1 Introduction

When perceptual beliefs are formed on the basis of perceptual experiences, it is not, in Heck’s memorable phrase, “in the same sort of way a bump on the head might cause me to believe that I am Napoleon (2000, 17)”. Unlike the bump in that case, perceptual experiences are reasons to hold beliefs, and subjects recognize them as reasons. Still, the fact that subjects treat their experiences as justifiers does not itself explain why they are justifiers. As Goldman (2008) puts it, it doesn’t explain what relations justifiers “bear to appropriate beliefs or propositions by virtue of which justification is conferred on those beliefs or propositions and not others”.

Goldman and other externalists, like Burge, deny that perceptual experiences themselves are the bearers of justificatory force, and offer an externalistic explanation of why perceptions are justifiers. This view has its own shortcomings. Since the internal story about justification is deemed insufficient, and the external one is not internally accessible, subjects are in the dark about what underwrites their perceptual beliefs. So the externalist’s story about justification ends up looking like Heck’s bump-on-the-head account of belief formation, from the subject’s perspective.

Externalists can grant that some of their perceptual justifiers are internally accessible states. But this doesn’t suffice for justification from the subject’s perspective, because what confers justifying force on the states on the externalist account still remains opaque to the subject.

In this paper, I describe a form of justification for perceptual beliefs that categorize objects. This form of justification suggests that subjects have a degree of internal access to the principles of externalistic perceptual warrant, and to features of perceptual states that give them justificatory force. If that is true, everyday subjects not only have internal access to states that externalists count as justifiers; they also have a degree of insight into why their perceptual experiences are reasons for their beliefs.

2 Justification and perceptual categorizations

When we classify objects under kinds, we can support our judgments by demonstrating low-level properties of the objects. I can point out that a tree is a certain type of pine by showing the slant of the trunk, or an oak by pointing out the texture of the bark, the shapes of the leaves or the pattern of the branches from a distance. Something similar applies to facial emotions. I can point out that someone looks angry or suprised by showing the slant of the eyebrows or the shape of the mouth or the eyes. We also point out specific low-level features to support judgments that identify individuals. In all these cases, beliefs with high-level contents (that a is a beech tree, that a is Sara) are justified by demonstrating low-level perceptual features. The low-level features include apparent shape, size, orientation, texture, colour, motion, position, pitch, volume, spatial and temporal relations like intervals, and combinations of features (eg colour patterns). The justifications seem to draw on characterization relations beween properties and kinds or individuals, or on some epistemic grasp of those relations. Call this form of justification “P”:

  • (P) We justify perceptual beliefs that classify objects under kinds or identify individuals by demonstrating low-level features of the objects.

How would P fit into available frameworks for analysing perceptual justification? In P, the beliefs justified are about worldly states of affairs (eg that an object a is an instance of a kind K). In classical foundationalist frameworks, such as BonJour’s (2001), only self-ascriptive beliefs are non-inferentially justified. So in such frameworks, the beliefs that P justifies would have to be justified inferentially. The beliefs could also be justified inferentially by basic beliefs whose contents were low-level properties (not self-ascriptions), or by perceptions of low-level properties, if perceptions entered into inferential relations. In all those cases, P would be a form of inferential justification:

  1. (1)

    Perceptions or beliefs with low-level contents inferentially justify beliefs with high-level contents. For instance, my belief that a is a beech is justified because I visually experience that a has a certain shape, slant, etc., and I believe that trees with such shapes, slants, etc. are likely to be beeches.

Beliefs of the relevant type (a is a K) would be non-inferentially justified according to the theories of perceptual justification provided by Pryor (2000), Huemer (2001), Feldman (2003), and McDowell (1994), which see perceptual experiences as direct evidence of the presence of worldly states of affairs:

  1. (2)

    Perceptual experiences with high-level contents, including kinds, suffice to provide defeasible non-inferential internal justification for beliefs with high-level contents. In that case, my belief that a is a beech is justified because perceptually it seems to me that a is a beech.

