Perhaps the most obvious reason for thinking that attention, that is, the process of attending, is a cognitive process is that when presented with some sensory input it is to possible to volitionally choose what we attend. It is also possible to instruct someone to attend to one thing versus another with immediate consequences for what the viewer ends up seeing (Mack and Rock, 1998; Ward and Scholl, 2015). Just as with many aspects of our cognition, attention is not under complete volitional control. Certain salient sensory events such as a sudden appearance of an object may cause people to automatically attend to the event whether they want to or not (Theeuwes, 2004; Theeuwes et al., 2004). Relatedly, attending to the same salient target has been shown to become easier when it is repeated (priming of pop-out)—a process at one time thought to be similarly automatic and not penetrable to an observers expectations or goals (Maljkovic and Nakayama, 1994).¹

Vision scientists once thought that it was possible to produce a set of features that are the targets of attentional mechanisms. In the visual domain, dimensions such as spatial frequency and motion direction do appear to be better targets for attentional selection than more complex attributes (Wolfe and Horowitz, 2004) and can thus be fairly viewed as “basic.” However, attempts to derive a complete set of features that form the targets of attentional selection and which divide pre-attentive perception from post-attentive perception have not been successful (e.g., Wolfe, 1998). Recent work has demonstrated that attention is not limited to any closed set of (ostensibly non-semantic) perceptual features such as a spatial frequency and orientation in the case of vision, but extends to clearly semantic attributes such as our knowledge of letters (Nako et al., 2014a), words (Dell’Acqua et al., 2007), and common objects (e.g., Lupyan, 2008; Lupyan and Spivey, 2010; Nako et al., 2014b). That people can attend to such clearly semantic categories means that attention makes use of learned object knowledge making it impossible to reduce attention to a process of selection of basic non-semantic features (see also Goldstone and Barsalou, 1998; Schyns et al., 1998).

Attending to different things has far-reaching effects on perception. At its most basic, cuing someone to attend to the left makes it easier to see what is on the left (Posner et al., 1980). Such spatial attention is often the sole focus in discussions of attention and CPP (Macpherson, 2012; Deroy, 2013), but it is also possible to attend to features in parallel across the visual field with the effect of improved ability to locate task-relevant stimuli (Maunsell and Treue, 2006), and, as further discussed in Section “Cuing Perception: Attention as a Mechanism by Which Knowledge Affects Perception”, to attend to semantic categories (Lupyan, 2008; Çukur et al., 2013; Nako et al., 2014a; Boutonnet and Lupyan, 2015)

Attending not only improves objective performance, but in some cases demonstrably changes subjective perception, enhancing contrast (Carrasco et al., 2004), saturation (Fuller and Carrasco, 2006), and changing perceived size of attended stimuli (Gobell and Carrasco, 2005). Failing to attend to something in the right way can make the difference between seeing and not seeing (hence the term ‘inattentional blindness’) (Mack and Rock, 1998; Ward and Scholl, 2015).

Attentional influences are observed “early” in both place within the visual hierarchy, and time, arguably precluding the existence of truly pre-attentive perception (Foxe and Simpson, 2002; Reynolds and Chelazzi, 2004; Hayden and Gallant, 2005). Although once controversial, it is now common knowledge that attention permeates perceptual processing through and through: from at least the thalamus in the case of mammalian vision (Reynolds and Chelazzi, 2004; Jack et al., 2006; Silver et al., 2007) and down to the cochlea in the case of audition (Smith et al., 2012). We can now say with certainty that many forms of attention work by altering the response profiles of neurons that respond to sensory inputs thereby altering (at least during certain temporal windows) visual representations (Gandhi et al., 1999; Lamme and Roelfsema, 2000; Corbetta and Shulman, 2002; Ghose and Maunsell, 2002; Maunsell and Treue, 2006; Silver et al., 2007). Although the present paper cannot do justice to the vast literature on the perceptual effects of attention (see Carrasco, 2011 for review), it would not be an exaggeration to say that no part of perceptual processing is immune from attentional effects.

