The Open University, UK.
k.frankish@open.ac.uk

There has been growing interest recently in the so-called cognitive conception of language – the idea that some human thought processes constitutively involve the representational resources of the language faculty – or, more colloquially, that we can think in natural language. This view is, I think, very attractive: there are theoretical reasons for endorsing it and introspection supports it, too. However, I shall not be defending the cognitive conception here. Instead, I shall be asking how a language-based cognitive system – a linguistic mind, as it were – might have evolved. I begin by outlining some evolutionary criteria which a satisfactory version of the cognitive conception must meet. I then look at some recent versions of the doctrine and ask which best satisfies them. This will give us a reason for preferring that version, should we decide to endorse the doctrine in the first place.

Many questions arise once we adopt an evolutionary perspective. For the present, however, I shall concentrate on just two:

1) How did natural language become involved in central cognition? It is widely accepted that the language faculty is a modularized, peripheral system, which is relatively encapsulated from the rest of cognition and which originally evolved for communicative purposes. How did such a system come to play a role in central cognition (that is, in flexible, intelligent, nonencapsulated, conscious thought)?

2) When did language-based cognition evolve? Most writers agree that grammatical language evolved some time within the last quarter of a million years. This does not leave much time for the subsequent development of language-based cognitive mechanisms – not, at any rate, if this would have involved alterations in neural anatomy.

Any satisfactory version of the cognitive conception must address these questions and any version which conspicuously lacks the resources to do so can be ruled out in advance. So, for example, we can rule out the hypothesis that a hardwired language-based cognitive processor evolved subsequently to the emergence of language. There would simply have been no time for such a system to develop. There remain, however, a number of viable candidates and the bulk of this paper is devoted to assessing them.

According to Derek Bickerton, language and central cognition co-evolved. Structured non-demonstrative thought ('off-line thought' as Bickerton calls it) requires a system of schematic representations and a set of combinatorial principles defined over those representations. Language, too, requires a set of schematic representations and a combinatorial syntax, and parsimony suggests that the same neural resources play both roles.

This suggestion has attractions. If Bickerton is right, then there is no problem of how language got involved in central cognition: it is central cognition – at least, in so far as central cognition involves off-line thinking. The ‘when’ question is also dispelled. There was no need for further adaptation after the development of language; the development of fully grammatical language was the development of structured off-line thought.

The proposal also has some serious drawbacks, however. I will mention two. First, it is unlikely that language possessed cognitive functions from the very start. For it would have been simpler to build a purely communicative language system than to build one which had both communicative and cognitive functions. A communicative language system requires only syntax, phonology, and comprehension systems, together with a lexicon. And while these could be adapted to play a role in cognition, they would not in themselves constitute a cognitive system. Additional subsystems would be required – in particular, some sort of central processor. (Syntax alone might give you structured thoughts, but not structure-sensitive thought-processing.) But if it would have been easier to construct a merely communicative language system than one that had cognitive functions, then we should expect the former to emerge before the latter – and not to co-evolve with it. Secondly, Bickerton ignores evidence for modular structure within central cognition. He tends to view the whole of human central cognition as language-based – and thus to suppose that it is a fairly recent system, with little inherited structure. Yet there is mounting evidence that we have lots of innate cognitive competences, realized in functionally distinct, partially encapsulated modules which have developed gradually over last million years or so, in many cases predating language. Such evidence tends to undermine Bickerton’s solution to the ‘how’ question. If the linguistic mind is not the whole of central cognition, then how is it related to the rest of it?

Is the cognitive conception compatible a modularist view of central cognition? Peter Carruthers suggests so, building on suggestions by Steve Mithen. According to Mithen, the human mind developed in three phases. In phase 1, it consisted of a rudimentary general purpose problem-solving system. In phase 2, this was supplemented by a number of self-contained domain-specific modules, which were fast but inflexible, and did not communicate with each other or with general intelligence. Finally, in phase 3, there was a growth of ‘cognitive fluidity’. The previously isolated central intelligences began to communicate with each other and with general intelligence – either through direct channels, or through the mediation of a metarepresentation supermodule. Carruthers argues that this picture naturally supports a version of the cognitive conception. Even while remaining internally isolated, he points out, the central modules would have formed input-output links with the language faculty. Natural language would then have been the obvious vehicle for inter-modular information transfer, once the internal barriers started to come down. Natural language would thus have come to serve as a cognitive lingua franca.

I have two worries about this proposal. First, it does not amount to a full-blooded vindication of the cognitive conception – not, at least, if we take that doctrine to involve the claim that language can act as a medium of inference as well as thought. In Carruthers's scenario all the real inferential work done within modules, in their own internal representational media, and natural language functions merely as a conduit between them. Secondly, the proposal does not fully resolve the 'how' question. A neural lingua franca may be a necessary condition for inter-modular co-operation, but it is not a sufficient one. Coherent trains of thought do not just spring into existence spontaneously – a problem has to be identified and the various modular resources deployed intelligently to its solution. Some kind of executive would thus be needed to marshal the problem-solving resources of the different modules and to co-ordinate their outputs. Moreover, this system would need to process sentences in a way that was sensitive to their semantic properties. But then it starts to look like the sort of hard-wired language-based cognitive processor whose existence we have already ruled out.

We been thinking of the linguistic mind and the processing mechanisms that support it as part of the brain. But perhaps this is wrong. Perhaps it is more like a program running on the brain – a feature of our mental software, not of our neural hardware. Such a view has been defended by Daniel Dennett. The modern human mind, Dennett claims, is not a biological system at all, but a virtual machine – the product of learned behaviours (Dennett calls them 'good tricks' or 'memes') which have reprogrammed our biological brains. The behaviours in question, Dennett suggests, are linguistic. We acquire virtual minds by talking to ourselves – by producing, rehearsing and rearranging sentences in overt or silent soliloquy. This stream of inner verbalization transforms the activity of the biological brain, causing its parallel multi-track hardware to simulate the behaviour of a serial, single-track processor, operating upon natural language sentences. Dennett calls this softwired system the Joycean machine.