These theories are framed as if beliefs and their direct perceptual justifiers have the same contents. (See, for instance, the formulations in Huemer, 2001, 99; Pryor, 2000, 532; and Feldman, 2003, 74. McDowell, 1994, is the most clearly motivated version of this view, although McDowell later adopts a view that would come under (4) below; see note 1.) The same-content assumption receives a boost from work by Siegel (2006) and Bayne (2009) arguing that kinds (and other high-level contents) are already represented by perceptual experience, or that there are perceptual experiences of seeming-K, where K is a kind. Sameness of content seems to be a necessary feature of the theories, safeguarding their non-inferentialism. If a perception with one content justified a belief with a different content, the natural assumption—in an internalist framework—would be that the justification is mediated by thoughts that connect the contents inferentially. Later I’ll reject this assumption, but for now, I want to point out that type (2) theories cannot explain practice P.Footnote 1 In P, beliefs about the presence of kinds are justified by states with different contents, namely, low-level perceptual properties. Since in P beliefs and their justifiers have different contents, P cannot be explained by theories of perceptual justification described in (2), such as phenomenal dogmatism or conservatism.

What about externalist theories like Goldman’s process reliabilism and Burge’s perceptual entitlement? In these theories, perceptual states justify beliefs directly, in an internalistic sense of that word, that is, without mediation of any conscious inferential processes (thoughts). Beliefs and their perceptual justifiers can, and generally do, have different contents on externalist theories. In an account of perceptual justification that focuses on the relevant details, Goldman (1976) defines conditions in which “percepts”, caused by the possession of “shape, size, color, and texture” properties by objects, non-inferentially cause subjects to believe that the objects have properties like “being a dog”, “being a barn” or “being Judy”—high-level properties that include kinds.Footnote 2 For Burge, perceptual warrant is from perceptions of “shapes and colors” and transitions from them to recognitional concepts for kinds (“concepts of tomatoes and cars”) (2003, 545–6). Thus, on the externalist accounts:

  1. (3)

    Perceptions of low-level features provide non-inferential external warrant for beliefs with high-level contents. In that case, my belief that a is a beech is justified because (a) I perceive certain low-level properties, (b) those perceptions cause the belief that a beech is present, and (c) perceptions of specific sets of low-level properties are correlated to the presence of beeches by reliable or truth-conducive processes.

That P uses experiences as justifiers does not conflict with (3); there is no reason to exclude conscious perceptual states from being among the justifier states on externalist accounts. Experiential states don’t suffice for justification on those theories, but as Goldman writes, there is “no incompatibility between reliabilism and appeals to evidential states as (partial) determinants of justification” (2011, 461). Burge describes perceptual entitlement to beliefs as a “warrant that need not be fully conceptually accessible, even on reflection” (Burge, 2003, 504). A subject would have partial internal access to warrant if she was conscious of a representational state that contributes to warrant while lacking access to other states also involved in providing the warrant. So Burge’s perceptual entitlement could include warrant by experiential states.

However, the externalistic (3) cannot account for practice P, because as a form of justification, P uses thoughts to connect low-level perceptual features to kinds—that is, to execute step (c) in (3). That a is likely to be a K because it has low-level features F, G, etc., characteristic of Ks, appears as an internally accessible reason in P. (In addition to the perceptual experience of low-level features, which is a distinct justifier.) Subjects must be capable of having such thoughts, since they express them with utterances connecting terms for kinds with complex demonstratives for shapes, colours, etc. (As when one says, pointing to the veins on a leaf, “That shape is typical of beeches”, or “It’s a beech, look at those patterns”, or simply “Beech leaves look like this”, meaning the shape of the veins.) If P expresses a form of justification, it is one in which thoughts mediate to connect kinds to typical features and those connections are represented by the subject as her own reasons for holding the belief.

Is P then a form of inferential justification? I can think of only one alternative:

  1. (4)

    Perceptions of low-level features provide non-inferential internal justification for beliefs with high-level contents.