The first reason for denying that attentional effects counts as evidence of CPP is to view attention as a process of selection happening after perceptual processing (often referred to as late-selection; Figure 1A). On such a view, perceptual processing may proceed in the same way regardless of what we are attending, with attention determining what contents are selected from perception. For example, Palmer et al. (1993) ask “to what extent attention affects perception rather than memory and decision?” As an illustration of a kind of attention that is well-characterized by post-perceptual selection, imagine someone scanning the walls of an art gallery trying to find the Picassos. To accomplish this, the visual system must process each painting to a sufficient degree so that, at minimum, Picassos can be distinguished from the rest. If one assumes that our knowledge of what Picassos look like resides outside of the visual system, then the best the visual system can do is deliver a ‘percept’ to whatever downstream system has the requisite knowledge. That system can in turn send a signal to examine the painting further, reject it outright as an obvious non-match, and so on. A classic example of a situation often characterized in just such a way is the process of attending to a conversation in a noisy room. Although we may have the impression that we are listening only to the voices of the people we are conversing with, on hearing our name, our focus of attention may suddenly be jerked away to another corner of the room. For this to happen, we must have been processing the ambient speech all along, at least to the level of distinguishing one’s name from all other words. Notably, such recognition of unattended conversation is hardly ubiquitous, happening only about a third of the time, and more so in people with poorer working memory (Conway et al., 2001). More generally, the locus of selection is not fixed, but depends on factors like perceptual and attentional load (e.g., Lavie and Tsal, 1994). Findings like these helped resolve the longstanding debates between early and late-selection (Lavie, 2005).

Still, to the extent that attention sometimes just selects stimuli that have already received full perceptual processing—“a subtle form [of] choosing what to perceive” (Macpherson, 2012)—one may conclude that it is therefore of little relevance to questions about effects on perception itself.

One reason why many researchers studying perception are so interested in attention involves modulation of perception rather than just a process of selecting amongst fully processed perceptual states. It is possible to be cued (by an experimenter or to cue oneself) to attend to a particular place, feature, or category, with the result that of being objectively better at perceiving. Not just remembering, not just better knowing what to do, but perceiving better (Carrasco, 2011).

Critics, however, have argued that although such effects clearly count as evidence of attention (a cognitive process) changing what we perceive, they do not count as cases of CPP because attention simply changes the input to the perceptual system. On this view, attention is something that happens before perception (Figure 1B). The perceptual system then goes on responding to the altered input in a reflexive and modular way encapsulated from the viewer’s knowledge, goals, and expectations. This argument is very clearly expressed by Firestone and Scholl (2016) who argue that attentional effects can be equated to more obvious changes in input like closing or moving one’s eyes:

To summarize the argument thus far: there are two broad objections to including attentional effects as instances of CPP. The first objection is that to attending to something involves selecting among already formed perceptual representations (Figure 1A). The second objection is that attention simply changes the input to perception. This of course changes what we see, but only because of a difference in input (Figure 1B). A related proposal is that attention “rigs up” perception without altering it (Raftopoulos, 2015b).

The first objection—attention is post-perceptual selection—faces two problems. First, although it may indeed be accurate to characterize some attentional effects in this way, it is abundantly clear that much of attention is not simply selection and operates by augmenting perceptual processing itself. Second, regardless of how “late” the attentional effect in question may be occurring and how complete the perceptual processing of unattended stimuli may be, one may wish to nevertheless include such cases as candidates for CPP if they concern behaviors that we wish to count as truly perceptual. For example, even if it could be shown that the unattended gorilla (Simons and Chabris, 1999) is fully processed, its phenomenological invisibility may be relevant if we wish to include being aware of what one sees as part of perception.