This story is particularly attractive from our current perspective. There is no special problem about how or when the virtual mind evolved: its development involved a process of memetic or cultural evolution, not the emergence of new neural structures. Still, the story will not do as it stands. The problem lies in the way the Joycean machine is supposed to work. According to Dennett, the key mechanism is one of self-stimulation. Inner speech is channelled through a feedback loop from speech production to speech comprehension. Internally generated sentences are then processed by the comprehension system just like externally produced ones, often evoking similar responses. So, for example, questioning yourself may prompt an instinctive verbal reply, containing information that would otherwise have been inaccessible. Dennett suggests that neural subsystems routinely compete for control of the vocal system and the self-stimulatory mechanisms it supports. As a result, the Joycean machine comes to act both as a bulletin board, where locally stored information is made globally available, and also as sort of virtual executive, focusing attention, marshalling resources and co-ordinating the activities of different subsystems.

It is likely that inner verbalization does have a self-stimulatory function of this sort. But self-stimulation cannot be all there is to the linguistic mind. For one thing, it is doubtful that it could generate sustained trains of intelligent thought. Self-stimulation might help to produce some regularity and consistency in one’s inner verbalizations, but it is hard to see how it could yield coherent inferential sequences of the kind involved in dedicated problem-solving. In such cases, it seems, our inferential subsystems are not merely competing for vocal control, each shouting out its favoured (and not particularly bright) solution. Rather they co-operating, each subordinating its activity to a global objective. And it is hard to see how they could be induced to do this without executive supervision of some sort. Secondly, self-stimulations will not have the cognitive role typical of linguistic thought. Consider the sort of cases that lend intuitive support to the idea that we can think in language. I notice that the steering on my car is uneven and say to myself ‘The wheel alignment needs checking.’ Here, it seems, I am not instructing or encouraging myself to think that the alignment needs checking; I am judging that it does. And this judgement may have long-term effects – say, in getting me to the garage the following day – which a transient self-stimulation would lack.

Dennett gives us part of the story, then, but important aspects of the linguistic mind remain unaccounted for. Can we complicate his picture in order accommodate these features? I think so. The trick is to think of linguistic reasoning as, to some extent, under personal control. The linguistic mind, I suggest, is indeed a virtual one, constructed by the discovery and transmission of good tricks. But these include, not just inner verbalization, but also various meta-cognitive and meta-linguistic skills. We do not only speak to ourselves, I suggest; we also adopt attitudes towards our inner verbalizations and perform explicit inferential operations upon them. In particular, we adopt some as explicit premises and others as explicit goals, and manipulate them so as to construct chains of reasoning, using learned inferential skills. I have suggested elsewhere that these activities constitute a distinct level of mentality, which is intentionally formed and sustained, and which constitutively involves natural language. I call this the supermind.

This proposal retains all the advantages of Dennett's. Like the Joycean machine, the supermind is the product of memetic and cultural evolution, rather than changes in the brain. Indeed, there is a plausible story to tell about just how it emerged. The metalinguistic and metacognitvie skills needed to develop a supermind – the ability to think about one's thoughts and words and to articulate cogent trains of argument – are just the skills needed for engaging in reasoned argument with one's peers. And as such, their possession would have carried huge benefits in early human society – for example in securing an equitable distribution of resources and in attracting mates (the ‘big man’ hypothesis). There would thus have been strong independent pressure for their development. And supermental abilities would have followed naturally in their wake. Humans would have begun to develop linguistic minds as soon as they started to internalize their skills in interpersonal argument, reasoning and debating with themselves. (The supermind, we might say, is not so much a Joycean machine as a Beckettian one.)

And supermind theory can, I believe, resolve the problems facing Dennett's account. Take the question of executive control. For self-conscious agents, equipped with metacongitive skills, problem-solving will assume a dual aspect. They will be able to think, not only about the first-order problem of what to do or think next, but also about the metaproblem of how to solve that problem. Suppose they have some general ideas about how to solve this metaproblem. So, they want to evaluate candidate hypotheses as they occur to them, preferring those that harmonize well with premises and goals they have previously endorsed, and rejecting those that conflict with them. And as a subgoal they want to trace out the implications of each hypothesis, searching for data that might confirm or refute it, or indicate how it should be revised. These desires then drive their subsequent attempts to tackle the first-order problem. As various candidate hypotheses occur to them (thrown up, let us suppose, by submodules), they set to work evaluating them, in line with their metacognitive goals. They persist in this, refining and complicating their hypotheses, until they reach a solution that satisfies their goals. In short, their metacongitive attitudes regulate their first-order problem-solving activities. In effect, they can act as their own central executives, marshalling and directing their low-level cognitive resources.

Supermind theory can also explain how inner verbalizations can assume a direct cognitive role, rather than a merely self-stimulatory one. We can decide that an inner verbalization will have the role of a thought by deciding to adopt it as premise and to use it as a basis for inference and action. Executing such a decision will involve using one’s explicit reasoning skills to make sure that the verbalized proposition has the appropriate inferential role – for example, by using it as a premise in one’s syllogisms. Language-based reasoning will thus be genuinely computational – though the computations in question will be carried out at an explicit, personal level.

Of the candidates reviewed, then, I suggest that supermind theory is the best placed to provide a full-blooded and neurologically plausible defence of cognitive conception. If we have indeed developed a linguistic mind, then this is form it is most likely to have taken, given the constraints on its evolution.

 Conference site: http://www.infres.enst.fr/confs/evolang/