A theory of implicit and explicit knowledge

This is the unedited penultimate draft of a BBS target article that has been accepted for publication (Copyright 1998: Cambridge University Press -- publication date provisional) and is currently being circulated for Open Peer Commentary. This preprint is for inspection only, to help prospective commentators decide whether or not they wish to prepare a formal commentary. Please do not prepare a commentary unless you have received the hard copy, invitation, instructions and deadline information.

A Theory of Implicit and Explicit Knowledge

Keywords

Implicit knowledge, consciousness, automaticity, memory, cognitive development, visual perception, artificial grammar learning

Abstract

The implicit-explicit distinction is applied to knowledge representations. Knowledge is taken to be an attitude towards a proposition which is true. The proposition itself predicates a property to some entity. Number of ways in which knowledge can be implicit or explicit emerge. If a higher aspect is known explicitly then each lower one must also be known explicitly; this parital hierarchy reduces the number of ways in which knowledge can be explicit. The most important type of implicit knowledge consists of representations that merely reflect the property of objects or events without predicating them to any particular entity or event. The clearest case of explicit knowledge of a fact are reflective representations of one's own attitude of knowing that fact. These distinctions are discussed in their relationship to similar distinctions like procedural-declarative, conscious-unconscious, verbalizable-nonverbalizable, direct-indirect tests, and automatic-voluntary control. This is followed by an outline of how these distinctions can be used to integrate and relate the often divergent uses of the implicit-explicit distinction in different research areas. We illustrate this for visual perception, memory, cognitive development, and artificial grammar learning.

Acknowledgements

We wish to thank Bruce Bridgeman, John Campbell, Peter Carruthers, Ron Chrisley, R.Carlson, Greg Currie, Tony Marcel, Gabriel Segal for invaluable discussions and Peter Carruthers, John Kihlstrom, Pierre Perruchet, and Carol Seger for their informative reviews.

Our basic strategy for meeting these objectives is to analyse knowledge as a propositional attitude according to the representational theory of mind (RTM Field, 1978; Fodor, 1978). Roughly speaking, if I know a fact (e.g., the animal in front of me is a cat) then, according to RTM, I have a representation of that fact and the internal, functional use of this representation constitutes it as knowledge of mine (rather than as a desire of mine, etc.). The central idea of how the implicit-explicit distinction applies is that knowledge can vary depending on what is represented (made explicit) and which aspects remain implicit in the functional use of representations. This application of the implicit-explicit distinction has several advantages.

The main advantage of our analysis is that it provides a common ground for the use of the implicit-explicit distinction in different fields of investigation. For instance, consider Schacter's (1987) influential definition of the implicit-explicit memory distinction: "Implicit memory is revealed when previous experiences facilitate performance on a task that does not require conscious or intentional recollection of those experiences; explicit memory is revealed when performance on a task requires conscious recollection of previous experiences." This definition may capture the phenomenal experience of implicit and explicit memory very well, but it leaves open how the definition is to apply to implicit and explicit knowledge in other fields. For instance, Karmiloff-Smith (1986, 1992) has argued that there are several steps of explicitation before consciousness is reached. Identification of explicit with conscious gives us no understanding of why Karmiloff-Smith's lower forms of explicitness have anything to do with this distinction. In other words, although it has been suggested to break up the implicit-explicit dichotomy into a series of levels of explicitness our analysis is needed to explain just what it is that becomes more explicit as one ascends levels and to relate proposed levels in one research area to different subdivisions of explicitness in other areas.

Existing problems of this kind with the implicit-explicit distinction are many. In memory research and subliminal perception research, explicitness has been linked to performance on direct tests in comparison to performance on indirect tests (Richardson-Klavehn & Bjork, 1988; Reingold and Merikle, 1993) because performance on direct tests seems to require conscious awareness. But the interesting question left open is why direct tests require consciousness. Or in visual perception it is found that touching an object is based on unconscious, implicit information whereas pointing to the object requires conscious, explicit information that is subject to visual illusions (e.g., Bridgeman, 1991; Milner & Goodale, 1995, Rossetti, 1997). Why? Also, more directly, what are the representational requirements for conscious awareness? What is the relation between knowledge we have voluntary control over and knowledge we are aware of? Why can we sometimes in limited ways control knowledge we are not aware of (Dienes, Altmann, Kwan, & Goode, 1995)? Can predictions be made for the conditions under which knowledge will be represented implicitly? With our analysis of the implicit-explicit distinction we are able to give some answers to these questions.

Another advantage of our analysis is that it is grounded in the natural use of the terms "implicit" and "explicit" as typically occurring in the context of verbal information (e.g.: "They didn't say so explicitly, it was left implicit"), whereas traditional ways of explicating this distinction have ended in defining it in terms of other related distinctions. As mentioned Schacter (1987, p. 501) defined implicit memory by its lack of conscious or intentional recollection, and Reber (1993, p. 5) defined implicit learning as "...the acquisition of knowledge that takes place largely independently of conscious attempts to learn and largely in the absence of explicit knowledge about what was acquired." These definitions of implicit memory/learning raise the question of why the terms implicit/ explicit are used at all. Why not call explicit memory or learning directly by their name, that is, conscious memory or conscious learning? (cf Reingold and Merikle, 1993, p. 42). Moreover, when using technical terms with an existing natural meaning, it seems to us, we should adhere to that existing meaning as far as possible and not impose some arbitrary `operational definition', or else we make it difficult for the scientific community to share the same meaning, since the natural meaning is likely to keep intruding. (Who still adheres -- or ever has adhered -- to the operational definition of intelligence as that which the WAIS measures?). So, it is not an unimportant feature of our use of the implicit-explicit distinction that it attempts to stay true to its natural meaning, which we believe was the unarticulated reason for introducing the distinction in the first place, and what partially motivated its acceptance and continued use.

From the natural meaning of implicit-explicit in the context of language we say that a fact is conveyed explicitly if that fact is expressed by the standard meaning of the words used. If something is conveyed but not explicitly then we say that it has been conveyed implicitly. We can discern two main sources of implicitness. One source is the contextual function/use of what has been said explicitly. A prime case are presuppositions. To use a famous example, the statement, "The present king of France is bald," presupposes that there is a present king of France. It does not express this fact explicitly because the function of the sentence (when uttered as an assertion) is to differentiate the present king of France being bald from him not being bald. For that reason the speaker of this sentence can claim that he DID NOT (explicitly) say that there was a king of France. Yet the presupposition does commit him to there being a king of France, or else his assertion of the king being bald becomes insincere. So in this sense he did (and thus we say: "implicitly") convey that there is a king of France.

The other source of implicitness lies in the conceptual structure of the explicitly used words. For instance, if one conveys that a person is a bachelor, then one conveys that this person is male and unmarried without making these features explicit. By using "bachelor" the speaker commits herself quite strongly to "male" and "unmarried" lest she shows herself ignorant of the meaning of the word bachelor in the particular language spoken. These are not rare cases. Whenever we say that something is an X (e.g., a bird) then we implicitly convey that it is also an instance of the super-ordinate category of X (e.g., an animal) on these same grounds as in the bachelor case.

Our analysis of knowledge locates the same distinction of implicit and explicit in terms of which parts of the knowledge are explicitly represented and which parts are implicit in either the functional role or the conceptual structure of the explicit representations. We define a fact to be explicitly represented if there is an expression (mental or otherwise) whose meaning is just that fact; in other words, if there is an internal state whose function is to indicate that fact. [1]Other, supporting facts, that are not explicitly represented but which must hold, in order for the explicitly known fact to be known, are implicitly represented .

2.1 Implicitness arising from functional role.

Our various mental concepts like knowledge are standardly analysed as propositional attitudes (Russell, 1919). That is the sentence "I know that this is a cat" consists of a person (I), a proposition (this is a cat) and an attitude relation between person and proposition (knowing). The representational theory of mind (Field, 1978; Fodor, 1978) says something about how such an attitude can be implemented in our mind. The suggestion is that the proposition is represented and the attitude results from how this representation is used by the person (functional role). That is, the representation "this is a cat" constitutes knowledge if it is put in a -- philosophers would say -- knowledge box or -- cognitive scientists would say -- data base . That means that the representation is used as a reflection of the state of the world and not, e.g., if it were in a goal box, as a typically nonexisting but desirable state of the world.

In this view we can say that the content of the knowledge is explicit since it is represented by the relevant representational distinctions (in analogy to explicit verbal communication). That is, there is an internal state whose function is to indicate the content of the knowledge. In contrast the fact that this content functions as knowledge is left implicit in its functional role [2] (like implicitly conveyed information is communicated by the functional necessities created by the explicit part). Also the fact that it is myself who holds this knowledge is not explicitly represented but is implicit in the fact that it is me who holds that knowledge. We have, thus, three main types of explicit knowledge depending on which of the 3 constituents of the propositional attitude is represented explicitly:

2.1.1 Content

The content of a propositional attitude, like knowledge, is that which the attitude is about. In our example of the cat that I see in front of me I know that it is a cat. The representation of the content of this knowledge as "this is a cat" identifies (1) a particular individual (i.e., the animal in front of me), (2) a property (or natural kind: catness), and (3) it predicates this property to the particular individual. To gain a more succinct and more general way of expressing these aspects we use predicate calculus notation, where F, G,... denote properties, a, b, ... denote particular individuals, and the syntactic combination of F and b into the formula Fb expresses that F is predicated to b (as opposed to `F,b' where the comma would indicated that F and b are just being listed and no predication takes place).

As an extreme case in which almost everything is left implicit we consider Strawson's (1959, p. 206) "naming game" in which a person simply calls out the name of a presented object, e.g., "cat" or "dog" depending on which kind of animal is presented. In this context the word "cat" expresses knowledge of the fact that `this (object in front of the person) is a cat' and it conveys this information to the initiated listener. We couldn't say anything less, e.g., that it only expresses knowledge of cat-ness, or of the concept of cat. Yet, what is made explicit within the vocabulary of this naming game are only the properties of being-a-cat, being-a-dog, etc. Consequently, since there is knowledge that it is the particular presented individual that is a cat or dog, that knowledge remains implicit. [3]

So, our use of Strawson's naming game provides an example of only the property (cat) being represented explicitly and the individual and predication of the property to this individual remaining implicit. It helped to introduce this issue with the naming game since it uses the publicly inspectable medium of language. However, when it comes to the question of which aspects can be made explicit independently of other aspects the naming game becomes an imperfect guide for explicitness of mental representations as the following shows.

In the naming game it is also possible to represent individuals explicitly and leave their properties implicit. This is the case for forced choices between two items, i.e., by pointing to that item that has a particular property, e.g., which one of two objects is a cat. In the case of the naming game one could argue that for this the response must explicitly distinguish the two items (a, b) by pointing right or pointing left, but not the property. The pointing thus conveys the information `This one is a cat' but makes only `this one' explicit and leaves `is a cat' implicit. In the case of the naming game, i.e., the information passing between two communicating parties, this is possible. But in the case of the knowledge that a single person must bring to bear explicitness of the individuals requires explicitness of the attributed property, because the person must be able to go into a cat/no-cat state for each individual in order to decide which individual is a cat and then respond correctly. Hence, for knowledge we have the constraint that explicit representation of the individual to which a property is attributed entails explicit representation of that property.

At this point one should be made aware that the notion of predication to a particular individual need not be restricted to particular objects or persons. It will be used later in extended form to events and even causal regularities. Traditional logic does not make this very explicit but Barwise and Perry's (1983) Situation Semantics offers an elaborate distinction between event types and individual events, in order to capture the facility of natural language to freely reference particular events, causal regularities, laws, etc. and then describe them as having certain properties or being of a certain type. For instance, a particular event (b) was a dance (F) and has the further features of having had me as a participant (G) etc.