To understand why the second objection—attention as a change in input—is compelling to some, and where it goes ultimately wrong, we need to examine some of its underlying assumptions. The objection rests on an analogy between a change in input caused by a change to the sensors, e.g., moving one’s eyes to the left to better see what is on the left, or squinting to blur out some details to see the larger picture, with changes in input caused by endogenous attentional mechanisms. The analogy is at least partially justified for spatial attention. Just as moving our eyes toward a target helps us see it, we have long known that shifting attention covertly—without moving one’s eyes—can likewise lead to perceptual improvements (Posner, 1980). Covertly attending to a spatial location enhances spatial resolution, improving performance on tasks that benefit from enhanced spatial resolution (Yeshurun and Carrasco, 1998). Covert attentional shifts are closely correlated with eye movements (e.g., Hart et al., 2013) and share common neural mechanisms. For example, electrical stimulation of the frontal eye fields can evoke both saccadic eye movements to specific locations and attentional shifts.² Such findings that make it sensible—on first glance—to conclude that perceptual changes due to attention are just like those caused by changes to changes to eyegaze. As we shall see, the analogy quickly breaks down when we go beyond spatial attention. The domain of spatial attention, however, allows us to better understand why a change in input (whether by moving one’s eyes or moving covert attention) would not constitute CPP. The reason is that the change in perception caused by such a change in input is not content sensitive. Insofar as looking to the left helps us see things on the left solely due to a change in what light now enters the eyes, it will be equally helpful for everything that is on the left. This improvement is independent of whether our intention was to look to check for oncoming cars or for pedestrians. In the literature on CPP, this is broadly referred to as a lack of semantic coherence between the cognitive state and the resulting percept (Pylyshyn, 1999; see Lupyan, 2015c; Stokes, 2015 for discussion).

As I will argue below, although some types of attentional shifts may lack semantic coherence, this is not the case for other kinds of attentional effects. It is one thing to find that attending to the left adds visual detail to anything on the left. But it is quite another to discover that one can attend to a certain object or object category with the perceptual consequence being changed perception of the content that is being attended. Note that even if one argued that the reason that attending to, e.g., cars helps one see cars better is through a change in input to perception, such a change would have to involve a content-specific change and is thus a qualitatively different kind of effect than simply seeing better anything in a particular location. This point is discussed in greater detail below, but first I would like to illustrate how easy it can be to confuse confounds with mechanisms when thinking about CPP.

If perception is cognitively penetrable, we should be able to find cases where the same physical input can be perceived differently depending on the cognitive state of the perceiver. The existence or ambiguous or bistable images of the kind shown in Figure 2 provide a natural starting point. That visual bistability is a perceptual phenomenon is supported by both the phenomenological potency of viewing bistable displays and by studies of its neural correlates (e.g., Tong et al., 1998; Meng and Tong, 2004; Kornmeier and Bach, 2005, 2012).

That there are images that can be perceived in multiple ways is not necessarily relevant to the CPP thesis. Consider what is perhaps the best known example of bistability—the Necker cube (Figure 2A). The Necker cube can be perceived as extending in depth in two mutually exclusive ways: as if the viewer is looking at it from the top or the bottom. The same 2-dimensional image has two different three-dimensional interpretations indicating that there is a many-to-one projection between 2-dimensional images and 3-dimensional objects. The situation becomes relevant to the CPP thesis if the same 2-dimensional image can evoke a different 3-dimensional interpretation depending on the viewer’s cognitive state.

An effective way of inducing a switch between the two interpretations of the Necker cube is to look at different parts of the image. For example, looking at the right cross in Figure 2A causes most viewers to perceive the cube as if looking at it from the top, while looking at the left cross causes most viewers to perceive the alternate perspective. Of course this is utterly unconvincing as a demonstration of CPP. The reason is an apparent lack of semantic coherence. One assumes that the effect of looking at the left or right cross would bias perception in the same way regardless of viewer’s cognitive state.