This theory looks like a non-starter because it collides with the same-content constraint, described earlier, on direct internal justification. Externalists can define non-inferential perceptual justifiers with different contents to beliefs, for the reason captured by point (c) in (3): perceptions of low-level features, and tokenings of recognitional concepts for kinds in judgments, could be connected in ways that are not internally accessible yet lend justificatory force to the judgments.Footnote 3 Internalists cannot hold that a perception with one content, p, just happens to directly justify a belief with a different content, q. One might be misled to think that they can say just this and that same-content is not an essential part of internalist direct justification. Classical foundationalists do not subscribe to it (they cannot, since their basic beliefs are self-ascriptive, unlike their justifiers). But this is misleading because classical non-inferentialists face a restriction on content which is substantially similar to same-content: they cannot claim that a basic belief I am having an experience that q is directly justified by an experience that p. As BonJour writes, a basic belief contains a “description of the very content involved in the constitutive awareness of content” of the justifier state (2001, 25, my emphasis).

In the next section, I’ll reject the same-content constraint, and argue that P is a case of (4). However, first I want to outline another account of P that some people might find tempting. On that account, even if P is a form of inferential justification, one can still be a non-inferentialist about perceptual justification.

Consider how this would work for Burge’s theory of perceptual entitlement. For Burge, perceptions of low-level features non-inferentially warrant perceptual beliefs (2003, 545–46). Burge grants that in some contexts, “one can use a back-up [inferential] reason to supplement an entitlement”, but such back-up reasons are “not necessary for being entitled to perceptual beliefs” (529): we’re entitled to hold the beliefs anyway by the truth-conduciveness of transitions between representational states in sensation, perception and belief. So in Burge’s non-inferentialist framework, P could be a form of inferential justification which provides back-up reasons for beliefs that we’re entitled to hold anyway.

Next, consider versions of (1) like Huemer’s phenomenal conservatism or Pryor’s dogmatism. Applied to judgments that classify objects under kinds, those theories hold that a’s looking like a K suffices to defeasibly justify the belief that a is a K. (For brevity: looking/seeming-K suffices to defeasibly justify K-beliefs.) The theories have to suppose there are such states as looking-K, looking-like-a-K or generally seeming-K since if they do not, being internalistic, they will collide with the same-content constraint. Some authors do argue that there are such states and that they are absent in one functional dissociation, associative agnosia (Bayne, 2009; Siegel, 2006).Footnote 4 If looking-K suffices to justify K-beliefs, it suffices even if we can (also) come up with inferential justifications connecting low-level properties to kinds, as P would do on an inferentialist reading. No inference is necessary for justification because our beliefs are justified on the strength of looks alone.

This same-content route around P can also be taken by externalists, provided they don’t discount experiential states from contributing to justification. An example of evidential justifiers in an externalistic framework is provided by Goldman (2008). He asks what justifies the recognitional judgments of an expert bird-watcher, as distinct from a novice. His reply is “differences in the cognitive processes by which they respectively proceed from visual experience to belief”.Footnote 5 Suppose for the sake of argument that there are visual experiences with high-level contents, like looking-K.Footnote 6 On this externalist picture, if the experiential states are reliably connected to other states (sensations and beliefs), and there are no defeaters, that suffices to non-inferentially justify beliefs. Again, practice P (construed as inferential) would not be necessary for perceptual justification.

Why not accept these non-inferentialist responses to the existence of P-type justifications? They make P epiphenomenal to justification, but that is not a reason to reject them. The fault is that they mistakenly concede P to the inferentialist, and in doing so, fail to identify features of perceptual justification generally.

