself must be able to generate truth conditions for descriptive utterances that involve ambiguous or context-dependent expressions (Ch. 15). For instance, Millikan handles the use of ambiguous expression types like ‘bat’ by tracing the lineage from which a particular token utterance derives: one lineage was proliferated for tracking an animal kind, another was proliferated to track an artefact kind (174). At the same time, Millikan stresses that semantic truth conditions are ‘skeletal’. To understand the semantic content of ‘horses have hooves’ both speaker and hearer need to understand ‘horse’ and ‘hooves’ via unicepts that have been tuned to the same real property clusters. In short, semantic content is directly referential (Ch. 2). But what individual speakers and hearers actually understand by ‘horse’ includes all of the different assumptions linked to their respective unicepts: for example, that horses are heavy, that they can be ridden, that only rich people can afford them and so on (176). Since everyone’s set of assumptions will be slightly different, this information cannot be part of the semantic content of ‘horse’. Yet often interpersonal communication turns on co-ordinating this extra-semantic content between speaker and hearer. How is this achieved? Successful communication happens when the hearer recovers at least as much content as the speaker intended to communicate, and does not recover any falsehoods that are not part of the semantic content (174). Beyond the directly referential semantic content, the hearer is free to include any extra-semantic information that can be gleaned from the current context and her stored information. In communication, then, the speaker constructs a linguistic sign that signifies (for him, in this context, intentionally or naturally) the facts he wants to convey. And the hearer then derives the content that the sign signifies for her, in this context, intentionally or naturally. Co-ordination between the speaker’s and hearer’s extra-semantic content is facilitated by the fact that humans engage in joint attention to features of the world, and engage in joint practical and theoretical projects, which facilitates convergence on similar extra-semantic contents in similar contexts (176–8). In contrast to Grice’s account of conversational implicature, the hearer need not know anything about the speaker’s communicative intentions. One final point Millikan makes is that the compositional structure of a linguistic sign may itself be indeterminate. In shared contexts, compositional phrases can gradually fuse into a new unitary expression with a new idiomatic content. For instance, as a popular metaphor like ‘dyed in the wool’ gradually forges its own reproductive lineage, the phrase becomes fused into a unitary sign with a new semantic content that differs markedly from the original compositional content. Similarly, common implicatures that originally relied on context to get the message across – for example, ‘go to the bathroom’ for ‘defecate’, ‘and’ for ‘and then’ – can gradually form their own reproductive lineages with their own idiomatic semantic contents. Languages are constantly evolving in this way, and there are many different overlapping traditions operating on the linguistic forms of a language at any one time. Millikan’s suggestion is that it may be indeterminate which lineage a given utterance belongs to: you may hear ‘dyed in the wool’ as a compositional phrase whereas I may hear it as a single fused idiom, and it may be indeterminate which lineage is influencing the speaker herself. A network of overlapping reproductive lineages for linguistic signs, Millikan concludes, ‘may render the semantics/pragmatics distinction indefinite, in various degrees, over a significant portion of any language’ (182). The observation that linguistic signs are simultaneously sent signs (with semantic content) and natural signs (with pragmatic content) also plays a role in Millikan’s account of how we interpret a variety of expressions, like incomplete definite descriptions, possessives and domain restriction (Ch. 16). If I say, ‘Jane’s map led everyone to the park’, you’ll normally know who and what I’m talking about. How is this done if not by tracking speaker’s communicative intentions? Millikan’s answer is that expressions can have a proper semantic function of cuing a hearer to identify a referent on the basis of any natural information available to her. What hearers understand is thus a mixture of semantic and pragmatic content.15 4. Epistemology and reading signs One important consequence of Millikan’s general theory of signs is that understanding language is treated as a kind of perception (Ch. 14). Both sensory perception and linguistic understanding are essentially sign-reading skills, in which an animal learns to recognize external signs as carrying information about a distinct state of affairs. Perception, I take it, is a form of sign reading. It attempts translation of the natural informational content carried by patterns in sensory data into inner intentional signs, either beliefs or representations of affordances. (185) Millikan emphasizes that perception does not involve anything like an inferential reasoning from a unicept about your own experiences to a unicept about an objective state of affairs. When you see a lemon, the only unicept involved is your ‘lemon’ unicept. Your unitracker simply funnels information derived from sensory same-tracking abilities directly to your stable ‘lemon’ unicept, without relying on any unicepts that represent your sensations as such. This is what Millikan means by saying you directly ‘translate’ the natural signs of shape, colour, texture etc. into an inner judgement (which has uniceptual structure). Perception does not involve a uniceptually structured inference. Sensory perception is a skill that uses a variety of sensory inputs (infosigns) to track facts about stable