But it turns out that the intention to see the cube in one way or another not only independently affects which 3-dimensional one sees, but has a considerably larger effect on the interpretation than where one looks (Hochberg and Peterson, 1987; Toppino, 2003; Meng and Tong, 2004; see also Peterson and Hochberg, 1983; Liebert and Burk, 1985; Peterson, 1986). What about covert attention? In contrast to eye movements, these are more difficult to control. One solution is to present a viewer with a very small Necker cube for which covert or overt attentional shifts ought to be less consequential. Toppino (2003) carried out this experiment and found that the viewer’s intentions had equally large effects when viewing a cube in which the critical areas spanned less than 1° of visual angle compared to a cube an order of magnitude larger.³

Recall that what makes Necker cube ambiguous is that the 2-dimensional rendering of the cube is equally compatible with two 3-dimensional interpretations. The fluidity with which we construct 3-dimensional percepts from 2-dimensional inputs makes it tempting to think that although which 3-dimensional percept we see at a given time can be influenced by expectations and task-demands, the generation of the percepts themselves is not subject to our knowledge and expectations (Pylyshyn, 1999). But is this true? The very ability to see shapes like the Necker cube as being 3-dimensional is not hardwired. It depends on having had sufficient visual experience (Gregory and Wallace, 1963). For example, one individual who regained sight after being blind between the ages of 3 and 43 described the Necker cube as a “square with lines” (Fine et al., 2003). Still, it may be argued that given sufficient visual experience, the visual system matures sufficiently to allow the process of computing depth from 2-dimensional cues to function in an automatic bottom-up way free of further influences of knowledge and expectations. But this is not so. Both images in Figure 2B are 2-dimensional and composed of all the same elements. To the “early” visual system the two objects should look much the same. Yet, the left object is readily seen as a 3-dimensional espresso maker casting a shadow while the object on the right continues to look 2-dimensional (Moore and Cavanagh, 1998). The availability of the 3-dimensional interpretation that is competing (and in this case, quickly winning) when viewing the espresso-maker is simply unavailable for the object on the right until one gains appropriate experience such as glimpsing an enriched grayscale depiction that makes its 3-dimensionality easier to see (see also Sinha and Poggio, 1996).

Figure 2C shows another example of an ambiguous image of striking simplicity. When shown this image, the majority (28/50) of participants recruited from Amazon Mechanical Turk reported seeing a 2-dimensional figure (Lupyan, unpublished data). Of these, 61% described it in terms of lines and angles and 39% described it in terms of higher level units such as a staircase or sideways alphanumeric characters: an L and two Zs, or two Zs and a 7. But there is a 3-dimensional alternative that was apparent to the remaining 22/50 observers: an embossed letter E. It is possible, of course, that the ability of the latter group to perceive the alternate interpretation is strictly due to differences in perceptual experiences. Perhaps people who see the embossed letter are those who have previously seen many more embossed letters and therefore are better at recognizing them. But when a separate group of 50 participants were presented with the very same image and informed that it was possible to see it as a letter, about 92% were able to see the embossed E, showing that—controlling for prior perceptual experience—simple verbal cues can affect what people see.

Figure 2D shows a different kind of ambiguity. Here, the bistability is between two meanings that can be constructed from the same visual input by assigning the same contours and feature to different parts: the chin in one alternative is the nose in the other. Switching between the two interpretations can be aided by selectively attending to different parts of the image, but can also be accomplished by nonvisual cues, e.g., hearing a voice of a young woman prior to seeing the display (Hsiao et al., 2012). Figure 2E poses a similar problem to Figure 2D except that seeing the alternative to the initially dominant parrot requires a more significant restructuring of the scene. The alternatives are now not between two kinds of faces, but between a typical-looking parrot and a very atypical woman in body paint. After accomplishing this restructuring, the viewer now has a second stable interpretation that can begin to compete with the initial interpretation (Scocchia et al., 2014). An interesting and to my knowledge untested possibility is that it is only after this initial restructuring that the second interpretation become a target for effective attentional selection.