3 Bodies of beliefs or perceptual memories?

K-beliefs (perceptual beliefs that categorize objects under kinds) depend for their formation on perceptual object recognition. When such a belief tokens a concept K for a kind, on theories of recognition, K is tokened because current perceptions of basic properties match representations of basic properties stored in perceptual templates for Ks. The externalist uses those very connections between perception, memory and concepts to justify K-beliefs. If P is inferential, it has to justify K-beliefs by using beliefs: ones that connect basic properties to kinds. The challenge for an inferentialist view of P is to show that subjects have such beliefs, about the typical properties of objects, which are adequate to justify the K-beliefs actually formed on the strength of recognition abilities.Footnote 7 The discussion that follows is restricted to visual perception because vision is the sensory modality on which theories of recognition focus.

Beliefs connecting kinds to low-level properties (shape, orientation, colour patterns, etc.) cannot stem from awareness of any inferential connections made in recognition. There are two main hypotheses about visual object recognition. On both, features are associated with kinds through visual matching. On neither are low-level features connected to object categories by drawing inferences from sentence-like representations. On one hypothesis, recognitional templates are viewpoint-dependent. Memorized views form a “view space” (Tarr & Bulthoff, 1998, 11), comparable to the representation of a quality space or similarity ordering, in which new views are “located” by an interpolation procedure (Poggio & Edelman, 1990). On a version proposed by Ullman, recognition is based on matching the stimulus to a set of views (Ullman, 1998, 42), giving visual templates a greater degree of abstraction. On the second hypothesis, recognition depends on stored visual representations of viewpoint-independent structural features (Marr, 1982; Marr & Nishihara, 1978) and topological relations between object parts (Biederman, 1987, 1995). This approach restricts the initial search for matches and supports generalization from past instances better than viewpoint-dependence, but its predictions fail for several fine-grained recognition tasks (see Tarr & Pinker, 1989; Bulthoff & Edelman, 1992; Ullman, 1998). So we’re likely to need both hypotheses to explain recognition generally (a point confirmed by neurological studies of brain regions that visually represent shape; see Kourtzi et al., 2003).

The most abstract templates are those posited by the second, structural hypothesis, but those templates still visually encode information about low-level features (how object parts relate spatially, visual cues like concavities, and the shapes of parts or “geons”). So on all the theories, recognition depends on visual matching, not on drawing inferences from sentence-like representations. Therefore, any beliefs we have about the typical features of kinds (which could be used by P to inferentially connect features to kinds) cannot stem from awareness of some body of propositional knowledge used to connect features to kinds in recognition.

For P to be inferential, it would have to rely on a separate body of propositional knowledge about the typical features of kinds. The natural place to look for such knowledge is the subject’s repertoire of concepts for kinds. Adapting a broad definition provided by Machery (2009, 12), we could describe a concept for a kind as a body of information or knowledge about a category of objects (eg about their causal properties, internal composition, and epistemic properties like appearance) which is used by processes underlying cognitive competences, like categorization and inference. Categorization includes perceptual categorizing, categorizing from descriptions, and “category production”: listing what we take to be typical properties of a category, including epistemic properties like being red and smooth for a tomato (Rosch et al., 1976). In the literature on concepts there is disagreement about whether perceptual categorization is a conceptual competence,Footnote 8 but it is uncontroversial that the category production competence is entailed by having a concept for a category. To that extent, we can assume that subjects can list some typical low-level features of kinds for which they possess concepts. This is a body of knowledge the inferentialist could draw on for beliefs about the typical features of objects.

There is a chance that the information supporting the ability to list low-level features is visual, and that the descriptions of features are produced by being extracted from visual representations. Theories of visual imagery argue that without any perceptual stimulation, following only semantic prompts, subjects have partial recall of generic low-level features for familiar kinds (Kosslyn, 1980, 1994; Tye, 1991). They hold that such recall unpacks information stored (in a non-imagistic format) in long-term memory and represents it visually by mobilizing parts of the brain used by visual perception, allowing us to verbally describe the content of the visual representations. (See Tye, 1991: 90, 99, for examples.) Some theories of concepts also hold that the prototypical features of objects listed in category production are retrieved from visual memories (Barsalou et al. 1999; Solomon & Barsalou, 2004; Prinz, 2002). But the idea that we can generate such visual imagery is not commited to neo-empiricism about concepts. Any theory of concepts that counts recognition (perceptual categorization) as a conceptual competence has to include perceptual skills in the possession conditions for concepts. The idea that our standing knowledge of the shapes, colours, etc., of objects and kinds depends on stored visual representations is compatible with any theory of concepts that does not, as a matter of principle, exclude perceptual skills from the conditions for concept possession. So it is compatible with all theories of concepts except Fodor’s (1998), which adopts that prohibition and concludes that there are no recognitional concepts.Footnote 9