Subliminal perception provides a psychological research example, as discussed in more detail in Section 3.2. The suggestion is that under subliminal conditions only the property of a stimulus (kind of stimulus) gets explicitly represented (e.g., the word "butter") but not the fact that there is a particular stimulus event that is of that kind. This would be enough to influence indirect tests, in which no reference is made to the stimulus event (e.g., Naming milk products), by raising the likelihood of responding with the subliminally presented stimulus (i.e., "butter" is listed as a milk product more often than without subliminal presentation). The stimulus word is not given as response to a direct test (e.g., Which word did I just flash up?) because there is no representation of any word having been flashed up. Performance on a direct test can be improved with instructions to guess (Marcel, 1993) since this gives leave to treat the direct test like an indirect test to just say what comes to mind first.

As mentioned earlier, even explicit representation of F being predicated to b ("Fb", or "This is a cat") leaves implicit the fact that Fb is a true proposition, i.e., a fact at the present time. Only the representation "Fb is a fact now" represents the fact that b is F at the present time completely explicitly. The reason for making these aspects explicit may seem superfluous.In particular, the addition "is a fact" may strike some readers as totally redundant and trivial, so let us briefly dwell on its significance.

Consider a simple mental system that does not represent truth explicitly but just contains a single model of how it perceives the world to be (Perner, 1991, described the young infant as having only this representational power). The model of the world is a type of knowledge box in that any proposition Fb that is in the knowledge box is taken (judged) as true, on the grounds of being in that box and the functional role this box plays in the mental economy. However, there is no possibility of representing propositions that are not true without creating mental havoc because all propositions in the box are acted upon as if they were true (Leslie, 1987, pointed this out in his analysis of pretence). To differentiate true from false propositions one could represent false propositions in a different functional box, as has been suggested for pretence and counterfactual reasoning (Currie & Ravenscroft, in press; Nichols and Stich, 1998). In concrete terms this means that a child who is pretending that the banana is a telephone represents, "this is a banana (Bb)" in its knowledge box and, "this is a telephone (Tb)" in its pretend box. This solution may be adequate for pretend play consisting of switching from a knowledge (serious action) mode into a pretend mode of functioning. Pretend actions are then simply governed by the representations inside the pretend box. It cannot account for the child knowing what it is pretending. To know that the pretend representations have to be in the knowledge box. That raises the problem of cognitive confusion (representational abuse, Leslie, 1987) and the pretend representations have to be quarantined in some sort of "metarepresentational [4]context" (Sperber, 1997). Such markers explicitly differentiate within the knowledge box what is to be taken as true from what is not to be taken as true. More generally speaking, for knowing what is true and what is not true the truth value has to be made explicit within the knowledge box, i.e., to represent "Fb is a fact" or "Fb is NOT a fact". [5]This distinction is also required for understanding change over time, i.e., to represent that Fb was the case and now Gb is the case (Perner, 1991; 1995, Appendix) and to interpret symbolic expressions and representations, e.g., to understand that objects in the world are also in the picture. [6]

The following table gives a summary of the different cases of the possible implicit-explicit combinations of facts that we have discussed so far. And we also claim that these are the only realistically possible ones.

This table excludes certain permutations of the four elements property, individual, predication and factivity. For the verbal exclamations in Strawson's naming game all combinations are possible, but for knowledge only the four cases listed above are possible. For instance, predication cannot be known explicitly on its own. It can be explicitly conveyed on its own in the naming game in response to the question "Does b have the property F?" The response "Has-it/doesn't have it" represents only predication explicitly. But, again, a system that can do this must make further internal distinctions, i.e., it must distinguish F from not-F in order to decide whether the presented object "has/doesn't-have" that property. Knowledge of the presented individual can remain implicit. This case is accounted for in 2(b) above.

In the case of factivity we are after the distinction between a state of affairs Fb being a fact or being fiction. The naming game can only be played with real objects. A system that can meaningfully distinguish between whether the predication of F to b holds in the real world or in a world of fiction, must have the representational resources to specify the property and the individual in question and the predication of this property to the individual in order to decide whether this predication holds in reality or only in fiction. Hence, if factivity is explicitly known then predication, individual, and property must also be explicitly known. Similarly, the time of a fact can only be left implicit for the present. A system that can meaningfully distinguish between whether the predication of F to b holds now or in the past, must have the representational resources to specify the property and the individual in question and the predication of this property to the individual in order to decide whether this predication holds now or has held previously. Hence, if time is explicitly known then predication, individual, and property must also be explicitly known.

Memory research provides a relevant example for these considerations. Explicit memory is not only conscious, but more to the point, a recollection of the past. For this it must represent past events as having taken place in the past. Only then can systematic answers be given to direct questions about the past. If a past event is only represented by its properties (event structure) then it can influence indirect tests and direct tests alike. Only when pastness of the event is represented explicitly can performance on a direct test that addresses the pastness directly outshine performance on indirect tests (see Reingold and Merikle's, 1993, criterion for explicit memory). So, we can see why and how directness of test relates to explicitness. In the next section we see how it relates to consciousness.

2.1.2 Attitude

All of these facts, possession, accuracy, judgement and causal origin (justification) , are supporting facts for any representation to constitute knowledge. E.g., "Fb is a fact" constitutes knowledge of the fact that b is F for a system only if (o) the system has the representation, (i) it is accurate, (ii) it is treated by the system as an accurate reflection of the world (the world is judged to be so) and (iii) it came about in a proper causal (justifiable) way. Hence all four facts are implicit in any knowledge until made explicit.

Only if the system can entertain R as a representation it possesses can the system represent what further properties (e.g. (1), (2), and (3)) this representation might have. But the three further reflections can be explicit independently of each other. Truth does not imply having been properly caused nor being taken for true, being taken for true does not imply either that it be true or that it was properly caused, and having been properly caused does not imply being taken for true nor that it must be true because, although generally reliable, even such a process can on occasion fail [7]. Note that some dependencies emerge if one represents that it is the same rational agent (e.g. oneself) who represents R as accurate and who represents R as being taken to be true.

2.1.3 Relating Explicitness of Content, Attitude and Self.

It is evident that explicit representation of self as holder of an attitude (e.g., "I know ...") contains an explicit representation of the attitude ("know"). The interesting question is to what degree explicit representation of knowing requires explicit representation of the content (e.g., "this is a cat"). That is: Is it possible to explicitly represent "I know" or "it is known" and leave implicit the fact that this is a cat (Fb) . In a variation of the naming game an expression like, "I know," can be implicitly conveying that the knowledge is of the fact that Fb. However inside a (rational) agent this explicit reflection on knowledge implies explicit factivity of the known, i.e., the agent must be able to judge the factivity of the known fact before coming to the conclusion that one knows that fact. Since explicit factivity implies explicitness of predication, individuals, and properties we can conclude that explicit representation of self or attitude implies explicit representation of the content.

The dependencies that we have discussed are summarised in Figure 1. If an aspect at a higher level is represented explicitly (at the origin of an arrow) then -- according to our analysis -- all aspects at a lower level (at the end of the arrow) need also be explicitly represented.

On the basis of this partial hierarchy we will later speak conveniently of "fully-explicit" knowledge when all aspect are explicitly represented, of "attitude-explicit" when everything up to the attitude is explicit, and of "content-explicit" if all the aspects of content are represented explicitly. Conversely, we use "attitude-implicit" to indicate that attitude and all higher aspects in the hierarchy are left implicit, and so on for the other aspects. Moreover, it is often convenient to differentiate between different levels within content: "fact-explicit" (equivalent to "content-explicit") when all aspects of content are explicit, "predication-explicit" when predication, individuals and property are made explicit (for simplicity sake we ignore the possibility of case 2b in Table 1), and of "completely implicit" if only properties remain explicit.

On an important cautionary note one has to point out that these hierarchical constraints only hold for a single representation. That is a single representation cannot make something explicit at the higher level and still represent aspects at a lower level implicitly. This, of course, does not preclude the possibility of there being two independent representations, one of which makes something explicit at a the higher level and the other representing something at the lower level implicitly. For instance:

2.2 Implicitness Due to Conceptual Structure.

This kind of implicitness ( structure implicitness ) arises typically in the case where the system represents (has a concept for) properties that can be defined as compounds of more basic properties, e.g., the property of being a bachelor has the components of being male and unmarried. So, if someone explicitly states that a person is a bachelor, then she implicitly conveys that he is also unmarried since being unmarried is a necessary, supportive fact for being a bachelor. Similarly, a person can explicitly know that someone is a bachelor, but not explicitly know that that person is not married. However, since not being married is a necessary fact for being a bachelor, this fact is known implicitly. In this example the structure of the component properties (male, unmarried, etc.) remain implicit in the explicit representation of the compound property (being a bachelor): a case of "property-structure implicitness". Roberts and MacLeod (1995) argued that concepts acquired incidentally and nonstrategically may have atomic nondecomposable representations; i.e., in which the property structure is represented implicitly in our terminology.

2.3 Summary.

We have so far developed a rich structure for describing different ways of how some knowledge can be implicit within the use of some other explicitly represented knowledge. That is, knowledge with explicit representations of part of its content can contain other parts of its content, the attitude and self as holder of the attitude implicitly. Also, explicit knowledge can consist of representation of compounds (typically: compound properties) that leaves the structure of its components implicit. We now explore how our analysis unifies the different distinctions that have traditionally been used to define or characterise or been brought in contact with the implicit-explicit distinction.

We have shown in the previous section that knowledge can differ in the amount of its functional and conceptual aspects that are represented explicitly. This puts us into a position to now show that the various distinctions that have been associated with the implicit-explicit distinction differ in the amount of explicit representation required. We start with consciousness since it has most prominently been used to define explicit knowledge (in memory, Schacter, 1987; in learning of rules, Reber, 1989). We will show that under a common understanding of "conscious" knowledge counts as conscious only if its content, the attitude of knowing and the holder of that attitude (self) can be represented explicitly. Hence, conscious knowledge is, indeed, prototypically explicit.

Consciousness has often been brought into close contact (even defined in terms of) verbalizability (e.g., Dennett, 1978) and the ability to address the content of one's knowledge verbally (direct tests) has often been used to characterise tests diagnostic of conscious and explicit knowledge. This makes sense in our analysis, since verbal reference requires very explicit representation of content. Furthermore, a close relative of verbally expressible knowledge has been "declarative" knowledge, which has often been put in opposition to "procedural knowledge." Although this opposition confounds several independent dimensions: procedural-inert, declarative-nondeclarative, and accessible-inaccessible, we can explain why these groupings appear natural and why they can be tied to the implicit-explicit distinction. Finally, the ability to exert voluntary control, in contrast to automatic action, has been tied to explicit, conscious knowledge. We can show that this linkage is justified, because--so the argument goes--voluntary control requires explicit representation of one's attitude which conforms to the requirement for conscious awareness, whereas automatic action can be sustained by procedural know-how.

We use "consciousness" (some philosophers might find the term "conscious awareness" more appropriate [10]) here as--we think--most people use it, i.e., that ones knowledge is available to oneself and that it is not necessary to prove its existence to one's own surprise through behavioural evidence. This is certainly the meaning given to the conscious-unconscious distinction in cognitive psychology, as we will see from the many research examples in the next section. For instance, implicit unconscious memory is exactly where I appear to have no knowledge (memory) of a past event but can be shown by behavioural evidence in an indirect test that I do have some (implicit) knowledge of that event.