features of the world. According to Millikan, understanding language is similar: it is a skill that uses various linguistic inputs (infosigns) to track facts about stable features of the world. When someone says ‘lemons are on sale today’, for instance, you’re directly disposed to form a belief with your corresponding unicepts (§7.5). You won’t normally form any prior uniceptual representation of the sentence’s surface form or the speaker’s communicative intentions. So beliefs based on testimony are not normally based on an inference. Instead, your unitrackers for features of the world (like lemons or being on sale) have been tuned to exploit linguistic infosigns – actual manifestations of non-accidental correlations between linguistic facts and facts about the target. Similarly, you become sensitive to speakers’ intentions to deceive, perhaps even forming an affording unicept targeted on deceptive intentions, without forming any factic beliefs about them – much less basing your reliance on testimony on prior beliefs about a speaker’s honesty or reliability (cf. §7.6). Millikan’s naturalistic epistemology thus places a great deal of importance on recognitional skills, as opposed to explicit inference. The ability to recognize non-accidental property clusters is the primary goal of learning. And this requires tuning one’s unitrackers to use easily available infosigns – manifestations of non-accidental correlations – to glean information about stable features of the environment relevant to one’s interests. A non-reflective creature like a frog relies on innate tunings, which were selected for over generations: sudden noises and looming shadows may be automatically channelled to the animal’s intentional sign for danger which cues a freeze or flee response. But more reflective creatures can fine-tune their unicepts through feedback from goals or feedback from consistency judgements (Ch. 5). These critical feedback mechanisms allow a creature to become more discriminating in identifying genuine (veridical) infosigns of the features they target (e.g. not every shadow indicates danger, not every assertion using ‘lemon’ is true). And consistency mechanisms provide needed empirical feedback to tune unicepts to entirely novel property clusters (e.g. these behaviours are an indication of schizophrenia). Stepping back, then, Millikan provides an integrated picture of how unicepts are designed to track information about the natural world. Unitrackers are honed by natural selection to use readily accessible infosigns to collect information about non-accidental property clusters. Many of these clusters will be affordances, features directly relevant to fulfilling an animal’s goals and interests. But many human unicepts are targeted on non-anthropocentric property clusters – real objects, kinds and properties that may (or may not) turn out to be relevant to fulfilling new and unforeseen goals. Unlike unicepts for affordances, unicepts for ‘real’ features are selected via mechanisms that test for consistency between judgements. In effect, then, these unicepts are tuned via purely intellectual feedback mechanisms, driven by intellectual curiosity. Thus the selection mechanisms governing factual unicepts are largely autonomous from direct selectional pressures to track features that promote specific goals. This, in effect, is Millikan’s answer to critics like Tyler Burge (Burge 2010), who doubt that teleosemantics can explain the ability to represent features of the world that are not directly relevant to an animal’s survival. This may be true of most animals, Millikan concedes, but it is not true of humans (4–5, 77–79). When our unitrackers are functioning as they were designed, both sensory perception and testimony afford non-inferential (‘perceptual’) knowledge of facts about the world. The resulting beliefs are directly referential – they represent real objects, kinds and properties without relying on reference-fixing criteria. And they are non-accidentally true. Presumably these unitracker-mediated beliefs constitute Millikan’s answer to her Kantian question, how is knowledge possible. Finer-grained epistemological evaluations can be made by citing the strength of the correlation between sign and signified. In effect, then, Millikan’s account of perceptual knowledge is a naturalistic version of virtue epistemology: when a unitracker is functioning normally, it generates knowledge. 5. Conclusion In this overview, I hope to have given some sense of the depth and originality of the account of mind, language and epistemology Millikan presents in this book. Of course, there is plenty of scope to raise worries about the details. But the overall picture Millikan paints of how our internal cognitive systems are structured and how they have been tuned by natural selection to exploit non-accidental correlations in the world is the most sophisticated and well-articulated naturalistic account of representation on offer. The book is an impressive intellectual achievement and it will be essential reading for any philosopher working on mind, language and information. The clear and concise writing make it an engaging read, but it will repay careful re-reading to appreciate the connections between the different chapters. Footnotes 1 Beyond Concepts: Unicepts, Language and Natural Information. By Ruth Garrett Millikan. Oxford University Press, 2017. viii + 240 pp. 2 Why ‘unicepts’? Millikan explains the terminology: ‘“Uni” is for one and “cept” is from capere, to take or to hold. The unicept takes various items of information about one and the same thing and holds them together so they can be used together’ (44). 