Why is it important if our beliefs about typical shapes, colours, etc. depend directly on conjuring up visual memories? The type of justification conferred by such beliefs would be affected. Consider how the beliefs would contribute to justification in the framework of P. For P to be inferential, subjects would need to have a body of beliefs like the one we have been seeking, ones about feature-kind associations, and from these, jointly with perceptions of low-level features, to inferentially justify K-beliefs (perceptual beliefs about the presence of kinds). But if the beliefs connecting features to kinds are descriptions of visual memories, which have to be consulted each time before the belief can be formed, then what bears the weight of connecting features to kinds is the visual memories, not the beliefs. No inferences are involved in making the connections, even though the connections can subsequently be couched in sentential or propositional form. But they are couched after the event of justification: the justification resides in the fact that the template for Ks contains memorized visual information that matches current visual information. The beliefs just report the matching—the fact that current percepts match the visual representations I have for Ks.

The suspicion that P is not inferential, but parasitic on the visual representations that support matching, is confirmed when we consider beliefs that make finer-grained classifications. Low-level features of objects can be demonstrated to justify beliefs that assign objects to basic-level categories (dog) as well as more specific categories (subordinate categories like Labrador), and beliefs that identify objects as individuals (my own dog). The more specific categories include the kinds of items used by studies on visual memory which discovered that humans have a surprisingly large capacity to store and retrieve exact visual details for recognition. In recent and older experiments (esp. Brady et al., 2008; Standing, 1973) subjects memorize thousands of items after brief single exposures, and later succeed in correctly identifying the items or discriminating them from very similar ones (even days later in Standing’s experiment). Experiments with less items but higher degrees of resemblance, focusing only on the fine-grained discrimination of exemplars, were conducted by Hollingworth (2004) and by others.

The process for retrieving this visual information has a hierarchical structure, described as follows in a review article by Brady and colleagues (2011). Suppose that a subject perceives an item for which she has detailed visual memories, “passively stored” in long-term visual memory, which are sufficient to enable recognition of the item. First the item is recognized as belonging to a basic-level category. This classification triggers the retrieval process; retrieval is triggered by coarse classifications that do not yet discriminate subordinate categories or individuals, not by matching specific low-level features. But once the retrieval process is activated, the detailed low-level features—highly specific shapes, colour patterns, orientations, spatial relations, etc.—are also retrieved. Complete retrieval is what allows the subject to make discriminations at more specific (“subordinate”) levels, as well as to identify objects as individuals they have seen before. If on the other hand the item only closely resembles one the subject has memorized, then she is able to discriminate it from the memorized object—again because detailed visual information has now been retrieved. Detailed visual memories of objects are also retrieved on perceiving closely resembling objects, supporting discrimination. In this sense, what the experiments show is that subjects store and retrieve information sufficiently well to make very fine-grained discriminations between subordinate categories and individuals.

The memory experiments capture features of everyday non-experimental situations in which we encounter new types, sub-types and individuals and subsequently re-identify or discriminate them. They show that the formation of the categorizing beliefs depends on the retrieval of detailed visual memories, and that, at the lower (later) end of the hierarchical retrieval process, retrieval of visual details is a stimulus-dependent process, ie, depends on exposure to stimuli very similar to the ones memorized.