The idea that consciousness has something to do with awareness of our mental states has a venerable tradition dating back to at least the writings of John Locke (cit. Tye, 1995, p. 5): "consciousness is the perception of what passes in a Man's own mind." And perhaps even to Aristotle (Güzeldere, 1995, p. 335). This intuition has recently been given prominence under the name of the Higher-Order-Thought Theory of Consciousness. Different versions of this theory differ as to the nature of the second-order state required. For instance, Armstrong (1980) sees it as a perceptual state--like Locke, as a higher order act of observing our first order mental states--, Rosenthal (1986) sees it as a more cognitive state, and Carruthers (1996) sees it as a potential for being recursively embedded in higher-order states (see Güzeldere, 1995). The basic insight behind these different approaches is that to be conscious of some state of affairs (e.g., that the banana in my hand is yellow) then I am also aware of the mental state by which I behold this state of affairs (i.e., that I see that the banana is yellow). There is something intuitively correct about this claim, because it is inconceivable that I could sincerely claim, "I am conscious of this banana being yellow" and at the same time deny to have any knowledge about whether I see the banana, or hear about it, or just know of it, or whether it is me who sees it, etc. That is, it is a necessary condition for consciousness of a fact X that I entertain a higher mental state (second order thought) that represents the first order mental state with the content X.

Of course, there is philosophical controversy as to whether this characterisation can capture the whole phenomenon of consciousness or at best some aspect. [11] We need only focus on the less controversial part of this theory, namely that the higher order mental state is only necessary. Although, in the following we will occasionally explore the potential explanatory power of the stronger theory that a higher order thought is both necessary and sufficient for consciousness. Moreover, in order to stay on the safe side with our claims we will principally pursue Carruther's potentialist version of the higher order thought theory in more detail. Because it does not require actual entertaining of a higher order thought but only the potential for forming such a higher order thought, it makes less demands on the cognitive complexity of routine conscious information processing than the other versions of this theory. This potentialist version, nevertheless, is sufficient for our objective of explaining why consciousness relates to explicitness, verbal expressibility, voluntary control, etc.

Carruthers (1996) sees consciousness as the potential of our mental content to be recursively embedded in higher order states. In other words, the content X of a knowledge state is conscious if it is recursively accessible to higher order thoughts, e.g., knowing that I know that X. In order to form this second order state one needs to explicitly represent the first order knowing. For this in turn, we argued, one needs to represent the content explicitly, in particular its factivity, i.e., "it is a fact that X". This is a necessary condition. Interestingly, it is not always required to have the first order attitude and self explicitly represented because those can be gratuitously inferred from the factivity of its content as Gordon (1995) has pointed out in the context of simulation theory. Within one's own perspective--and that is all we are concerned with here--there is a one to one correspondence between what is a fact for me and what I know. Gordon speaks of ascent routines that allow us to go from descriptions of facts to knowledge attributions for oneself, e.g., from "X is a fact" I can go to "I know that X". That means that once factivity is represented explicitly, explicit representation of attitude and self is also possible. Of course, other conditions may have to be met (e.g., it must be in a short term memory store), but explicit representation of factivity (and thus all other aspects of content) is often all that is required.

In sum, on the weak version of the higher-order thought theory where potential access for higher order thoughts is only a necessary condition, we can conclude that explicit representation of self and attitude is necessary for conscious knowledge and sometimes explicit representation of factivity is all that is necessary for conscious knowledge. On the stronger version where access for higher order thoughts is also a sufficient condition, explicit representation of self and attitude or factivity is sufficient for conscious knowledge. The for us critical implication of this view of consciousness is that the required higher order states represent the attitude and holder of the first order state explicitly. As we have seen earlier, this in turn demands explicit representation of the content of the first order mental state. In sum, that means that to have conscious knowledge one must represent all three aspects of this knowledge explicitly (or be able to form such explicit representations). For instance, to consciously know that the banana is yellow, I must explicitly represent that it is a present fact that the banana is yellow, that this fact is known and (be able to explicitly represent) that it is me who knows it. Consequently, this analysis makes clear why most definitions of explicit knowledge involve consciousness, since it imposes the clearest, most extreme case of explicitness. It also puts us in good stead for understanding why verbal access to knowledge and other features to be discussed below are tied to consciousness.

Verbal communication (for transmitting information) proceeds by predication. A referring expression (or an ostensive gesture) is used to identify an individual (topic) and then further information about this individual follows. Hence, verbal report requires knowledge with explicit predication. An even stronger requirement of explicitness is necessary for the following reason. Unlike perceptual information linguistic information cannot be taken uncritically at its face value. As Gibson (1950) has emphasised visual perception is highly reliable under most normal circumstances and thus can -- barring the few visual illusions -- be taken as true. This strategy applied to linguistic information would lead to a highly unstable knowledge base (Perner, 1991, chapt. 4). For this reason verbal information needs to be interpreted without being taken as true at first. Only after evaluation (checking compatibility with other available information) should the information be accepted as true. To do this a distinction has to be made between 'is a fact' and 'not yet clear', i.e., factivity has to be represented explicitly.

In research on implicit memory (Richardson-Klavehn & Bjork, 1988) and subliminal perception (Reingold & Merikle, 1988) a critical distinction is made between direct and indirect tests of knowledge. A direct test is one that refers to the fact in question. An indirect test does not refer to the fact in question, but the answer to some unrelated question or reaction to some stimulus shows that some information about the fact must still be present. In both literatures, the fact in question is the spatio-temporal context of the presentation of a particular stimulus. The key methodological difference between implicit memory and subliminal perception is in terms of how long after the presentation of the stimulus, knowledge of this fact is tested (Kihlstrom, Barnhardt, & Tataryn, 1992). In implicit memory, the fact in question could be the fact that a particular word was studied 10 minutes ago in the laboratory, and typically the word is consciously perceived at the time of study. The implicit memory case is considered in more detail in section 5.2 below. In subliminal perception, the fact in question is whether a particular stimulus has just been presented. According to the normal approach (e.g. Holender, 1986), perception is regarded as subliminal or implicit (Kihlstrom et al, 1992) if the participant performs at chance on a direct test of some aspect of this fact (because it was not consciously perceived), but the stimulus still indirectly affects processing.

Our analysis makes clear why performance on indirect and direct tests has anything to do with implicit-explicitness and consciousness of the probed knowledge, provided the test questions are answered bona fide, i.e., participants say that X is the case only if they have a representation stating that X is a fact. The analysis makes also clear, however, that one cannot equate test performance with type of knowledge, since there is no guarantee that test answers are given bona fide, i.e., participants might say that X is the case even though they just act on a feeling that that might be right.

Even knowledge without explicit predication can influence indirect test responses, since the test does not refer to the event in question. For example, if after a brief (e.g., 10 msec) presentation of the word "doctor" or "table" followed (within, e.g., 50 msec) by a patterned mask (backward masking: a frequently used technique for achieving subliminal perception), a clearly visible word (e.g., "nurse") or nonword (e.g., "nurge") is presented and observers have to judge whether this item is a word or not, this lexical decision provides an indirect test of knowledge of the presentation of the first word. Although the task instructions refer only to the clearly visible word, it has been found (e.g., Marcel, 1983a) that if the first word is semantically related (i.e., "doctor") then identification of "nurse" is faster than if the first word is unrelated ("table"). For this processing advantage to occur it is sufficient to take in only the property of the presented stimulus, i.e., "doctor" without any representation that there was a particular event that had that property. For instance, the semantic processing triggered by the word form "doctor" will activate the semantic field of medical profession which then gives the ensuing "nurse" a greater processing advantage than "table".

In contrast, a direct test refers to the event in question. There are different ways of making this reference. The question can refer to the event, e.g. "What was the word on the screen?". A bona fide answer (it certainly is a fact) "doctor" can be given to this question only if the event has registered as a fact . So, we see that bona fide performance on such a direct test requires explicit representation of factivity which, on Carruthers potentialist higher-order thought theory of consciousness is at least a necessary and possibly also sufficient condition for consciousness. This provides a theoretical justification for using direct tests to assess conscious knowledge if all answers were bona fide. Unfortunately, there is no guarantee for that. Co-operative participants in our experiments try to give the best answer, and then even knowledge with implicit predication (far removed from meeting the criterion for consciousness) may help them give correct answers (correct guesses) to direct tests, a known problem in the field ( e.g., Roediger and McDermott, 1996).

Performance on indirect tests can be influenced by conscious knowledge as well as implicit knowledge lacking explicit predication. One could only infer the use of implicit knowledge that lacks consciousness from the difference between performance on an indirect test over a direct test (even if non bona fide answers are given on the direct test). This conclusion is warranted especially if performance on the direct test outstrips performance on the indirect test under conscious processing conditions so that any lingering issues about sensitivity differences (Shanks & St John, 1994) are eliminated (Reingold and Merikle, 1993, p. 53 ).

Since direct tests do not typically involve reference to one's subjective mental state of seeing, Cheeseman and Merikle (1984; see also Greenwald, 1992) referred to the threshold conforming to this test as the "objective threshold": If the interstimulus interval between a stimulus (e.g. a word) and a mask is reduced so as to make perception more difficult, the objective threshold is defined by the interstimulus interval at which the participant performs at chance on a direct test of the nature of the stimulus presented. However, our analysis suggests that this might not reflect a single threshold, since there are at least two theoretically significantly different ways of making such a reference (cf Dagenbach, Carr, & Wilhelmsen, 1989). One way is to stipulate that an event occurred and the observer's task is to determine of which type the event was, e.g.: "What was the word on the screen?" This way of questioning puts the focus of the observer's mental search on finding a suitable property for an answer. A predication implicit representation of the perceived property will serve that purpose.

A different way of phrasing the question is to stipulate a particular event type, e.g., the occurrence of a word, and the observer's task is to decide whether such an occurrence took place or not, i.e., to judge the existence or occurrence of a word. Marcel's (1983a, Experiment 1) question whether a word (any word) was present or absent to determine the detection threshold appears to be of that kind. Here the observers had to judge whether the occurrence of a word took place or not. Such a judgement would require a predication-explicit representation of the perceived event. A mere representation of the property 'word' without explicit predication to the observed event would not provide a natural answer to the observer's mental search initiated by the presence-absence question. Interestingly, several studies inspired by and attempts to replicate Marcel's work used the other approach for determining the detection threshold, i.e., "Which colour word was it (one of four possible colours)?" (Cheeseman & Merikle, 1984) or "Was there a word or a blank?" (Dagenbach, et al., 1989). In this case a predication implicit representation of the event type ("red" or "word" or "blank") provides an answer to the mental search. This may be one reason why these studies had only partial success in replicating Marcel's original finding that detection (absence-presence) has a higher threshold (i.e. occurs at a longer stimulus onset asynchrony, SOA, between stimulus and mask) than graphic or semantic similarity judgements (also see Fowler, Wolford, Slade, & Tassinary, 1981).

Finally there is also the possibility of formulating a direct test by referring to the target event as a perceptually experienced event: "What was the word that you just saw? ". For the observer to give a bona fide answer the stimulus event needs to be encoded explicitly as a visually perceived event. Without that encoding the observer can but answer "I didn't see anything". [12]Since reflection on one's state of seeing is required, this detection criterion corresponds to the "subjective threshold" introduced by Cheesman and Merikle (1984, 1986; see also Merikle, 1992); i.e the point at which participants know they know what they saw.

The purpose of this discussion was mainly to show that the known problems in this field can be formulated in our framework. The contamination of explicit (direct) tests through implicit knowledge and of implicit (indirect) tests by explicit knowledge has been debated particularly intensively in memory research. Jacoby (1991) proposed as a solution his process dissociation procedure which brings in conscious voluntary control as an arbiter. We will discuss the relation between the implicit-explicit distinction and consciousness and volition in the next two sections.

The notions of procedural and declarative knowledge have been brought into contact with the implicit-explicit distinction by several authors. For instance Karmiloff-Smith (1986, 1992) characterized implicit knowledge as procedural that is severely limited in its accessibility to other parts of the system. Accessibility has been emphasised as the central issue in the distinction between procedural and declarative knowledge by Kirsh (1991). Squire (e.g., 1992) characterized the knowledge of the past that is typically impaired in amnesics as declarative memory (where declarative is considered largely a terminological variant of explicit memory or knowing that) and contrasts this to nondeclarative (implicit, knowing how) memory that includes procedural memory (habits, skills and conditioned reactions) but also memory of facts revealed by priming.