Another example of a basic visual process being affected by knowledge is shown in Figures 2F,G. Lest our visual system be limited to processing a small number of fixed inputs, it is critical to have a way of parsing inputs into constituent (and generative) parts. The most basic way to parse a visual input is by distinguishing the from the ground. How can we tell what is the figure and what is the ground in Figure 2F? The solution originally conceived by the Gestalt psychologists is to formulate a set of perceptual ‘laws’ (or biases) such as: objects occupy less area than the ground, objects are generally enclosed and form contiguous regions, objects often have symmetrical contours. Notice that none of these make any mention of object meaning and do not take into account prior experience with the candidates for object-hood. As predicted by these Gestalt grouping principles, in Figure 2F-left it is easier to see the center black region as the figure than to see the white “surround” as the figure. In Figure 2F-right, the situation is more ambiguous; the black and white regions appear to make equally good figures. But consider what happens when the figures are rotated by 180° (Figure 2G). The Gestalt dispreferred regions now appear as figure in Figure 2G-left, while in Figure 2G-right, the white and black regions are now unambiguously perceived as figure and ground, respectively (Peterson et al., 1991). This basic finding and the subsequent work by Peterson and colleagues (Peterson and Gibson, 1994; Trujillo et al., 2010; Cacciamani et al., 2014) provides an obvious challenge to explaining figure-ground segregation using perceptual laws that are not sensitive to content. The relevance of such findings to CPP is that they show that figure-ground segregation does not operate in a content-neutral way and is sensitive to at least some aspects of meaning (see Peterson, 1994; Vecera and O’Reilly, 1998 for discussion). ⁴ Results such as these also challenge accounts on which perception proceeds through a series of serial operations with earlier ones informationally encapsulated from the results of later ones. Indeed, the idea that object knowledge affects figure-ground segregation appear downright paradoxical if one assumes that the process of figure-ground segregation is what provides the input to later object recognition processes (see also Lupyan and Spivey, 2008; Kahan and Enns, 2014). But finding that recognition can precede and influence such “earlier” perceptual processes is exactly what one would expect if the goal of vision to provide the viewer with a useful representation of the input (Marr, 1982), and to do so as quickly as possible (Bullier, 1999).

Figure 2H provides another example of the role that prior knowledge can play in constructing meaning from an otherwise meaningless visual input. Often called “Mooney images” (Mooney, 1957) such two-tone images can be seen perfectly well, but the majority of people, most of the time, are unable to perceive anything of meaning in the image.⁵ In the case of Figure 2H, approximately 10% of viewers spontaneously perceive the meaningful object. The situation changes dramatically when people are provided with a verbal hint. Told that there is a musical instrument in the image, about 40% quickly see the trumpet. Such verbal cues not only improve recognition, but have additional perceptual consequences. Perceiving the image as meaningful helps people perform a simple perceptual task—determining whether two Mooney images are identical or not. These behavioral improvements were related to differences in early visual processing (specifically, larger amplitudes of the P1 EEG signal, Samaha et al., 2016; see also Abdel Rahman and Sommer, 2008). Contra Pylyshyn’s (1999, p. 357) statement that “verbal hints [have] little effect on recognizing fragmented figures”, we find that not only do verbal hints greatly enhance recognition, but they facilitate visual discrimination.

Despite the important differences between the cases shown in Figures 2A–H, there is something to be gained by attempting to unify them through the lens of perception as an inferential process—a process of generating and testing hypotheses (Gregory, 1970; Barlow, 1990; Rao and Ballard, 1999; von Helmholtz, 2005; Enns and Lleras, 2008; see also Clark, 2013; Hohwy, 2013 for overviews). For example, at the level of object representations, the Necker Cube generates three hypotheses: (a) a 2-dimensional collection of lines, (b) 3-dimensional cube extending up, and (c) a 3-dimensional cube extending down. Hypothesis (a) is dispreferred because it leaves too much unexplained. Accepting (a) would mean that the angles and lines are arbitrary. Hypotheses (b) and (c) offer a simpler description: what explains the arrangement of the lines is that they correspond to a cube. These two hypotheses are equally good at accounting for the arrangement of the lines, but yield mutually exclusive percepts and as a result begin to oscillate (see Hohwy, 2013 for general discussion; see Rumelhart et al., 1986; Haken, 1995; Sundareswara and Schrater, 2008 for examples of computational models).