Turning now to justification, the inferentialist has to show that we have beliefs, connecting categories to low-level properties, which are adequate to justify the perceptual beliefs that we actually form on the strength of recognition. For each K-belief formed, the inferentialist needs the subjects to have beliefs connecting kind K to the low-level features that characterize Ks as distinct from other categories. For the identification of individuals or exemplars, the subjects need to have beliefs connecting each discriminated individual a to features that characterize a as distinct from other individuals. But the subjects had never seen the specific objects and types before the experiment, and only accessed their visual memories at the end of a process of retrieval that depends on perceptual exposure to identical objects and types or to very similar ones (permitting discrimination). So when the subjects formed beliefs that categorized the items, they would not already have had a body of beliefs (connecting individuals and categories to low-level features) adequate to justify those categorizations. If they had such standing bodies of beliefs, no process of retrieval of visual details would be required for them to form the judgments in the first place.

Therefore, if K-beliefs are to be justified by beliefs, these can only be beliefs formed by drawing on visual information retrieved during recognition—on condition, and to the extent, that the visual memories retrieved are partially accessible to awareness. Brady et al. 2008 report that participants in their experiments “volunteered information about the details that enabled them to pick the correct items” among two exemplars of the same kind, and among two states of the same object. That does suggest that the participants were able to consciously match low-level features of the stimulus to retrieved memories of visual details. To the extent that visual details can become conscious on retrieval, we can give the following explanation of practice P—that is, of how subjects justify classificatory judgments with utterances that connect them to perceptions. Drawing on awareness of the contents of retrieved visual memories, for example shapes, subjects can think that shapes S characterize Ks, or say “It’s a K—that shape [meaning S] is typical of Ks”. This proposition does no justifying of its own, it only connects S to Ks on the strength of memorized visual representations of how Ks look. If practice P works like this, it is not a distinct justificatory step. It is a distinct step in expressing the single justificatory step involved: the justification of perceptual beliefs by the processes that support recognition, which themselves are not inferential.

4 Internal access to external warrant

We can salvage an evidentialist account of P from this criticism of inferentialism. Its scope would depend on how much internal access we can gain to object recognition processes and their outputs. Some outputs which are internally accessible would not be useful to the evidentialist. For example, the sense of familiarity caused by recognition (Mandler, 1980) must count as evidence for some judgments, but it isn’t useful for explaining P, which draws specifically on feature-kind associations. The ideal scenario for an evidentialist construal of P would be the one described at the end of the last section: we gain partial awareness of visual memories during the retrieval process when exposed to certain stimuli, ones that match visual memories or very closely resemble them. But it would be speculative to generalize this scenario. The most suitable experiments to support it are those by Brady and colleagues because they focus on discriminations from sensory detail, not gist or “recollection” (of the context of encounter). That subjects reported which visual details enabled them to identify objects shows that during retrieval they became aware of memorized details at the most specific level. Unfortunately, it wasn’t the aim of that study to establish any conclusions about awareness. So at best once can assume that there are some cases in which very fine-grained classifications can be justified by reporting which features of a perceived object match, or fail to match, visual memories.

More conservatively, the evidentialist’s claim could be restricted to cases in which there is systematic evidence of conscious matching. These cases are provided by theories of visual imagery, described in Sect. 3, which argue that even without perceptual stimulation we have partial visual recall of generic low-level features for familiar kinds; and theories of concepts for which category production depends on consulting stored perceptual imagery some of which is conscious (see Prinz, 2002, Ch. 6). The reason these grounds suit the evidentialist is that they rely on conscious visual imagery. (Not all visual imagery is conscious; see Nanay, 2018 for both a review of this literature and arguments in favour of unconscious stimulus-independent modal imagery.) This is a valuable point for the evidentialist, even if its scope is more restricted than that of stimulus-dependent retrieval. To the extent that subjects have conscious access to memorized visual information about familiar kinds or individuals, there is nothing to prevent them from consciously comparing current visual perceptions to memorized perceptions. The evidentialist doesn’t have to (and shouldn’t) claim that conscious matching is required for recognition; if it can follow recognition, it provides experiential grounds for justifying K-beliefs.