Now our suggestion is that at least four different dimensions: knowledge contained in a procedure vs. knowledge not in a procedure, declarative vs. nondeclarative, accessibility, and implicit vs. explicit, are in play that need to be kept conceptually distinct. However, the goal is to show that there are some necessary relations between these dimensions and the types of knowledge form natural clusters: procedural knowledge tends to be implicit and, therefore, inaccessible, whereas declarative knowledge involves quite explicit representation of its content, tends therefore to be conscious and accessible for different uses.

To some, implicit knowledge may simply mean inaccessibility. Apart from being an arbitrary conceptual stipulation this definition of implicitness also lacks precision. Inaccessible in what way? All knowledge has to be accessible in some way or else it would not qualify as knowledge (on views like those of Millikan, 1984; Dretske, 1988) and, in any case, there would be no evidence that there was any knowledge at all. Our framework indicates how the implicitness of different aspects of knowledge makes the knowledge inaccessible in different ways, as indicated in our discussion in section 3.2 on direct and indirect tests and verbalizability, and in our treatment of procedural knowledge, which we now discuss.

The distinction between procedural-declarative knowledge was introduced in artificial intelligence (McCarthy & Hayes, 1969; Winograd, 1975) and later taken over into psychological modelling by Anderson (e.g., 1976). It concerned how to best implement knowledge: Should one represent the knowledge that every man is mortal as a general declaration "for every individual it is true that if that individual is human it is also mortal". The prime use of this general information would be to be consulted whenever knowledge of a human individual is introduced in the data base to then infer by general logical inference rules that this individual must also be mortal. The alternative is to have a specialised inference procedure: "Whenever an individual is introduced that is human then represent that this individual is mortal." [13]

Now we can see in what sense declarative knowledge is explicit. It represents explicitly that the regularity of 'human then mortal' is predicated to individuals and its generality of applying to every individual is also marked. Moreover, (provided the data base provides the required expressive power) it states that this regularity is a fact. In contrast, the procedure that adds 'is mortal' to every human individual it encounters, also knows something about this regularity but its knowledge is implicit in its application; its generality is implicit in the fact that it is applied to every encountered individual. But there is no distinction made in the system that represents that it is applied to individuals and that it is applied to every individual. The analysis also brings out the intuitive meaning of declarative knowledge as knowledge that declares what is the case (e.g., Squire, 1992, p 204: memory whose content can be declared) because it represents explicitly that something is a fact. Nondeclarative memory can be given precision in our analysis either as the stronger form of knowledge that does not make predication explicit or as a weaker form of knowledge that makes predication explicit but leaves factivity implicit.

The implicit nature of procedural knowledge also makes clear why it has limited accessibility. For instance the implicit nature of the procedural representation of the fact that all humans are mortal, does not allow the distinction between whether this rule applies to a current case and my thinking about the rule. For, the only internal distinction available is whether the rule is being activated or dormant. It being activated can represent that there is a current case to which it applies OR that one is thinking (deliberating) the rule. In order to separate these two cases one needs some internal distinction that (explicitly) represents whether the application of the rule applies or not. Only then can one distinguish whether one is just thinking about the rule without it actually applying, or whether one is thinking about it because it applies. This distinction, in turn, is a prerequisite for hypothetical reasoning. Moreover, there is no way to check on the adequacy of procedural knowledge. Such a check requires explicit representation of factivity in order to represent the result of the inference as a hypothetical possibility which is then compared with other available evidence. [14]Hence without the possibility of explicitly representing whether something is a fact or not, one cannot engage with procedural knowledge in hypothetical reasoning and planning or check on it's validity. This puts a severe limitation on the usability of procedural knowledge.

The advantage of procedural knowledge is its efficiency. Procedures need not search a large database since the knowledge is contained in the procedures. Knowledge that resides in the application of a procedure, as we have seen, leaves predication and factivity implicit. As a result it is limited in its accessibility in a way that has been claimed for modularity (Fodor, 1983), e.g., modular knowledge only applies to a specific input modality, cannot use knowledge from other domains, etc. Implicitness of procedural knowledge is, therefore a natural source of modularity in--as originally proposed--our input modalities that do not require fact explicit representation (as we will argue in detail for visually guided action later). In that context modular knowledge can be called implicit. However, implicitness is a less natural ally of modularity in case of central processes (Fodor, 1987, "modularity gone mad").

Modularity or quasi-modularity of central, conceptual processes has been proposed, for instance, by Cosmides (1989) for reasoning processes that use a cheating detector module. Sperber (1996) considers quasi-modularity as general feature of central cognition. Smith and Tsimpli (1995, ch. 5) posited a quasi-modular central language module to explain the highly developed insular foreign language ability in an otherwise handicapped individual. The stipulated central language module is not the same as the usual linguistic input processing module, since it is not used to converse in different languages, but to playfully translate from one language into another. Such central modules are unlikely to operate purely procedurally without explicit predication or factivity. This is very clear in the proposal by Leslie (1987, 1994) of a theory of mind module to explain the relative ease and speed with which children develop a theory of mind. Since a theory of mind does not just process factual information but has to represent the content of people's beliefs and desires, explicit representation of factivity is tantamount. Clearly, modular knowledge in this sense cannot be implicit as defined in this paper. [15]

In sum, knowledge contained in the application of a procedure (procedural knowledge) is active and efficient knowledge, but it leaves predication and factivity implicit, hence it is nondeclarative and limited in its range of applicability (hypothetical reasoning, checking validity) and far from being accessible to consciousness. In contrast, knowledge that states its predication and factivity explicitly cannot be contained in the use of a procedure. It thus loses efficency but becomes more flexible, to be used in hypothetical reasoning, evaluation of truth, and conscious awareness. The distinction between procedural knowledge and declarative knowledge provides a good basis for understanding why voluntary control of action is tied to explicitness and consciousness.

The dominant view in philosophy of what differentiates our intended actions, for which we are responsible, from other movements is that actions must be caused by our desires and beliefs (Davidson, 1963). Heyes & Dickinson (1993) in pursuit of the question whether animals act or just respond, argued that intentional action--unlike responses--must be based on an understanding of why one does them, i.e., one has to represent the goal one pursues and that the action leads to that goal. Searle (1983) even argued that intentional action must be causally self referential, i.e., one has to intend that the action be caused by one's intention.

A useful model for pursuing this phenomenal distinction between automatic (responses) and controlled, or willed action is that of Norman and Shallice (1980). It distinguishes two levels of control. There are the horizontal strands that operate at the level of implementing schemas which consist of complex conditional action tendencies (productions like in Anderson's, 1976, ACT model) with automatic control through activation by triggering stimuli and mutual inhibition of simultaneously triggered schemas ( contention scheduling ). The vertical strands of control come from the supervisory attentional system (SAS, a close relative of the central executive, Baddeley, 1986). The two control systems are supposed to capture on the one hand the phenomenal distinction between automatic responses and intentional action and on the other hand explain why a particular set of actions becomes difficult for patients with problems of voluntary control (e.g., patients with frontal lobe insult). These "SAS tasks" are typically (1) the setting up of new action schemas upon task instructions, (2) monitoring of novel or dangerous actions, or (3) the inhibition or monitoring of interfering existing action schemas.

Action schemas or productions are complex versions of responses to stimuli. They incorporate procedural knowledge about event contingencies in the world that (as discussed in 3.3) leave predication and factivity of these regularities to instances implicit in their application. The stimuli that trigger them can be declarative, or nondeclarative representations of features of the environment or internal states. The control exerted at the level of contention scheduling as well as that exerted by the SAS is in terms of boosting or inhibiting the activation of schemas. For instance, in order to ensure that a single schema produces coherent action the dominant schema might get its activation boosted even further at the cost of the activation of less dominant schemas.

Our claim is that contention scheduling directs this control purely on the basis of the schemas as representational vehicles (the amount of activation is a feature of the schema as vehicle not of its representational content). In contrast, the SAS directs its control on the basis of the schemas' representational content. In support of this contention one can show that such content oriented control is necessary for the 'SAS tasks' listed by Norman and Shallice. For instance, in a version of the Wisconsin Card Sorting test for children a three year old child (like a frontal lobe patient) who has learned to sort cards by colour, has now to sort the same cards according to a new rule, e.g., the shape of symbols on the card. Without SAS the once learned colour sorting rule is dominant and will suppress execution of the new rule. Three year old children, even though the child knows the new rule and can verbally state it will perseverate by sorting according to the old rule (Zelazo, et al., 1995), like frontal lobe patients tend to do on the traditional test (Shallice, 1988). If the SAS to be of use here, it has to boost the new schema and inhibit the old, dominant schema. But this cannot be done on the basis of vehicle features like amount of existing activation or strength (too many weak schemas would be boosted) but the SAS has to be able to address the new schema by its content, i.e., that stimulus-response sequence that the new rule requires (see Perner, 1998, for discussion of other SAS tasks).

Control of schemas via their content requires representation of that content. In order to avoid confusion, this content has to be explicitly marked as not being factual (i.e., explicit representation of factivity), but being something that is desired or intended (explicit representation of attitude). This means that the SAS must be (or contain) a second-order mental state (one that represents desires) which is the important prerequisite (or even sufficient condition) for being a conscious state according to the higher-order thought theory of consciousness (see 3.1). So, this analysis suggests, that the need to represent content and attitude explicitly distinguishes controlled or willed action from automatic action. We can identify intentional action with action (be it automatic or willed) that is in line with the explicit representations of the SAS (it is under control). If automatic action contravenes those representations then it is experienced as an unintentional lapse or "slip of action" (Reason & Mycielska, 1982). The analysis also makes clear why willed action is conscious--because it is based on a second order mental state. And with this we have a theoretical justification why in the quite different areas of research on implicit memory and subliminal perception voluntary control is used as a criterion for consciousness. Note that, however, not all aspects of the content of a schema have to be explicitly represented to allow control by the SAS; only sufficient aspects to indicate that the action of the schema is desired. Only those aspects of the content which are explicitly represented will be conscious; the remaining aspects may in principle embody knowledge which the person is not aware of having, and whose details of application they could not control. Our argument requires a conscious representation to be made by the SAS (e.g. `I want that I play Fur Elise on the piano'), but the overlap in content between this representation and a body of knowledge (e.g. about piano playing) could allow that knowledge to apply, even if the factivity of the knowledge is not explicitly represented; that is, a fully explicit representation in the SAS can co-exist with implicit representations in a knowledge base. We will see an example of this in section 4.4 below.

Jacoby's (1991) process dissociation procedure uses voluntary control of knowledge in order to provide better estimates of implicit (unconscious) or explicit (conscious) memory. The procedure can be used not only for memory but also for, e.g., subliminally presented information (Debner & Jacoby, 1994). One critical part of this procedure is the exclusion condition, in which participants in an indirect test of memory (e.g., to complete word stems) are instructed to not use words that were presented in a list. Unconscious knowledge, in particular, knowledge that leaves predication implicit (e.g., the word form "butter" of the word that was on the learning list) can influence the indirect test and escapes exclusion in the exclusion test, since the word form does not fall under the description "word on that list". So, the number of words from this list that are, despite instructions, used as an answer is a better indicator of implicit memory than performance on the indirect test without exclusion instruction, since on the indirect test there is no control for participants using words that they can remember explicitly.

