The predictive coding model of brain function is a deeply important development for neuroscience, and Andy Clark does the field a service with this careful, thorough, and accessible review. We are concerned, however, that Clark's account of the broad implications of model – and in particular his attempt to turn it into a Grand Unified Theory (GUT) of brain function – may be at least four dogmas of empiricism out-of-date (Anderson Reference Anderson2006; Chemero Reference Chemero2009; Davidson Reference Davidson1974; Quine Reference Quine1951). Clark's adoption of a thoroughgoing inferential model of perception, his neo-neo-Kantian view of the relationship between mind and world, and his insistence that every sensory modality operates according to the same underlying causal-epistemic logic – all (individually and severally) threaten to return us to the bad old days of epistemic internalism (e.g., Rorty Reference Rorty1979) that the field, including the author of Being There (Clark Reference Clark1997), rightly left behind.

Here we suggest that Clark (although not he alone) has made an error in conflating different senses of “prediction” that ought to be kept separate. The first sense of “prediction” (henceforth prediction₁) is closely allied with the notion of correlation, as when we commonly say that the value of one variable “predicts” another (height predicts weight; education predicts income; etc.). Prediction₁ is essentially model-free, and it comes down to simple relationships between numbers. The second sense of “prediction” (prediction₂), in contrast, is allied instead with abductive inference and hypothesis testing. Prediction₂ involves such cognitively sophisticated moves as inferring the (hidden) causes of our current observations, and using that hypothesis to predict future observations, both as we passively monitor and actively intervene in the world. It is theory laden and model-rich.

We have no trouble believing that a fundamental part of our exquisite attunement to environmental contingencies involves sensitivity to (and the ability to make use of) inter- and cross-modal correlations in sensory signals. Sensitivity to temporal and spatial (e.g., across the retina) correlations could underwrite many functional advantages, including the ones Clark highlights, such as reducing sensory bandwidth and drawing attention to salient departures from expectations. In this sense we share Clark's belief that predictive₁ coding is likely to be a ubiquitous and fundamental principle of brain operation; neural nets are especially good at computing correlations.

However, we don't think that evidence for predictive₁ coding warrants a belief in predictive₂ coding. And it is only from predictive₂ coding that many of Clark's larger implications follow.

Clark makes the move from predictive₁ coding to predictive₂ coding largely by relying on an innovative account of binocular rivalry offered by Hohwy et al. (Reference Hohwy, Roepstorff and Friston2008). In Clark's somewhat simplified version of their proposal, the experienced alternation between seeing the face stimulus presented to one eye and the house stimulus presented to the other is explained by a knowledge-driven alternation between rival hypotheses (face at location x, house at location x) neither of which can account for all of the observations. According to Clark, the reason the images don't fuse and lead to a visual steady-state is because we know that faces and houses can't coexist that way. If this knowledge-driven account is the correct way to understand something as perceptually basic as binocular rivalry, then predictive₂ coding can begin to look like a plausible, multilevel and unifying explanation of perception, action and cognition: perception is cognitive and inferential; inference perceptual; and all of it is active.

But while the predictive₂ coding model of binocular rivalry may be consistent with much of the data, it is far from the only possible explanation of the phenomenon. Here is an outline of a reasonable predictive₁ coding account: Given the generally high-level of cross-correlation in the inputs of our two eyes, the left eye signal would predict₁ greater correlation with the right eye than is currently in evidence; this would weaken the inputs associated with the left eye, unmasking the inputs associated with the right eye, which would predict₁ cross-correlated left eye signals . . . and so on. However far this particular proposal could be taken, the point is one can account for the phenomenon with low-level, knowledge-free, redundancy-reducing inhibitory interactions between the eyes (see, e.g., Tong et al. Reference Tong, Meng and Blake2006). After all, binocular rivalry also occurs with orthogonal diffraction gratings, indicating that high-level knowledge of what is visually possible needn't be the driver of the visual oscillation; humans don't have high-level knowledge about the inconsistency of orthogonal gratings. In general, although not every pair of stimuli induce bistable perceptions, the distinction between those that do and those that don't appears to have little to do with knowledge (see Blake [Reference Blake2001] for a review). Adopting a predictive₂ coding account is a theoretical choice not necessitated by the evidence. It is hardly an inconsequential choice.

Using predictive₂ coding as a GUT of brain function, as Clark proposes, is problematic for several reasons. The first problem is with the very idea of a grand unified theory of brain function. There is every reason to think that there can be no grand unified theory of brain function because there is every reason to think that an organ as complex as the brain functions according to diverse principles. It is easy to imagine knowledge-rich predictive₂ coding processes employed in generating expectations that we will confront a jar of mustard upon opening the refrigerator door, while knowledge-free predictive₁ coding processes will be used to alleviate the redundancy of sensory information. We should be skeptical of any GUT of brain function. There is also a problem more specific to predictive₂ coding as a brain GUT. Taking all of our experience and cognition to be the result of high-level, knowledge-rich predictive₂ coding makes it seem as if the world that we experience and think about is a projection of our minds. Western philosophy has been down this lonely and unproductive road many times. It would be a shame if the spotlight that Clark helpfully shines on this innovative work in neuroscience were to lead us back there.