Eating and drinking are amongst life’s most multisensory experiences (see [136], for a review). Indeed, according to eminent neuroscientist Gordon Shepherd, flavour perception (e.g., when wine tasting) engages more of the human brain than most activities [123]. Visual information has been shown to influence our expectations concerning the sensory properties of food and drink, with colour acting as a particularly salient sensory cue (e.g., [50, 137]). It has long been suggested that “we eat first with our eyes” [5], and visual cues can dramatically influence both the palatability of food as well as its perceived taste, aroma/odour, and flavour [29, 153, 159, 189].

A long history of research has investigated the impact of adding and/or changing the colour of food and (more commonly) drink on taste/flavour perception (see [137, 139], for reviews), as well as on hedonic responses [200]. Separately, however, researchers have also been interested in the more abstract crossmodal correspondences demonstrated between colour patches or colour words and tastes, aromas, and/or fully-fledged flavours [99, 155, 184]. Crossmodal correspondences have been defined as the sometimes surprising associations that we all experience between seemingly-unrelated sensory features in different sensory modalities (see [133, 144] for reviews). Such correspondences could arise as a result of similarities in the neural coding of the corresponding stimuli, the statistical co-occurrence of sensory features in the environment, and/or the use of the same linguistic/semantic terms to describe different sensations. There is, however, also growing evidence for the emotional mediation underlying many crossmodal correspondences, such as in the case of the crossmodal matching of music and visual stimuli including colours or paintings [87, 148]. Emotional mediation has also been shown to play an important role in the crossmodal matching of colours with odours (see [113]; see also [149] for a recent review). That said, the cognitive/neural mechanisms underlying crossmodal correspondences with colour that involve taste and flavour (olfaction) may not necessarily be the same.

While significant effects of adding/changing the colour of food and drink have been reported in many studies, no significant results have been reported in certain others (e.g., see [138] for a review). As such, the empirical research that has been published to date has been somewhat unclear as to the precise conditions under which changing the colour modifies the taste and/or flavour of food and drink. One attempt to explain the inconsistency in the literature on food colouring was put forward by Shankar and her colleagues [119–121]. These researchers highlighted how people often have (or internalize) different flavour (aroma) associations with specific colours. For example, Europeans typically associate the smell/flavour of cinnamon with the colour brown (i.e., their association is with the dried spice) whereas North Americans often tend to associate it with the dark red of the hard-boiled cinnamon-flavoured sweets instead (cf. [30, 64, 109, 118, 132]). In other words, the colour associations that people hold with aromas or flavours may differ. Shankar et al. [119] conducted a series of studies showing that the degree of discrepancy between the aroma/flavour of a coloured drink, and the flavour that participants associated with that colour, determined whether or not, and how exactly, the colour of the drink biased the participant’s identification of its aroma. When the discrepancy was small, assimilation (or visual dominance) was observed. When the discrepancy was larger, however, the colour cue was seemingly ignored. Given that most of what we think we taste we actually smell (via retronasal olfaction; see [138] for a review), Shankar et al.’s ‘degree of discrepancy’ account of coloured aromas should likely also apply in the case of colour’s influence over flavour perception (though see [61]).

Cultural differences in how ingredients are used in cuisine may also influence taste and flavour expectations. In the US, for example, pumpkin is often used in the context of sweet foods (e.g., pumpkin pie, pumpkin spiced latte Frappuccino at Thanksgiving), whereas it is more typically considered a savoury ingredient in other parts of the world (see also [14], on the differing cultural associations with nutmeg, and on the cultural differences in the basic tastes that are associated with food, or rather with the aromas of various spices). Within the literature on both colour and olfaction, the evidence suggests that when multiple possible associations are consistent with a specific colour or odour stimulus, people’s preferences are better predicted by a model that uses a weighted average of all of people’s individual preferences for the associated objects than are those preferences that are based on a single-associate model (e.g., [86, 114]). When people’s own idiosyncratic object associations (e.g., a colour might remind someone of their favourite sweater) are included, as well as “standard” associations generated across a group (e.g., the colour blue for the sky), the model’s fit improves still further; thus further reinforcing the role of experience in developing associations [115, 116].

One might wonder whether this approach, known as ‘Ecological Valence Theory’ [86] could also be extended to help predict the crossmodal correspondences that have been documented between colours and odours/tastes, and not just people’s hedonic preferences within a sense. Here, it is perhaps worth noting how the colours that synaesthetes associate with letters have been shown to be predictable, at least in part, based on the colours of the objects that they happen to associate with words that start with the letter. So, for example, the letter ‘A’ is often associated with ‘Apple’, with apples typically being green or red [71], thus perhaps explaining the varying colour associations that have been reported in this particular case.