[16]

Our analysis of the different aspects of knowledge that are represented explicitly and those that are left implicit provides a basis for relating different criteria that have been brought into contact with the implicit-explicit distinction. Knowledge that represents its content, its attitude, and its holder (self) explicitly is on the higher-order thought theory conscious. Explicit representation of factivity might be sufficient, since from being a fact knowledge can be inferred. Explicit representation of predication (and often of factivity) is required for being able to refer in verbal communication and thus a link emerges between direct tests (where reference is made to the known fact) and explicitness and consciousness. Similarly, procedural knowledge leaves predication implicit in its application. Therefore it remains unconscious. Declarative knowledge represents predication and factivity explicitly and thus qualifies for conscious access. Automatic action is based on schemas (productions) that, like procedural knowledge, leave predication implicit, while controlled action (SAS) represents the content of these schemas explicitly together with the attitude. Willed action is therefore conscious while automatic action can remain unconscious. This justifies the use of voluntary control to help distinguish conscious from unconscious elements in task performance.

Visual information is not processed in a unitary way. At least two functionally different systems exist. Traditionally it was thought that the functions were for perception of objects and perception of the spatial relations between these objects ('What' versus 'where', Ungerleider & Mishkin, 1982). Recently, Milner & Goodale (1995) have moved from a distinction in terms of encoding different aspects of the visual array to reconceptualising the distinction in terms of the system's purpose of either forming a perceptual representation ('what' there is) or exerting visuo-motor control ('how' to act). This reconceptualisation has been prompted in large part by functional dissociations in brain injured patients and normal people (e.g., Milner & Goodale, 1995; Rossetti, 1997). As one example we describe a series of experiments by Bruce Bridgeman on the induced Roelofs effect.

Bridgeman (1991, Bridgeman, Peery & Anand, 1997) reports that for human observers a stationary dot within a rectangular frame appears to move opposite to a movement of the frame. After a brief exposure to this apparent movement the display vanished and the observer had to either indicate verbally at which of five marked locations the dot had been after the movement or to point to the location of the dot. In their verbal responses all observers were susceptible to the illusion and reported the dot's last location as having moved opposite the frame's movement. In contrast, only half the observers were susceptible to the illusion in their pointings, the other half pointed quite accurately to the dot's actual location. Bridgeman interprets the results as showing the dissociation between a cognitive (perceptual) system used for verbal report and a system for visuo-motor control that steers the pointing finger.

This interpretation can be refined within our conceptual framework. Visually guided behaviour can be procedural and nondeclarative, i.e., it doesn't need to explicitly represent a distinction between facts and non-facts. It is a system that registers object (features) in egocentric space and everything which is represented is a fact. An interesting question is whether predication needs to be represented explicitly. It seems that the object that one grasps does not need to be represented as a re-identifiable individual. Representation of its visible features suffices [17] as Campbell's (1993) analysis shows that orienting oneself in relation to landmarks can be done within a pure feature placing system without the necessity of conceptualising the landmarks as physical objects that have these features. So, no predication of the visible features to the objects that have them needs to be represented. However, this still leaves the question of whether the visible object features need to be predicated to the spatial positions, i.e., "dot-ness in position x, y, z" which amounts to predication of the feature 'dot-ness' to that position. Or is it sufficient to simply have a conjunction of feature and position? A plausible answer might be that a mere conjunction is sufficient if only a single object needs to be tracked. Then the predication of feature to position can remain implicit in the tracking. For keeping the position of a second feature in mind while tracking the first, explicit predication is required. We know of no data that speak to this issue [18]but the question of whether visually guided action leaves only factivity and time or also predication implicit is testable.

In contrast to visually guided behaviour, to give a verbal response is to make a judgement, that that's where the dot really is. The information in this system needs to explicitly represent predication and factivity. Since these are preconditions for consciousness, this explains why the information used for the verbal response is what is consciously experienced. The analysis also makes clear a certain ambiguity in the pointing condition. Pointing is on the one hand a movement of the finger to the target (a visually guided movement), on the other hand it is a declarative act that states what is the case. The bimodal distribution could be due to this ambiguity. From our analysis it follows that if the instructions are not to point but to move one's finger to touch the dot, then no observer should be susceptible to the Roelofs effect. Bridgeman (personal communication) carried out this condition and obtained the predicted results.

Bridgeman's experiment also illustrates the other interesting parameter of the visuo-motor system that its information persists only for a few seconds. When the response is delayed for 8 seconds then all observers show the Roelofs effect just as in their verbal response ( and this also holds for the condition where observers had to move their finger to the target, Bridgeman, personal communication). Representations that do not mark factivity and time are only useful to represent the here and now, since they do not differentiate what is the fact (here and now), what is not a fact but a mere hypothetical assumption, or what was a fact but isn't any more (see Perner, 1991, for developmental convergence of the abilities to represent hypothetical scenarios and represent change over time). So, because the visuo-motor system leaves time and factivity implicit, it can only update its information about the current state of the environment but not keep track of past state of affairs and compare them with the present state of affairs. For this factivity and time need to be represented explicitly (see alsoWong and Mack 1981).

In sum, what these results demonstrate is that there are two visual information processing systems. One is identified neurophysiologically with the dorsal path from the primary visual cortex (V1) to the posterior parietal cortex (Milner & Goodale, 1995). Its information is unconscious, it cannot be used for statements (verbal or gestural) about the world, it is not susceptible to certain illusions and is used for action in the world but is of limited duration. Our interpretation is that this system leaves factivity and time implicit (and perhaps also predication--see above). The other system is identified with the ventral path from V1 to the inferotemporal cortex. It's information is conscious, susceptible to illusions, it is used for statements about the perceived world, and is used for action in the world after some delay. Our interpretation is that this system represents predication and factivity explicitly and, thus, makes its content accessible to consciousness. (see alsoAglioti, DeSouza, and Goodale, 1995, Gentilucci, Chieffi, and Daprati , 1995, Milner & Goodale, 1995, chapter 6; Rossetti, 1997).

Also the spared capacities in blind-sight and numb-sense patients (tactile analogue to blind-sight, Paillard, et al., 1983) depend on similar parametric variations. For instance, Marcel (1993) reported that blindsight patient G.Y. was better able to detect an illumination change in the blind field when the response was made quickly than when it was delayed by 2 or 8 seconds, when the response consisted of an eye blink (interpretable as a nondeclarative response) than a verbal "yes-no" (a declarative comment), and when the patient was invited to guess than when instructed to give a firm judgement (where bona fide responses require judgement explicit representation). Marcel also found that people of normal vision responded to near-threshold changes in illumination in the same way as blindsight patients. That is, in people with normal vision, detection was better when responses consisted of an eye blink rather than a "yes-no" verbal response, and when people were invited to guess rather than make a firm judgement.

A particularly interesting point about the last result is that the response shift from judgement to the guessing condition consisted not of a criterion shift to saying "seen" more often, but of an increase in discrimination accuracy (increase in hit rate and decrease in false alarm rate). A shift in criterion towards "seen" responses would be expected if the stimulus was encoded explicitly as a fact about which one is uncertain in one's judgement. Then being given leave to guess would simply lower the rejection criterion resulting in an increase in the willingness to say "yes". In contrast, when a stimulus is encoded fact implicitly, there is a representation "illumination change" but no information as to whether it occurred or did not occur, or whether it occurred on the current or an earlier trial. Thus there is no proper information for a judgement (hence low detection accuracy). With leave to guess, however, one is free to let oneself be influenced by the fact-implicit information that happens to be correct, which results in higher detection accuracy.

Memory has many different facets. To help focus our discussion we distinguish the wider use of memory as the availability of information acquired in the past (e.g., remembering/ still knowing that 2x2=4) from the narrower meaning of memory as availability of information about events in the past acquired in the past. As a concrete example we use the typical memory experiment in which one is read a list of words, among them the word "butter", and we look at the consequences if various aspects of this event are being represented explicitly or left implicit. The consequences we consider are in terms of memorial state of awareness, retrieval volition, and test responses.

As the first possibility we consider strong implicitness. At learning, the word "butter", designed to represent the fact that "the word `butter' occurred on the list" is stored so that only the word form "butter" is represented explicitly and all the rest is left implicit. This supports no particular memorial state of awareness. It could support a 'feeling of familiarity', if that word had been encountered the first time on that list. This representation cannot be voluntarily accessed, and not used bona fide in any direct test, since no reference to any particular occurrence can be made. It can, however, influence indirect tests. The mere presence of the word form "butter" can for instance enhance the likelihood of answering with "butter" to the request to list dairy products. It could also account for participants including `butter' on an exclusion test without any accompanying feelings of familiarity (Richardson-Klavehn, Gardiner, & Java, 1994).

It is also likely that there are cases where it is not just the word form "butter" that has been represented, but also the perceptual details by which that word form was perceived. That is, a representation of the conjunction of various contextual features is formed, but this feature-complex need not be predicated as having occurred on the list. Such a representation could enhance perceptual identification and produce familiarity effects without supporting recollection (e.g. Jacoby & Dallas, 1981). Such a representation could also be involved in the "mere exposure effect" in which exposure to a stimulus, for example a novel shape, can lead to high affect ratings for the stimulus in the absence of recollection of having seen it before (Zajonc, 1968; Bornstein, 1989; Gewei and van-Raaij, 1997).

When the occurrence of the word "butter" is explicitly predicated, i.e., "the word 'butter' occurring on that list", then it can come under direct voluntary control since now reference to the particular event of being on the list is possible. As a consequence, performance on a direct test can be better than on an indirect test (Reingold and Merikle's, 1993, control for differences in test sensitivity). However, voluntary control remains as an educated guess and does not result from a considered judgement, since the occurrence is not represented as a fact.

Explicit representation of the occurrence as a fact, makes the event accessible under the description of being a fact and participants can now give a considered judgement that the word "butter" is part of that list. With explicit representation of time, participants can then also give a considered judgement that "butter" occurred at a particular reading of the list in the past. They can experience memory of a past event. It can be a conscious experience of memory of the past according to the higher-order-thought theory, since explicit representation of factivity entails a higher order thought about one's knowledge. However, even with such a representation participants may remember no details of seeing/hearing the item.

An important next step comes with explicit representation of the experiential source of one's knowledge: `I know that "butter" was on the list because I saw it there'. Only such encoding -- encoding of having been in direct contact with the known event -- constitutes genuine episodic memory according to Tulving (1985; Perner, 1991). [19] Tulving (1985; and later others, such as Gardiner, 1988) distinguished two types of recognition responses: Those accompanied by simply an experience of Knowing that the item occurred earlier in the context of the experiment ("K" responses); and those based on truly Remembering the prior experience of the item ("R" responses).

"K" responses may arise for various reasons, e.g., because the word form `butter' is encoded predication implicitly and simply comes to mind readily (whether the participant does give a positive recognition response depends on his theory of why the word came to mind) or because a predication explicit representation has been formed and so the participant guesses that the word had been on the list. In both cases, the participant may give a "K" response with low confidence. On the other hand, if the participant experiences strong familiarity when he comes across the word "butter" he may give a "K" response with strong confidence. However, in all these cases there is no genuine knowing that "butter" was on the list just guesses that carried more or less conviction. Researchers in the field (Conway et al, 1997) have now started to give participants also a choice between "K" responses and "guesses". This may separate predication and fact implicit knowledge from knowledge that represents factivity (and past-ness) of the event in question explicitly. Unlike "guesses", "K" responses should not be just produced but be produced as the reflection of a fact."R" responses differ from "K" responses in that they need not only be seen reflecting facts but also as products of one's direct experience.

Table 2 summarises the different levels of explicitness, which memorial state of awareness, voluntary control and kind of test performance they support. Our analysis yields distinctions that reassuringly map onto distinctions that have emerged from the empirical literature. In particular, it can address the distinction between retrieval volition and memorial state of awareness (Richardson-Klavehn, Gardiner & Java, 1996; Schacter, Bowers, & Booker, 1989), it honours the distinction between "implicit" memory and the distinction between "know" and "remember" judgements as two kinds of explicit memory in the spirit of Tulvings (1985) original distinction, where "know" judgements are supposed to cover 'knowledge of the past' and "remember" judgements memories of experienced events as experienced (Perner, 1990). This analysis indicates that both "R" and "K" count as declarative knowledge (both involve explicit predication) and familiarity can be purely procedural (predication left implicit).