Partial awareness of matching and retrieved memories would shed light on claims that we have perceptual experiences of looking-K (Bayne, 2009; Siegel, 2006). It contributes to explaining a point made by Siegel: ability to recognize an object alters the visual experience of the object, not just the way we think about the object. An object a would look like a K when the subject is aware of current visual perceptions of a matching visual memories of Ks which have become conscious, through stimulus-dependent retrieval or cognitively driven visual imagery. Here, the immediate relata of looking-K, or looking-like-a-K, are (a) currently perceived low-level features and (b) memorized perceptions of low-level features of Ks, and the relation between them is matching. In that case, looking-K is not a same-content perceptual justifier—a state that can function as a justifier by virtue of having the same, high-level, contents as beliefs. It is a perceptual justifier that connects low-level to high-level contents through matching. In terms of the justification options enumerated in Sect. 2, seeming-K is not suitable for internalist dogmatism or conservatism (2), but for the kind of justification described in (4): it provides perceptual justification which is different-content, non-inferential, and internally accessible.

That P provides subjects with internal grounds for holding beliefs should be welcomed by internalists. Does it count as an internalistic form of justification? If internalism is the view that internal access is only a necessary condition for justification, then P can count as internalistic.Footnote 10 In any case, the justification has to comprise more than the internal access. Earlier I explained why any utterances occurring as part of P are not a distinct justificatory step but essentially piggyback on the single justificatory step provided by visual recognition processes. An analogous point applies to awareness. Justification is given by states and processes to which awareness gives partial access. If the connections between perceptions, memories and tokenings of recognitional concepts were inaccurate, no awareness of them, and no descriptions of them by P, could generate justifying power. That they are accurate is a property of process and state types and thus goes beyond any token state, including the evidentialist’s current experiences:

Present perceptual representation makes a contribution to the warrant of a perceptual belief only insofar as the perception involves the perceptual application of representational types whose natures are individuated by types of perceptual paths that connect them reliably to the referents of the perceptual belief (Burge, 2003, 547).

So the experiences, or their descriptions by P, have justificatory force only by being placed in an externalist framework for perceptual justification, one on which, in Goldman’s words, “there is no incompatibility between reliabilism and appeals to evidential states as partial determinants of justification”.

Consider how Burge’s account of entitlement can provide such a framework. Burge denies that beliefs with high-level contents are inferentially justified by beliefs with low-level contents (2003, 547).Footnote 11 Perceptual warrant is from perceptions of “shapes and colours” and transitions from them to tokenings of recognitional concepts (“concepts of tomatoes and cars”). The transitions consist of percepts being “associated or matched, in learned but automatic ways” to recognitional concepts (2003, 545–6). In the vocabulary used here, perceptions of basic properties warrant K-beliefs by virtue of the percepts being reliably matched to perceptual memories in recognitional templates for Ks. This entire picture can be drawn, and is mainly drawn by Burge, without using the concepts of experience or consciousness. On the evidentialist construal of P proposed here, P describes the subject’s awareness of current perceptions (of shapes, colours, etc.) matching memorized perceptions of Ks. The awareness amounts to partial internal access to the states (visual perceptions and memories) and processes (visual matching) that provide Burge’s warrant.Footnote 12

Earlier I claimed that practice P shows that subjects have a degree of internal access to the principles of perceptual warrant, and some insight into why their perceptual experiences are reasons for their beliefs. Now I can explain what I mean by these claims. It is a principle of perceptual warrant for K-beliefs that for the belief to be perceptually justified, current perceptions of, eg, shapes have to match memorized shape perceptions associated with the concept for Ks specifically, not with concepts for other kinds. (Alternatively, memorized shape perceptions are part of the concept for Ks as perceptual information required to possess the concept as a body of information underpinning competences like perceptual categorization.) Practice P shows that subjects have internal access to this principle, because the utterances used by P (“It’s a K—that shape is typical of Ks”) express the fact that the subject takes the belief to be justified only when the principle is met.