In the language of predictive-coding, for someone with normal viewing history, hypothesis (a) has higher surprisal (lower ‘goodness’) than hypotheses (b) or (c). In Figures 2F,G, a segregating the figure from the ground should take object semantics into account because semantics affects the likelihood that a given feature corresponds to an actual object. We can apply the same principles of predictive coding to better understand what is happening in Figures 2C,H. Representing these as a meaningless collection of arbitrary lines results in a less compressible representation than representing them as meaningful objects (see Pickering and Clark, 2014 for a discussion of the relationship between predictive coding and compressibility). This attempt to ‘explain away’ sensory inputs in as compact way as possible is a common foundation of the various predictive-coding models of perception (van der Helm, 2000; Huang and Rao, 2011; Friston et al., 2012), with preference for simplicity going well beyond perception (Chater and Vitanyi, 2003; Feldman, 2003).

In attempting to ‘explain away’ Figures 2C,H, however, a hypothesis corresponding to meaningful objects is simply unavailable to most people. As soon as one becomes available, e.g., as a result of a verbal hint, the hypotheses dominates perception and we see the previously meaningless collection of lines as something meaningful percept (an embossed E, a trumpet) (see Christiansen and Chater, 2015 for a discussion of this same idea of continuous re-coding of input into chunks in the domain of language processing).

Although beyond the scope of this paper, it is worth noting that Gestalt principles and other “laws of perception” are not in conflict with theories focusing on minimization of prediction error. The latter theories can be seen as attempting to explain perceptual laws in more general terms. For example, a perceptual “law” such as common fate (wherein separate features all moving together against a background are likely to be grouped into a single object) can be thought as minimizing surprisal/ prediction error by positing a hypothesis that the moving parts can be predicted by a single cause—their belonging to one object. This hypothesis is preferable to the more complex alternative (corresponding to higher surprisal/prediction error) of there being multiple independent causes to the common motion. Our resulting percept of a single moving object is the phenomenological consequence of that simpler hypothesis being preferred.

Learning to associate certain visual inputs with meaningful categories: faces, letters, espresso makers, body-painted women, trumpets, etc., makes these richer hypotheses available as potential alternatives. We (our visual system) can evaluate the likelihood that an input corresponds not just to a visual object, but to a trumpet, or the letter E. These alternatives are preferred to the extent that they offer stronger predictive power, explaining for example, the observed placement of the various visual features. Allowing vision to benefit from these higher-level hypotheses helps make meaning out of noise.

In discussing how visual knowledge can lead observers to perceive the same sensory input in different ways, I conflated two kinds of effects of cognition on perception. The first concerns the finding that previous experiences with letters, faces, and various objects look like can influence the operation of even basic perceptual processes like figure-ground segregation and construction of 3-dimensional structure. The second is that it is sometimes possible to change what one sees through various cues. For example, the likelihood that people perceive Figure 2C as a single three-dimensional object is affected by being told that it is possible to see it as a letter. Some critics of CPP contest that CPP of the first type (sometimes called “diachronic penetrability”, see McCauley and Henrich, 2006 for discussion) is not really evidence of CPP because it merely shows that such visual knowledge has become incorporated into the visual system over time at which point it (apparently) no longer counts as cognitive. I will forego discussing this rather odd argument. Instead, in this section, I elaborate on the second kind of effect—sometimes called synchronic penetrability—wherein similar or even identical inputs are perceived in different ways depending on the cognitive state of the viewer at the time the input is perceived (see also Klink et al., 2012).