Laid down representation of fact that Fb	Memorial state of awareness	Retrieval volition	Reference by:	Recognition test response
Property	"F"	none	involuntary	nothing	correct guess.
Compound	"F-X"	feel of famil.	--"--	nothing	recogn. by famil.
Predication	Fb	--"--	direct vol.	"part of list"	--"--
Factivity + Time	"Fb happened"	knowing past	--"--	"was on list"	"K" (past event)
Origin	"I experienced Fb"	remembering	--"--	"remember!"	"R"

The thrust of our framework is that there is not a simple dichotomy between implicit and explicit knowledge. This owes much to Karmiloff-Smith's (1986, 1992) insistence that the basic dichotomy should be embellished by further levels of explicitness. It is reassuring that our framework that logically unfolds from the conceptual analysis of knowledge yields a plausible correspondence to Karmiloff Smith's empirically motivated classification. Her initial level (I) of implicit knowledge where the information is only in the system maps onto procedural knowledge that leaves predication implicit. Her first level (E1) of explicit knowledge results from a redescription of the original information encoded in procedural format, so that the information becomes information to the system, useable by different parts of the system. This maps onto knowledge that makes predication explicit (thus can be referenced felxibly by different user systems) but leaves factivity implicit. At the next level of explicitation (E2) the knowledge becomes conscious, and at the final level (E3) also verbally expressible. The once clear progression from E2 to E3, has later been collapsed into a level E2/3 (1992, p. 23) due to the lack of a clear empirical demonstration of such a progression. The level E2/3 corresponds to knowledge that makes factivity (and source) explicit. Moreover, since explicit factivity tends to make knowledge conscious and verbally accessible our analysis actually suggests the merging of the original levels 2 and 3.

Whereas Karmiloff-Smith's research emphasises how implicit knowledge becomes increasingly explicit with development, also dissociations between two competing knowledge bases have been found -- reminiscent of the dissociations in visual perception (e.g. Diamond & Goldman-Rakic, 1989; Goldin-Meadow, Alibali, and Church, 1993; Clements & Perner, 1994). Goldin-Meadow et al review studies that show that, for example, the acquisition of concepts of quantity (Piaget & Inhelder, 1974/ 41) can be more advanced in children's gestural comments than in their verbal responses. One of the interpretations of this finding was (Church & Goldin-Meadow, 1986) that the multidimensional spatial medium of hand gesture makes it easier to express novel ideas than the unidimensional temporal medium of linguistic expression. However, one can think of the gestures as spontaneous (mostly unconscious) concomitants of the thinking process. In that case the earlier emergence of advanced knowledge might be the sign of thoughts about reality that have not yet been recognised as being about reality (factivity implicit). This interpretation fits a parallel finding in children's developing "theory of mind".

Clements and Perner (1994) reported that understanding of false belief emerges in children's visual orienting responses as early as 2 years and 11 months, a year earlier than in their verbal responses to questions. Children are told enacted stories in which the protagonist does not see how his desired object is unexpectedly transferred from one (A) to another location (B). Children in the interesting period around 3 years of age answer the question about where the protagonist will go to get his object wrongly by pointing to the current location of the object. However a majority of these children look (visual orienting responses) in anticipation of the protagonist at the empty location where the protagonist mistakenly thinks the object is.

Further research (Clements & Perner, 1996) indicates a remarkable similarity to the dissociations observed between the two visual systems (see Section 4.1). When instructed to move a welcoming mat for the mistaken story protagonist who was on his way to get his object, then children who move the mat spontaneously tend to move it correctly to where he thinks the object is (A), while children who need prompting (thus with some delay) move it to where the object is (B). We see, there seems to be a stage in children's developing understanding of belief where two different knowledge bases dissociate. One of them is a more accurate, and developmentally advanced knowledge base (in analogy to the dorsal visual path) that supports only non-declarative action (looking and moving a mat) that is carried out without delay (spontaneous mat move) while a less accurate and less developmentally advanced knowledge base (analogous to the ventral visual path) is used for declarative responses (verbal and pointing) and delayed action (prompted mat movings). We do not know, of course, whether the more advanced knowledge is conscious and the other unconscious, since one cannot ask 3 year old children to report on such a distinction but otherwise the similarities are remarkable.

Such a similarity between dissociations in processing visual information about the environment and understanding another person's false belief suggests that the characteristics of the two types of knowledge are not primarily determined by the brain regions in which the information is processed (dorsal vs. ventral path) but by more general functional differences that apply to visual information processing as well as a theory of mind. Our analysis shows how these functional distinctions could arise from which aspects of knowledge are represented explicitly. An interesting speculation about functional differences in the theory of mind case is, that the explicit understanding comes with (something of) a real theory, i.e., a causal understanding of belief formation and how belief determines action. Whereas, the implicit understanding of where the protagonist will go may be based on abstraction of situational regularities. Within our framework this assumption gives a quite coherent picture of the existing data and leading to new, testable predictions (Perner & Clements, in press).

One can learn that certain events tend to go together and form a typical sequence. Such filtering of statistical patterns of possible combinations does not need representation of individual events and inferences from individual events to all possible events. Rather it is a process of pattern formation and recognition for which connectionist systems are good (e.g., to classify different feature patterns into letters, e.g., Bechtel & Abrahamsen, 1991). The encountered combinations of letters in artificial grammar tasks have a similar effect and can be particularly well modelled by connectionist networks (Dienes, 1992). Although individual instances shape the connections between units and, hence, the association between the properties that these units represent, there is no representation of the individual instances. [20]Connectionist work also shows that such pattern generalisation leads to pattern completion. If many elements of a typical pattern are present then the network tends to generate representations of the missing bits. This is important, because such pattern completion processes can produce expectations of what is to come on the basis of what has so far happened. And the, for us, important implication is that such associative expectation is possible without explicit predication.

This makes it possible to anticipate correctly where the protagonist will go to get the desired object in our false belief stories without explicit predication to a particular occasion, i.e., without representing that he will go there. So, according to our above discussion, such a representation of the mere event form 'protagonist going to location A' and hence, 'protagonist at location A' as part of a pattern completion process, can guide visual orienting responses and spontaneous actions because such a representation can trigger an existing action schema waiting to be executed. It cannot be used for communication because it lacks predication to an individual event which can be re-identified across mental spaces explicitly marked as, e.g., "facts", "anticipation", or "verbal description." It cannot sustain uncertainty, since it does not support a self-reassuring check about where the protagonist will come down since without explicit predication there is no representation stating that he will go anywhere. And that is the pattern of results we observed in the precociously correct responses: they were high only in spontaneous action and visual orienting responses.

In contrast, a theory of belief goes beyond mere generalisations of observed regularities and constitutes genuine causal understanding of the underlying processes (see Gopnik, 1993; Perner, 1991, for indications of theory use). Causal understanding cannot be achieved by mere pattern matching and pattern completion but must employ explicit predication since causal reasoning is counterfactual supporting (Lewis, 1986; Salmon, 1984). Counterfactual support means that one understands that if the conditions were different then the result would be different, and such reasoning requires different mental spaces for contrasting the actual facts with their counterfactual oppositions. For these reasons, responses that are based on a causal theory of belief should also be accessible to communication (answers to questions) and be robust against doubt (hesitating action).

On the basis of this reasoning one can predict that implicit knowledge should be primarily shown in the situation described above, where the correct response can be based on situational, behavioural regularities, such as "people look for objects where they last put them, where they last saw them, where they told someone to put it, etc.". In the traditional scenario all these regularities -- if they apply -- point to the same, correct answer "A". In a variant scenario (Perner, Leekam & Wimmer, 1987) the protagonist, who has put the object into B, tells a friend to move the object from B to A, but the friend forgets. Here, behavioural regularities give different predictions. "Last seen" or "where put" indicate location B while "told to put" indicates correctly A. Hence signs of implicit understanding should be hampered in this scenario. Indeed, Clements (1995, Chapter 5) reports that children show fewer orienting responses to location A than in the traditional scenario. In contrast, their verbal responses show little difference in the two scenarios, replicating the original result by Perner, et al. (1987). This is to be expected if explicit responding is based on a causal understanding of belief formation.

Another prediction is that verbal explanations of why the protagonist believes the object is still in location A (in the original scenario) in contrast to observing behavioural regularities (seeing the protagonist look for the object in A) should affect implicit and explicit understanding differently. Causal explanations should primarily affect explicit understanding while observation of regularities should have a stronger effect on implicit understanding. The part for explicit understanding of this prediction has been tested. Clements, Rustin & McCallum (1997) report that causal explanations affect verbal responses but observation of regularities does not. The corresponding data on visual orienting responses or action responses are still outstanding.

Our framework also elucidates the different ways in which knowledge can be implicit in the standard implicit learning paradigms. The paradigm explored most thoroughly in the implicit learning literature is artificial grammar learning (see Reber, 1989, and Berry, 1997, for overviews). In a typical study, participants first memorize grammatical strings of letters generated by a finite-state grammar. Then, they are informed of the existence of the complex set of rules that constrains letter order (but not what they are), and are asked to classify grammatical and nongrammatical strings. In an initial study, Reber (1967) found that the more strings participants had attempted to memorize, the easier it was to memorize novel grammatical strings, indicating that they had learned to utilize the structure of the grammar. Participants could also classify novel strings significantly above chance (69%, where chance is 50%). This basic finding has now been replicated many times. So participants clearly acquire some knowledge of the grammar under these incidental learning conditions. But is this knowledge implicit? We will now theoretically and empirically analyze the case of artificial grammar learning in terms of the different aspects of being a fact or being knowledge that can be made explicit, or left implicit, according to our previous analyses. (See also Dienes and Perner, 1996, who explore whether participants represent the property structure of a grammar implicitly or explictly, an issue not dealt with in the following.)

When participants learn the structure of an artificial grammar by exposure to the exemplars, they may not explicitly represent the particular grammar to which the properties are predicated. Consider a person who uses the mental rule that "M can be followed by T". This statement represents the fact that, according to the grammar one was trained on 10 minutes ago, M can be followed by a T . Yet, the fact that it is a particular grammar which has this property is not explicitly represented since there is nothing in the expression "M can be followed by T" whose function it is to covary with that fact. This fact can be made explicit by forming the mental expression: " g has the property that M can be followed by a T", where g denotes a particular grammar, e.g., the grammar that I was just being trained on. The critical feature here is that different properties, like "me having just been trained on" and "being a grammar in which M can be followed by T" can both be predicated to g. This extended expression makes the implicit predication of 'M is followed by a T' to a particular grammar explicit, because the whole expression does have the function to covary with the fact that the identified particular grammar is characterized by the property in question.

Whether participants represent the individual grammars and the predication relationship explicitly can be revealed by the volitional control participants have over the application of their knowledge. Consider a test of volitional control given to participants by Dienes, Altmann, Kwan, and Goode (1995). Participants were given 7 minutes to try to memorize exemplars generated by one grammar, and then another 7 minutes to try to memorize exemplars involving the same 6 letters generated by a second grammar. Participants were then informed that two grammars were involved, and given a test set in which a third of the items followed the first grammar (but not the second, e.g.: xmxrtvtm), a third followed the second grammar (but not the first, e.g.: xmvrxrm), and a third violated both grammars (e.g.: xmtvvxrm). Participants were asked to choose items that followed only one of the grammars; half the participants were asked to endorse only the items consistent with the first grammar, and the other half of the participants were asked to endorse only the items consistent with the second grammar. Participants were perfectly able to distinguish the grammars to the usual level of performance in such tasks and they showed no tendency to endorse the grammar they were asked to ignore. How could this performance be achieved?