It is also a principle of external perceptual warrant for K-beliefs that they have to be caused by process types that reliably cause such beliefs to be tokened in the presence of instances of K. Visual matching processes are such types: they connect the visual detection of low-level properties to retrieved visual memories, and to associations of those memories with concepts, in such a way that recognitional concepts for Ks end up being reliably tokened in the presence of Ks. When subjects justify their belief that a K is present (or not) by reporting that the object’s shape matches (or fails to match) visual memories for a specific category, they use visual matching as their criterion for whether or not an object of a certain kind is present. So they subscribe to that principle, since they correctly identify matching as the process that reliably leads to tokenings of recognitional concepts for Ks in the perceptual presence of Ks.

To offer some perspective on these claims, suppose that instead we adopted the view of perceptual justification attributed to phenomenal dogmatists (option 2 in Sect. 2). On that view, the contents of our perceptual experiences include kinds, and such experiences defeasibly but non-inferentially justify beliefs, by virtue of having the same contents as the beliefs. On such a view of justification, subjects treat perceptual experiences as grounds for their beliefs. But they have no access to any principles that make those experiences justifiers. This is because in (2), the sole evidential basis is the perceptual experience itself. There is no access to the processes that connect perceptions, memories and concepts; as a result, there is no identification of the process-types that reliably connect tokenings of concepts to the presence of instances (visual detection of features, retrieval, matching, and reliable associations with concepts). On the proposed view, by contrast, subjects do have insight into why their experiences are justifiers. They correctly identify the processes (recognitional ones) that reliably cause concepts for kinds or individuals to be tokened in the presence of instances and individuals. They correctly identify the roles played by their experiences (perceptions, memories) in those processes (the matching process). They recognize the outputs of recognition (retrieval and matching) as relevant to justification. They correctly identify the ways in which those outputs support or fail to support beliefs (they use the experience of visual matching to support or refute a judgment). Briefly, subjects have a significant degree of internal access to the nature of the externalistic perceptual warrant that justifies their perceptual beliefs. They have more access than they would have on the internalist view.

5 Conclusion

I argued that P-type justifications, which justify beliefs with high-level contents by using perceptions of low-level features, exploit a partial awareness of connections made by object recognition processes. The recognitional processes provide justification; P’s utterances just report whatever degree of awareness we have of the processes and their outputs. To reach this conclusion, I argued that we are unlikely to have standing bodies of beliefs adequate to justify the classifications that we actually make when we form perceptual beliefs. I relied on empirical work which shows that subjects can have partial awareness of visual memories in two ways: from cognitively driven visual imagery, and from stimulus-dependent retrieval. There is a lack of empirical work focusing on whether we can have stimulus-independent retrieval of visual details (from massive visual memory storage) which parallels our ability to discriminate objects and types in stimulus-dependent conditions. If it turns out that we do have such a capacity, inferentialists could use it against my view.

The point of rebutting inferentialism was not just to support externalism. It was to argue for a form of evidentialism that tracks externalistic warrant, through a partial awareness of the recognitional processes which provide warrant for classificatory judgments. This would give subjects some insight into why their perceptual experiences justify their beliefs—into the nature of perceptual justification. The justification itself is externalistic, but that does not mean that as everyday subjects we have no grasp of it and its principles. In fact I argued that we have more access to the nature of justification on this view than on internalist views like dogmatism.

If subjects know the principles of perceptual justification to the extent suggested in Sect. 4, what prevents their justifications from consciously applying that knowledge, and being inferential after all? Even if we knew and applied an externalistic theory of justification in our everyday judgments, it would make no difference to the outcome of our justifications. The theory would be general, while each particular judgment gets its justification from specific perceptually formatted information (in perceptions, memories and matching) to which we have varying degrees of access. So even if subjects know about the nature of perceptual justification, their justifications won’t be inferential.