One way to change perception is by using a perceptual cue. For example, to help people see the embossed E in Figure 2C, one can cue them with a conventional letter “E” (Figure 3, top row). To bias people to see the young woman in Figure 2D, one can cue them with a biased version of the figure (Figure 3, middle row), and to help people see the trumpet in the Mooney image in Figure 2H, one can cue them with a more conventional picture of a trumpet (Figure 3, bottom row) or else trace out the outline of the trumpet in the original image.⁶ If such perceptual cues were the only way to affect how an ambiguous or under-determined image is perceived, such cueing effects would be of little relevance to CPP. But there are other ways of cueing perception. For example, I suspect that simply hearing “Eeee” immediately prior to or during seeing Figure 2C would increase the likelihood of perceiving it as a single three-dimensional letter E. Similarly, an auditory cue—the voice of a younger or older woman biased people to perceive the younger or older woman in Figure 2D, respectively, an effect that was additive with effects of spatial attention (Hsiao et al., 2012). Finally, although not empirically tested to my knowledge, it is conceivable that hearing a trumpet sound can help people see the trumpet in Figure 2H.

Such cross-modal effects are sometimes excluded from counting as instance of CPP because, it is argued, they merely show automatic influences of one perceptual modality on another—an intraperceptual effect (e.g., Pylyshyn, 1999, sect. 7.1) rather than an effects of cognition on perception. Drawing such an intraperceptual boundary strikes me as self-defeating, for it would mean that the knowledge as what an “E” sounds like, male and female voices, and musical instruments would all become part of the perceptual system. Someone holding the view that audiovisual integration does not count as CPP may point to the findings of Alsius and Munhall (2013) which show that audiovisual integration can occur in the absence of conscious awareness of the visual stimulus (see also Faivre et al., 2014) as evidence that such integration occurs in a completely automatic way. But this automaticity is not inevitable: even conventional audiovisual integration can be interfered with by having participants engage in an attentionally demanding task (Alsius et al., 2005).

We need not restrict ourselves to literal perceptual cues. The next two columns of Figure 3 show examples of general and more specific linguistic cues to perception. As mentioned above, being informed that it was possible to see a letter in Figure 2C more than doubled the likelihood of people seeing the embossed E, a result that one can speculate would only increase if one was given more precise information via language of what the letter was. In the old-woman/young-woman case, although many people quickly see the ambiguity, the possibility of biasing naïve viewers to one interpretation or another purely through language (i.e., without any overt perceptual cues), and that linguistic instructions continue to be effective in biasing one’s perception (e.g., Hsiao et al., 2012), speaks to the power of language to guide perception in the absence of any overt perceptual cues. In case of the Mooney image depicted in the last row of Figure 3, even superordinate linguistic cues like “animal” and “musical instrument” aid in recognition of the images. More specific cues (e.g., the word “trumpet”) are predictably more effective (Samaha et al., 2016). In other work, we have shown using hearing a verbal cue affects visual processing within 100 ms. of visual onset (Boutonnet and Lupyan, 2015), results that we interpret as showing that verbal cues activate visual representations, establishing “priors” that change how subsequent stimuli are processed (Edmiston and Lupyan, 2015, 2017; Lupyan and Clark, 2015).

Here, one may again ask whether the power of language to guide and bias perception is due to changing the input to perception via attention. The answer is that it depends on the cue. A location cue like “LEFT” is highly effective in changing where someone attends (Hommel et al., 2001), but because its effect is (presumably) content neutral, it is possible to think of it as merely a change in input. Other linguistic cues, however, have much richer semantic content: hearing “dog” helps people perceive dogs (Lupyan and Ward, 2013; Boutonnet and Lupyan, 2015). One may argue that such effects simply show that language is a good way to “rig up” perception (Raftopoulos, 2015b). But rather than being an alternative to CPP, such an argument speaks to the mechanism by which language has its effects. The fact of the matter remains that a person presented with the same sensory input can perceive it in different ways depending on a word they had previously heard (see Lupyan, 2015a for review).