One way to succeed in such a test is to have direct volitional control over one's knowledge, in the sense that one can decide to use or not to use IT because IT has been explicitly labelled as the particular body of knowledge one wishes to use or not use. That is, we assume that for direct control it is necessary to represent the individual grammar explicitly. But there are alternative ways of controlling which body of knowledge to use that does not require such explicitness. For example, Whittlesea and Dorken (1993) argued that participants could distinguish different grammars by familiarity. One account of the Dienes et al (1995) results along these lines is that the choice of grammar can be done by means of a compound property, e.g., in-context-A,-M-can-follow-T. Context A could be, for example, a particular time at which a string was studied . If context A is reinstated by task demands or imagination, the knowledge of a particular grammar can be isolated (through association) without having to explicitly predicate these properties to any particular grammar. Even though this scenario of indirect control over particular grammars without explicit representation of the grammar is often possible or even plausible, there may be situations in which one can plausibly decide that volitional control was actually mediated (at least partly) by explicitly representing the individual grammar. For example, if, with a sufficiently sensitive test, measures of familiarity (such as ratings, speed of stimulus identification) do not predict classification response, then these alternative scenarios (that do not represent the individual explicitly) are not supported. In fact, Buchner (1994) found that grammaticality judgements were not related to speed of identification. If this type of observation is supported, it follows from the volitional control experiments that participants do represent the individual grammar (and the predication relationship) explicitly. Of course, as we have mentioned previously (Section 2.1.3), the presence of knowledge in which the predication relationship is represented explicitly does not rule out the possiblity that there is in addition other knowledge about the same topic which is predication implicit.

To clarify how explicitly participants can reflect on their knowledge it is necessary to be clear about what piece of knowledge participants may be reflecting on (e.g., Shanks & StJohn's, 1994, information criterion). We distinguish two different domains of knowledge. The first domain we call grammar rules . These are the general rules of the grammar that the participant has induced; e.g. "M can be followed by T". The second domain pertains to the ability to make grammaticality judgements . This arises when the grammar rules are being applied to a particular string and pertains to the knowledge whether one can judge the grammaticality of the given test string independently of any knowledge that one knows the rules that one brings to bear for making this judgement.

Knowledge of artificial grammars and of natural language may differ with respect to these two types of domains. We seem to lack explicit knowledge of grammar rules for English (we can't represent any sort of attitude towards most rules of English grammar, so such rules are at least attitude implicit) as well as for the quickly acquired artificial grammars. In contrast, we are fully aware and have explicit knowledge of our ability to judge the grammaticality of English sentences but we may lack such explicit knowledge for the nonsense strings produced by an artificial grammar (and perhaps therefore in the early stages of learning a first or second language as well).

There are various possible relationships between the knowledge of rules and grammaticality judgements. Reber (e.g.1989) showed that people did not use the rules to respond deterministically; that is, when retested with the same string, participants often responded with a different answer. Extending this argument, Dienes, Kurz, Bernhaupt, and Perner (1997) argued the data best supported the claim that participants matched the probability of endorsing a string as grammatical to the extent to which the input string satisfied the learned grammatical constraints, and that this probability varied continuously between different strings. The result of learning is to increase the probability of saying grammatical to grammatical strings and decrease it for nongrammatical strings. As people begin to learn, the probabilities start to covary with probability of success, with higher probabilities for saying grammatical occurring for strings that actually are grammatical. This means that the probabilities actually imply the epistemological status of the grammaticality judgement ranging from a pure guess to reliable knowledge. The probabilities have the function to capture this information, since without this correlation the system would ipso facto not be successful, and the relevant learning mechanism would not have evolved. However, the mechanism responsible for producing these probabilities need not explicitly represent that there is knowledge i.e. that the representations induced by training and test have the properties given in section 2.1.2. For example, there is no need for the mechanism to represent that there is something that is taken as reflecting the accuracy of the judgements, nor that the accuracy of the judgements is well-founded in the learning history, nor that the self is the possessor of the knowledge.

Although participants' response probabilities suggest only a structure-implicit representation of the accuracy of their judgements, we do not know whether participants might, in fact, have a more explicit representation thereof. One method for determining if participants can explicitly represent the epistemic status of their judgements is to ask participants to state the confidence they have in making each classification decision, say for example on a scale which ranges from 'guess', through degrees of being 'somewhat confident', to 'know'. If the confidence rating increases with the probability of correctly responding to each item, with random responding given a confidence of 'guess', and deterministic responding given a confidence of 'know', then the propositional attitudes implied by the probabilities have been used by the participant to explicitly represent the epistemological status of the grammaticality judgements; if confidence ratings are not so related to response probabilities, then actual epistemological status has only been implicitly represented.

The above procedure only tests whether participants represent their ability to make judgements as knowledge. It is possible, like in the natural language case, that participants know when they have the knowledge to judge grammaticality and know when they are guessing, but still their knowledge of grammar rules is not represented as knowledge. This could be tested if we knew the actual content of participants' grammar rules. If the rules have been induced over time by some type of optimal learning rule, then the epistemological status of the rules must be greater than just guesses. If participants, despite stating rules freely, or endorsing presented rules, nevertheless, believe that they are just guessing, then the representations of the rules have not been appropriately represented as knowledge. Also, if the rules had not been represented as knowledge, they may not be offered as descriptions of the grammar when participants are asked, because participants would not know that they knew anything. Of course, failure to state the rules in free report could also arise for other methodological reasons due to the normal failings of free recall.

Establishing whether participants explicitly or implicitly represent their attitude of knowing towards their grammaticality judgments rather then grammar rules is methodologically easier, and the relevant research to date has focused on judgements. As noted above, one way to determine whether participants explicitly represent their ability to make judgements as knowledge would be to determine for each test item the probability with which it is given the correct response. If a plot of confidence against probability is monotonically increasing line going through (guess, 0.5) and (know, 1.0) then participants have fully used the implications of the source of their response probabilities to infer an explicit representation of their state of knowledge. If the line is horizontal, then the state of their knowledge is represented purely implicitly. If the line has some slope, but participants perform above chance when they believe that they are guessing, then some of the knowledge is attitude explicit and some of the knowledge is attitude implicit.

Typically in artificial grammar learning experiments, participants make one, or sometimes two, responses to each test item so it is not possible to plot the confidence-probability graph just described. In fact, it is not strictly necessary to plot the full graph. Lets take the case where the participant makes just one response to each test item. We divide the items up into those to which the participant makes a correct decision ('correct items') and those to which the participant makes an incorrect decision ('incorrect items'). If accuracy is correlated with confidence, then the correct items should be a selective sample of items given a higher confidence on average than the incorrect items. Conversely, if participants do not give a greater confidence to correct rather than incorrect items, then that is evidence that the slope of the graph is zero; i.e. participants do not represent their state of knowledge of their ability to judge correctly. If participants give a greater confidence rating to correct rather than incorrect items, then that is evidence of at least some explicitness. If in this case, participants perform above chance when they believe that they are literally guessing, then that is evidence of some implicitness in addition to the explicitness.

Note that the argument made in the previous paragraph presumes a certain theory of how participants apply their knowledge: namely, probabilistically, rather than deterministically (as we have mentioned); but also that the knowledge is largely valid. Consistently, Reber (1989) has argued that people's incidentally acquired knowledge of artificial grammars is almost entirely veridical. If people had deterministically applied partially valid rules, then there would be no difference between confidence in correct and incorrect decisions, regardless of whether the knowledge was attitude explicit. Thus, application of the procedure in different domains requires careful consideration of how knowledge is applied in that domain.

Chan (1992) was the first to provide data that tested whether participants explicitly represented their attitude of knowing towards their grammaticality judgements. Chan initially asked one group of participants (the incidentally trained participants) to memorize a set of grammatical exemplars. Then in a subsequent test phase, participants gave a confidence rating for their accuracy after each classification decision. Chan found that these participants were just as confident in their incorrect decisions as they were in their correct decisions, providing evidence that the attitude of knowing was represented only implicitly. He asked another group of participants (the intentionally trained participants) to search for rules in the training phase. For these participants, confidence was strongly related to accuracy in the test phase, indicating intentionally rather than incidentally trained participants more explicitly represented their attitude of knowing. Manza and Reber (1997), using stimuli different from Chan's, found that confidence was reliably higher for correct rather than incorrect decisions for incidentally trained participants. On the other hand, Dienes et al. (1995) replicated the lack of relationship between confidence and accuracy, but only under some conditions: the relationship was low particularly when strings were longer than three letters and presented individually. Finally, Dienes and Altmann (1997) found that when participants transferred their knowledge to a different domain, their confidence was not related to their accuracy.

In summary, there are conditions under which participants' represent their attitude of knowing towards grammaticality judgements implicitly on most judgements, but there is sometimes evidence of some judgements having an explicit attitude of knowingEven in the latter case, there is usually evidence of implicit knowledge: Both Dienes et al. (1995) and Dienes and Altmann (1997) found that even when participants believed that they were literally guessing, they were still classifying substantially above chance.

Dienes et al (1995) provided evidence that this type of implicit knowledge seemed to be qualitatively different to knowledge about which the participants had some confidence. When participants performed a secondary task (random number generation) during the test phase, the knowledge associated with 'guess' responses was unimpaired, but the knowledge associated with confident responses was impaired (to a level below that of the knowledge associated with 'guess' responses). That is, this criterion is not just another curious way of categorizing knowledge: It may separate knowledge in a way that corresponds to a real divide in nature.

In summary, when participants learn artificial grammars, there is evidence that for at least some of the acquired knowledge, participants represent the grammar to which the knowledge is predicated, and thus can exert intentional control over which body of knowledge to apply. This intentional control indicates, by our analysis in section 3.4, that the participants have conscious knowledge of some content predicated to that grammar - in particular, the content that they use to choose the grammar. But there is no need to suppose that participants were conscious of any further aspect of their knowledge (e.g. what the rules of their induced grammar were). If, based on task instructions, participants form the representation `I am thinking that I should apply the first grammar I studied' they are conscious of their desire to apply the first grammar. If the knowledge pertaining to this grammar is represented predication explicitly, the mental specification that it is that grammar that they want to apply may be sufficient to ensure that the grammar does apply, and so the participant has volitional control because of the predication explicitness of the representations formed during learning. But the representations of the knowledge about the grammar may not make explicit that the rules are facts, or that the knowledge is knowledge. In that case, participants may have volitional control but regard their responses as guesses, an outcome found by Dienes et al (1995). In several studies, there was evidence that participants did not explicitly represent the attitude of knowing towards many of their grammaticality decisions, and thus they were not conscious of this knowledge as knowledge. We suggest that the reason for this is precisely that participants did not have conscious knowledge of their grammar rules, and thus could not know that their grammaticality decisions were based on sound knowledge.

These comments illustrate how one can empirically tease apart whether the knowledge is predication implicit or not, and whether it is attitude implicit or not.This enables future research to determine which aspects of knowledge are left implicit in the representations formed during different types of learning. Such research could address the question of whether different types of implicitness correspond to qualitatively different learning systems. In addition, future research needs to address other implicit learning paradigms (see Dienes & Berry, 1997, and Stadler & Frensch 1998 for detailed reviews of implicit learning generally.)