Over the years, a number of researchers have suggested that the phenomenon of sensory dominance might help to explain (or at the very least to describe) the crossmodal influence of colour upon people’s taste/flavour judgments [197, 198]. As a case in point, according to Zampini and his colleagues, the flavour that is associated with the colour of a food or drink product will sometimes determine (or dominate) the experienced taste/flavour (see also [33] and [135], for a developmental angle to sensory dominance in the world of multisensory flavour perception). Complete visual dominance is by no means always the outcome that has been observed in the literature. There is sufficient evidence to show that adding/changing the colour of a food or drink can sometimes result in the experience of flavours that are neither present in the food nor drink itself, nor necessarily associated with the colour, but emerge from the integration of visual flavour expectations with chemosensory cues [188]. Wang and Spence demonstrated the impact of adding colour in a study on the perceived aroma/flavour of a white wine that had been artificially coloured to look like a rosé wine when compared to the ratings given for unadulterated white and rosé wines (see Figure 1, for a summary of the results). The study was conducted in a large group of participants (N = 168) having different levels of expertise in the world of wine. The participants tasted three wines—a white wine, a rosé wine, and the white wine dyed to match the rosé (the fake wine in Fig. 1). The participants selected three aroma and three flavour descriptors from a list of descriptors. They also rated their liking for the wine, flavour intensity, and how difficult they found it to describe each wine. Linguistic analysis of the results revealed that those with wine tasting experience judged the miscoloured wine to be much more similar to the rosé than to the white wine, even though the fake wine and the white wine were one and the same. As can be seen in Fig. 1, red fruit descriptors were attributed to both of the rosé-coloured wines, especially in terms of flavour. Quantitative ratings revealed that the miscoloured wine was liked less than either the white or rosé wines. The participants also found it more difficult to describe the miscoloured (or fake) wine than they did to describe the rosé wine. Notice how tropical fruit aromas were only really perceived in the miscoloured wine condition. Such ‘emergent’ responses can therefore be taken to suggest that the flavours we perceive result from some sort of combination (or integration) of the unisensory cues. However, understanding the nature of that combination is by no means straightforward.

Given the rise of Bayesian causal inference as an explanatory framework for understanding multisensory integration in the spatial senses (namely vision, audition, and touch; [36]), it is natural to ask whether the same approach can perhaps be used to explain colour’s influence over taste/flavour perception. (See also [72] on the use of information theory to try and model multisensory flavour perception.) According to the simplest version of such frameworks (e.g., Maximum Likelihood Estimation), multiple cues each have their own sensory reliabilities that then determine the relative weights of each source of information in the final percept/decision (e.g., [3, 37]). (Note that two of the assumptions underlying the MLE approach are that the noise in each of the sensory channels is independent and has a Gaussian distribution. However, Marks and colleagues have argued that, in the case of olfaction and gustation, the noise is pooled across the flavour senses [72].) Recent work [70] has demonstrated that this approach to modelling the combination of olfactory and gustatory cues can be successfully used to predict flavour choice (i.e., preference) in rats. In humans, meanwhile, those combinations of olfactory and gustatory cues that commonly co-occur in the foods that we consume have been shown to exhibit enhanced integration, including increased ‘oral referral’ [66, 67] (see also [140], for a review), as well as a greater crossmodal, or multisensory enhancement of smell by taste [17, 27]. However, there have not yet been any published papers that take a computational approach to colour’s influence on flavour perception.

It is, however, important to stress that the crossmodal influences of orthonasal olfaction and vision on multisensory flavour perception would appear to be qualitatively different [7]. In part, this may be because of the somewhat confusing interaction/overlap between orthonasal and retronasal olfaction [106, 193]. While strictly-speaking orthonasal olfactory cues help to set flavour expectations (i.e., rather than being part of flavour experience), there is evidence that such cues are nevertheless sometimes integrated into multisensory flavour experiences (see [150], for a review).