3 On Millikan’s view, information about a referent that is derived from testimony or from pictures on the internet plays the same role in cognition as information derived from direct visual or auditory tracking. In fact, she takes reliance on testimony to be a special type of indirect perception. An explanation of this approach, however, depends on her general account of information and signs elaborated in the second part of the book. 4 For instance, John Perry and Laura Schroeter use ‘mental file’ in this way (Perry 2001, Schroeter 2008). However, other theorists define ‘mental files’ as the totality of bound information (Evans 1982, Lawlor 2001), or take ‘files’ to be individuated in part by a proper subset of the bound cognitive states or relations (Recanati 2012, Dickie 2015). Most mental file theorists restrict them to singular thought – a restriction that does not hold for unicepts. 5 An intentional shift can happen, for instance, if a unicept’s associated unitracker is selected for tracking something new – which will result in a unicept’s content becoming equivocal. 6 In talking about affording judgements, I’m departing from Millikan’s preferred terminology. She reserves the term ‘judgement’ for mental acts that produce representations with subject/predicate structure. Instead, she speaks of the recognition of affordances – where this is not taken to imply that the affordance actually exists. I’m following the meta-ethical literature, where ‘judgement’ is a neutral term for an occurrent attitude. 7 To play the subject role in judgement, each substantive unicept for a real kind must be associated with a cluster of characteristic projectable attributes of its referent (e.g. cats have a characteristic shape, habits, diet etc.). To play the predicative role, an attributive unicept is associated with (i) a range of substantive unicepts it can modify (e.g. ‘being carnivorous’ can modify ‘cats’ and other animal unicepts, but cannot modify ‘rocks’) and (ii) a set of attribute unicepts it is taken to exclude (e.g. ‘being carnivorous’ excludes ‘being herbivorous’). This exclusion relation is crucial to Millikan’s account of negation (67–68). 8 It’s worth stressing, that on Millikan’s account, the mechanisms that reinforce individual and group learning are also natural selection mechanisms. 9 As in traditional philosophy of science, her processes of hypothesis formation in ‘the context of discovery’ are sharply distinguished from the processes of hypothesis testing in ‘the context of justification’. Discovering a good hypothesis is a matter of trial and error (80–81). 10 Millikan addresses one further issue in this chapter, the problem of location-detached signs (5.7). Roughly, a location-detached sign is a sign that does not ‘show’ its spatio-temporal relation to what it signifies. Here is her summary of the problem: How do we locate things to same-track and then learn about them without first locating and following them in space-time relation to ourselves (79)? Millikan’s broad answer to this question is that we can form unitrackers that rely on multiple sources of indirect evidence (‘proxy signs’) and we then refine those unitrackers by means of consistency mechanisms. This process, like scientific theorizing, can allow us to home in on real kinds without first tracking them in our immediate environment. 11 That is, a properly functioning intentional sign will manifest a non-accidental correlation between sign and signified (established by natural selection), which is used by both sender and receiver to keep track of facts about the signified state of affairs. 12 Millikan also takes perceptions and beliefs to be intentional signs. The ‘senders’ in these cases are subsystems of the animal: perceptual same-trackers and cognitive unitrackers. The ‘receivers’ are distinct subsystems: either motivational systems that use the signs produced by perception and cognition to guide action, or reasoning subsystems that modify beliefs through inference or checking for consistency. 13 Millikan handles prescriptive language in a parallel way, as fixing satisfaction conditions for the prescription. Successful prescriptive communication involves the sender performing the very act prescribed by the sender – which requires the two to co-ordinate on the same action-guiding uniceptual content. 14 The discussion of self-signs is relevant to Millikan’s later characterization of mechanisms involved in keeping track of identities in thought through ‘situated signs’ (Ch. 15) and her account of how elements of the context can be recruited in interpreting natural language (Ch. 16). 15 Given Millikan’s views about contextual factors being incorporated into semantic content, however, I suspect these are cases where the semantic/pragmatic distinction becomes fuzzy. References Burge T. 2010 . Origins of Objectivity . Oxford : Oxford University Press . Dickie I. 2015 . Fixing Reference . Oxford : Oxford University Press . Evans G. 1982 . The Varieties of Reference . New York, NY : Oxford University Press . Grice P. 1957 . Meaning . Philosophical Review 66 : 377 – 88 . Google Scholar Crossref Search ADS Lawlor K. 2001 . New Thoughts about Old Things: Cognitive Policies as the Ground of Singular Concepts . New York, NY : Garland Publishing . Millikan R.G. 1984 . Language, Thought, and Other Biological Categories . Cambridge, MA : MIT Press . Millikan R.G. 2000 . On Clear and Confused Ideas . Cambridge : Cambridge University Press . Peacocke C. 1992 . A Study of Concepts . Cambridge, MA : MIT Press . Perry J. 2001 . Reference and Reflexivity . Palo Alto, CA : CSLI Publications . Recanati F. 2012 . Mental Files . Oxford : Oxford University Press . Schroeter L. 2008 . Why be an anti-individualist? Philosophy and Phenomenological Research 77 : 105 – 41 . Google Scholar Crossref Search ADS © The Author(s) 2019. Published by Oxford University Press on behalf of The Analysis Trust. All rights reserved. For permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
TI - Beyond Concepts
JF - Analysis
DO - 10.1093/analys/anz015
DA - 2019-04-01
UR - https://www.deepdyve.com/lp/oxford-university-press/beyond-concepts-Flb5UQubXZ
SP - 363
VL - 79
IS - 2
DP - DeepDyve
ER -