In this paper, the natural language meaning of the implicit-explicit distinction was applied to knowledge representations, where knowledge was taken as a propositional attitude held towards a proposition. A series of different ways in which knowledge could be implicit or explicit directly followed from the approach. The most important type of implicit knowledge consists of representations that merely reflect the property of objects or events without predicating them to any particular entity. The clearest case of explicit knowledge of a fact are reflective representations that represent one's own attitude of knowing that fact. We argued that knowledge capable of such fully explicit representation provides the necessary and perhaps sufficient conditions of conscious knowledge. This view is consistent with the suggestion of Kihlstrom et al (1992) that it is bringing knowledge representations "into contact with" the representation of the self that enables consciousness, because that connection defines the self as an experiencing agent in possession of the knowledge. Kihlstrom et al suggested that it is this connection to the self that is lacking in implicit perception; we agree, and also suggest that the lack may be even deeper, the perceptual knowledge may lack not only representation of the self, but even predication to a particular event (e.g. what happened a few seconds ago).

Our analysis also corresponds in places to some recent analyses by Cleeremans (1997) and Dulany (1991, 1996). According to Cleeremans (1997) , "knowledge is implicit when it can influence processing without possessing in and of itself the properties that would enable it to be an object of representation (p. 199)". Knowledge can be an `object of representation' if the participants can metarepresent their representation of the knowledge as having various properties; for example, if they can metarepresent it as accurate (or inaccurate), as judged to be true (or false or undecided), or as properly caused (or not). Thus, Cleeremans' criterion corresponds to one aspect of the distinction between attitude implicit and explicit; in particular, to whether the metarepresentation (0) given in section 2.1.2 is formed - (0) being the representation that "the representation of Fb is a fact is possessed by the system" [21]. If the content of a piece of knowledge, acquired by a reliable process, can be specified by the participant even as a guess, then it is not implicit by Cleeremans' criterion. As we argued in the section on artificial grammar learning, the participants behaviour may indicate that a grammatical decision has been taken to be accurate (by the participants' consistent responding), but the person may judge the decision to be a guess. Thus, the attitude of knowing implied by the participants' behaviour has not been explicitly represented. The piece of knowledge `this string is grammatical' is unconscious as knowledge, but it is conscious as a guess, because the participant can entertain higher order thoughts about it (`I guessed that this string is grammatical'). A deeper form of implicitness occurs when one cannot even entertain a higher order thought about the knowledge; this corresponds to Cleeremans' definition of implicit and to complete attitude implicitness in our terminology.

Cleeremans argues that connectionist networks are particualrly suitable for producing implicit knowledge, an analysis that agrees with our own (see Dienes & Perner, 1996). In a connectionist network, the only information available for further transmission through the system is the activation of units (by assumption, for a real connectionist network, not a simulated one). Thus, knowledge embedded in weights is simply not available to be represented as accurate or inaccurate knowledge, so it naturally satisfies Cleeremans' definition of implicit. On the other hand, Cleeremans argues that in a symbol system representations appear to have at least the potential to be attitude explicit because the system that uses them could always decide whether or not it possesses them. Dulany (1996) makes a stronger claim. Like us, he describes consciousness as involving an agent (I) holding an attitude towards some content; but according to Dulany propositional content is always conscious.

Our analysis makes a distinction between predication explicit (which could be a symbolic representation `Fb') and, among other things, explicit representation of attitude; only the latter representation would produce consciousness of the content Fb. It may be true as a matter of empirical fact that any predication explicit representation also allows attitude explicitness, and Dulany's claim would be true. This is a bold empirical hypothesis, but our analysis makes clear that there is no a priori reason for believing it to be true - why should a representation formed, for example, for some local need by a part of our perceptual system inevitably enable attitude explicit representations? In section 4.1 we indicated that the predication explicitness of some types of (factivity implicit) perceptual knowledge is an open testable question.

Both Dulany (1991, 1996) and Jacoby (e.g. Jacoby, Lindsay, & Toth, 1992) argued that implicit processes change subjective experience (see also Perruchet & Gallego, 1997). In our analysis, predication implicit knowledge (i.e. maximally implicit knowledge) can change behaviour and we take it for granted that such behavioural change is accompanied by conscious experiences. In a subliminal perception experiment, for example, the activation of the word form 'red' may lead to a 'red' response on a forced choice objective test. This behaviour would be accompanied by the thought 'red pops into mind', or something similar. But the perceptual event was not consciously experienced as a perceptual event; this would require the representation `I am seeing the word red on the screen' (fully attitude explicit knowledge) to be produced directly by the act of seeing the word red on the screen. The predication implicit representation `red' might trigger inferential thoughts that `I must have seen the word red on the screen'. These higher order thoughts enable the participant to be conscious of the possiblity of having seen red, but these inferences do not constitute the conscious perception of red. So, like Jacoby, Dulany, and Perruchet we do suppose that implicit knowledge is often accompanied by conscious experience; one just has to be clear about what it is that the person is conscious of. But we do not claim that all implicit knowledge leads to conscious experiences. For example, it is possible that the perceptual system considers various perceptual hypotheses (e.g. predication implicit features, concepts, or schemata) before settling on one (e.g. Marcel, 1983b), predicating it to an individual. The other hypotheses may never influence conscious experience at all (albeit they had the potential). Also, a representation may not itself lead to conscious experience, but it may cause other representations downstream of processing that produce conscious experience.

Similarly, an attitude implicit rule may lead one to feel good about a particular part of an English sentence or other grammatical string; this is a conscious experience, but not of the rule. A participant implicitly learning an artficial grammar might induce the rule `T can follow M', without predicating it to a grammar, representing it as a fact or not, or representing an appropriate attitude towards it. Nonetheless, the knowledge may make the bigram `MT' look familiar, e.g. induce a conscious experience that `MT looks natural'. The participant might infer a further thought `in this grammar, perhaps T can follow M'. If this happens, the participant, by observing his own behaviour has induced a piece of explicit knowledge and this explicit knowledge co-exists with the implicit knowledge the participant already had. Within the participant's knowledge box is the unconscious representation `T can follow M', not predicated to any particular grammar or represented as a fact. In addition, there is in the knowledge box the conscious representation `I see that MT looks natural'. Sometimes the unconscious and conscious representations will contradict each other, as in the experiment by Bridgeman (1991) reported in section 4.1.

Our analysis of the meaning of implicit is in itself neutral on the question of whether different systems are responsible for producing knowledge of different degrees of implicitness. However, different degrees of implicitness will be useful for different purposes, and our view of the evidence is that different systems often do realize different degrees of implicitness in their knowledge (for example see Section 4.1). Dienes and Berry (1997) reviewed the field of implicit learning and concluded that a natural divide was between learning that produced knowledge about which participants either explicitly represented the attitude of knowing or did not (as we indicated in section 4.4 on artificial grammar learning). Dienes and Berry recommended picking out attitude implicit knowledge by the use of confidence ratings, e.g. looking at whether participants performed above chance when they claimed they were just guessing (Dienes and Berry called this the guessing criterion). The guessing criterion was found to be useful in separating types of knowledge qualitatively different in other respects (e.g. guessing knowledge was found to be resistant to secondary tasks as compared to knowledge about which participants had confidence); but it is still a testable empirical matter if it is attitude implicitness vs explicitness that distinguishes different learning systems. We suggest that implicit learning is a type of learning resulting in knowledge which is not labelled as knowledge by the act of the learning itself. Implicit learning is associative learning of the sort carried out by first-order connectionist networks (Clark & Karmiloff-Smith, 1993; Shanks, 1995; Dienes & Perner, 1996; Cleeremans, 1997). Explicit learning is carried out by mechanisms that label the knowledge as knowledge by the very act of inducing it; a prototypical type of explicit learning is hypothesis testing. To test and confirm a hypothesis is ipso facto to realize why it is knowledge. Participants in an implicit learning experiment are quite capable of analyzing their responses and experiences, drawing inferences about what knowledge they must have. These explicit learning mechanisms when applied to the application of implicit knowledge can lead to the induction of explicit knowledge. This results in the guessing criterion being an imperfect (but still informative) guide to picking out implicit knowledge; it is not the guessing criterion but the nature of the underlying representations that defines the knowledge as implicit.

[7] We are grateful to Peter Carruthers for having pointed out in response to an earlier draft that without this addition "through a generally reliable process" our criterion (3) and with it our definition of knowledge become otiose. The practical point of criterion (3) is to distinguish reliable from unreliable sources, but even the most reliable source can in principle fail. If one requires that process to be so reliable that it necessarily follows that it produces true representations then criterion (3) would imply criterion (1) but at the cost of a practically useless criterion (3).

The following implications hold between these three types of self-explicitness for a rational agent that takes himself to be rational: (1s), (2s), and (3s) each imply (0s). (2s) implies (1s) since representing oneself as believing Fb implies that one represents Fb as true. In other words, one can't represent oneself as believing something that one represents as false. Conversely, (1s) implies (2s) since if one represents R as true one should treat it as true. (3s) strongly suggests but does not strictly imply (1s) (and hence (2s)) since representing that the knowledge was properly caused implies that it ought to be accurate, i.e., that I should take it to be accurate.

[9] Conditions (0), (i), (ii), and (iii) capture the everyday use of the word `know'. Cognitive scientists generally use a broader definition, namely, they only require conditions (0), (ii) and (iii) to hold; simply being false is not sufficient reason to prevent a piece of knowledge from being knowledge (eg Newton's Laws). Removing conditions (i) and (1) would not alter any of the conclusions that follow; note that (1s) given in footnote 6 should still be included, as it follows from (2s), so our characterization of fully explicit knowledge stands as is.

[13]More technically expressed the issue was whether one should represent the knowledge that every man is mortal as (1) a declarative axiom "

x (Human(x)

Mortal(x))" and then apply the general inference procedure "[

x (F(x)

G(x)) and F(b)]

G(b)" which means roughly: If in the data base you find for Variables F, G, x and b the expressions "

x (F(x)

G(x)" and "F(b)" then add "G(b)" to the data base, or (2) should one encode the relevant knowledge directly in a specialised procedure: "Human(b)

Mortal(b)". Our interest lies with the difference between representing the regularity that being human implies being mortal either by means of the declarative implication sign "

" or by means of an inference procedure (production) symbolised as "

[16]Although Jacoby's method constitutes a clear methodological improvement, one needs to point out a remaining weakness. There is no guarantee that all participants will use the same criterion for excluding information. Consider: Is knowledge that makes predication explicit but leaves factivity implicit (e.g., 'the word "butter" being on the list") sufficient for exclusion? Probably not, it needs to be represented as a fact. But is even that sufficient? Consider the possibility that the origin of this piece of knowledge is not explicitly represented and consequently, no justification for one's judgement can be given, then a person under justification pressure, unsure of her intellectual competence, might not consider it a reliable fact and not bring it under the exclusion criterion. In sum, although Jacoby's procedure undoubtedly provides a methodological advance in dissociating implicit from explicit memory, it still suffers from the ambiguities inherent in indirect and direct tests as measures of implicit and explicit knowledge. We will briefly return to the issue of resolving such ambiguity in our discussion of intentional control of knowledge of artificial grammars.

[17] To claim that visually guided action can be based on predication implicit representation may be too radical since Evans (1975) has shown how limited linguistic communication would be without predication. However, visual perception of and action in ones immediate surroundings may be different since relations in ones egocentric space are much more constrained than between linguistically communicating partners. In Campbell's (1993) words this is possible because the features can be used in a causally indexical way which linguistic communication cannot exploit to the same degree as people typically do not stand in exactly the same causal relation to what they communicate about.

[20] For this reason one can speak of association but not of inference. For, inferences go from state of affairs to state of affairs, i.e., reasoning of the form 'whenever X is the case then Y must be the case.' But that means X and Y are predicated to particular occasions. That associative processes are possible implicitly and without consciousness but not inferences is reminiscent of Sloman's (1996) suggestion that implicit knowledge is tied to associative processes and explicit knowledge to rule governed inference processes.

represented
	explicitly	implicitly
1.	property	individual + predication + factivity
2.(a)	property + individual	predication + factivity
(b)	property + predication	individual + factivity
3.	property + individual + predic.	factivity
4.	property + ... + factivity	none