By including priors, Bayesian models can accommodate the role of expectations. For example, when the brain processes information from different senses, assumptions about whether the sources arise from a common cause can be modelled as a prior [59]. When information from, say, sound and vision seem to be plausibly related (e.g., one hears a voice at the same time that one sees a moving mouth), the brain may be more likely to integrate that information than when they seem unlikely to be related (e.g., one hears an alarm clock while viewing a moving mouth). Spatiotemporal coincidence can also be treated as a prior, at least when combining cues/information from the spatial senses (e.g., [23, 79]; though see also [119]). According to Bayesian models, the likelihood that different sensory inputs (such as a given colour and a particular taste/flavour) belong together can be represented as ‘coupling priors’, sometimes referred to as a ‘prior of common cause’ (e.g., [35, 110, 117]). These priors are combined with the likelihoods of each input. As such, the Bayesian approach provides one means of formalizing different degrees of certainty regarding the unity of two or more unisensory inputs as a continuous (rather than as a discrete) variable (e.g., [59]). For the spatial domain, this is relatively straightforward to imagine—the individual sensory inputs for visual and auditory locations would each have their own uncertainty, and the coupling prior would represent the probability that the two inputs shared a common cause. If the uncertainties in the individual estimates are relatively high, and the coupling prior is strong, people are likely to perceive a unified audiovisual signal, even if the two signals come from different locations (as with the ventriloquism effect); however, if the uncertainties are smaller, then a spatial discrepancy can become more perceptible (consistent with the ‘degree of discrepancy’ account [120]). So it is the relative uncertainties in the coupling prior and the individual sensory likelihoods that determine whether integration or segregation is more likely to occur.

So, how might one conceptualize coupling priors in the case of multisensory flavour perception [23]? Given the large diversity in flavour experiences across both culture and over time [50], it would seem unlikely that people have evolved highly specific innate associations between particular colours and tastes/flavours. The one possible exception here might relate to the tendency to match more saturated colours with a more intense taste/flavour; one might think of this as an example of an innate correspondence based on intensity matching [133, 169]. However, it is worth noting that people may also have internalized a statistical regularity that is perhaps also present in manufactured food products (i.e., foods with a more intense taste/flavour are often given a more saturated colour in the marketplace), thus making the origins of any such intensity/saturation mapping hard to determine definitively.

Because cognitive neuroscience methods typically require participants to remain quite still, there are particular challenges associated with attempting to conduct ecologically-valid experiments with food. For example, in functional magnetic resonance imaging (fMRI), typical flavour stimuli consist of liquids, to avoid chewing and minimize swallowing, presented with controlled timing and temperature [18], often washed down with a squirt of artificial saliva! All the practical challenges described above apply, with the addition of confined movement for both fMRI and electro-encephalography (EEG), and a very restricted space for the participant to lie in the fMRI machine itself.

As this review of the literature has hopefully made clear, colour affects taste/flavour perception in a multitude of ways. There is converging evidence that the colour of food and drink helps set taste/flavour expectations and that these expectations, which would appear to be generated in a fairly automatic manner prior to tasting, bias the subsequent multisensory flavour experience. There is substantial empirical evidence that colours are associated with specific flavours, and that these associations are mostly learned as a result of the internalization of the statistics of the environment. What is more, the research shows that there are perceptual consequences of these crossmodal associations, provided at least that the degree of discrepancy between the flavour expectations set by colour does not differ too much from the actual flavour.

At a more abstract level, we do not yet have the right kinds of data, nor do we understand the dimensionality of flavour space. However, as both modelling and neuroscientific methodologies advance, we are optimistic about the insights to be made in coming years. It is possible that machine-learning and the rapidly growing field of computational gastronomy may also deliver some relevant insights concerning the way in which flavour space should be conceptualized (see [151]). Given the success of Bayesian causal inference in other areas of multisensory perception research, it is to be hoped that the same approach can be fruitfully applied to explain/predict colour’s influence over multisensory flavour perception. As highlighted earlier, this will likely require the collection of extensive psychophysical datasets, involving unimodal and multisensory judgments of expected/actual taste/flavour, preferably with the unimodal judgments obtained at different levels of reliability. An individual’s subjective degree of confidence in the meaning of colour may also provide useful information (cf. [102]).

Perhaps unsurprisingly, despite all the research on colour’s influence over taste and flavour that has been conducted over the last 80 years or so (e.g., since Moir’s early study [74]), there are still a number of questions that we still do not have a satisfactory answer to. These include the following: Is the crossmodal influence of colour over taste mediated by flavour? What, exactly, is the relation between the strength and ambiguity of colour-flavour mappings? Under what conditions does product extrinsic (and/or intrinsic) colour affect taste/flavour perception? Is it possible to model colour’s influence over multisensory flavour experience in terms of Bayesian causal inference? Answering such questions will likely help in food product development (and experiential marketing) in the years ahead (e.g., [2, 15, 73]).