Biology & Philosophy

, Volume 29, Issue 5, pp 731–745 | Cite as

A critique of the principle of cognitive simplicity in comparative cognition



A widespread assumption in experimental comparative (animal) cognition is that, barring compelling evidence to the contrary, the default hypothesis should postulate the simplest cognitive ontology (mechanism, process, or structure) consistent with the animal’s behavior. I call this assumption the principle of cognitive simplicity (PoCS). In this essay, I show that PoCS is pervasive but unjustified: a blanket preference for the simplest cognitive ontology is not justified by any of the available arguments. Moreover, without a clear sense of how cognitive ontologies are to be carved up at the joints—and which tools are appropriate for the job—PoCS rests on shaky conceptual ground.


Comparative cognition Parsimony Complexity Animal cognition Cognitive evolution 


A widespread assumption in experimental comparative (animal) cognition is that, barring compelling evidence to the contrary, the default hypothesis should postulate the simplest cognitive ontology (mechanism, process, or structure) consistent with the animal’s behavior. I call this assumption the Principle of Cognitive Simplicity, or PoCS. In this essay, I show that PoCS is pervasive but unjustified: a blanket preference for the simplest cognitive ontology is not justified by any of the available arguments. Moreover, without a clear sense of how cognitive ontologies are to be carved up at the joints—and which tools are appropriate for the job—PoCS rests on shaky conceptual ground.

PoCS assumes that for a given behavior, the possible cognitive systems driving the behavior can be ordered by complexity. Crucially, the notion of a scale of cognitive complexity guides scientific decisions despite the fact that concepts like complexity and cognition are often left vague or ambiguous. Some researchers take a catholic view of cognition, on which any information-processing activity that mediates between an environmental or internal input and a behavioral output is cognitive (Shettleworth 2012; Heyes 2012). Others prefer a more restricted view, on which only those systems that produce behaviors consistent with a particular view of rationality are meaningfully cognitive (Dickinson 2012). In this essay, I will use the term, cognitive, in the broadest possible sense in order to avoid committing myself to any single distinction between the cognitive and the non-cognitive. Adding to the confusion in the sciences, the concepts of rationality and other supporting concepts, such as representation and intentionality, are themselves too commonly left vague or ambiguous. Although it is standardly assumed that a representational system is more complex than a nonrepresentational system, few researchers back up this assumption with a sustained argument (e.g., Zentall 2001). Comparative cognition researchers cope with the confusion by labeling some features of cognition as simpler than others. For instance, putatively simpler processes include associative processes, while most allegedly representational activities, such as reasoning and certain kinds of memory, are treated as complex.1

The pervasiveness of PoCS-like thinking can be explained historically. Animal cognition researchers have long been wary of an illicit overattribution of complex (sometimes “human-like”) abilities and faculties to nonhuman animals. This worry has been enshrined in a misinterpretation of Lloyd Morgan’s famous “canon.” The canon states that:

(MC) In no case may we interpret an action as the outcome of the exercise of a higher psychical faculty if it can be interpreted as the outcome of the exercise of one which stands lower in the psychological scale. (Morgan 1894, 53)

A common misinterpretation of MC holds that it offers an Occamist parsimony principle for comparative psychology. Despite numerous rebuttals of this interpretation (e.g., Thomas 2001; Sober 2005; Fitzpatrick 2008), the parsimony interpretation continues to be widely cited. For example, Elsevier’s Dictionary of Psychological Theories (2006) states:
$$\begin{aligned} {\text{PARSIMONY}},\;{\text{LAW/PRINCIPLE}}\;{\text{OF}}. & = {\text{Lloyd}}\;{\text{Morgan's/Morgan's}}\;{\text{canon}} \\ & = {\text{Occam's}}\;{\text{razor}} \\ & = {\text{Occam's}}\;{\text{principle}} \\ & = {\text{economy}},\;{\text{principle}}\;{\text{of }}\left[ \ldots \right] \\ \end{aligned}$$
Highly respectable researchers continue to use modified versions of MC to caution students to err on the side of cognitive simplicity. In the most recent edition of her textbook on the principles of comparative cognition, psychologist and biologist, Shettleworth (2012) writes:

A reasonable modern interpretation of Morgan’s Canon is that explanations in terms of general processes of learning along with species-typical perceptual and response biases should always be sought before invoking more complex or specialized cognitive processes. (Shettleworth 2012, 11–12; emphasis added)

Note that Shettleworth assumes that less complex processes should not be postulated where presumably simpler process can explain the observed behavior. Like Shettleworth, other modern versions of MC similarly replace the problematic “higher/lower” distinction with variants of the “complex/simple” distinction. These modern versions drive hypothesis choice across comparative cognition, as I show next in “PoCS in action: three case studies”. Then, in “Defenses of PoCS”, I evaluate three arguments for PoCS, rejecting all three.

PoCS in action: three case studies

PoCS is best explained through illustration. The following three case studies will each examine how PoCS functions in a range of experimental contexts. In each of these case studies, explanations imputing the simplest cognitive ontology are given preference.

Metacognition in rats

Crystal and Foote (2009) critique the conclusion of their 2007 study, which used a duration-discrimination task to test for awareness of uncertainty among rats. In the 2007 study, rats were presented with audio tones of different durations and trained to classify the tones into the categories of “short” or “long.” The rats were then presented with a range of tones, some clearly short and others ambiguous. Correct responses were rewarded with food, and incorrect responses were not rewarded at all. Next, the rats were given the same test, but were given a third option: to decline a test. Declined tests allowed the subjects to move on to more tests with the prospect of getting more food. When given the choice to decline tests, the rats consistently opted to decline the ambiguous (“harder”) tests but not the unambiguous tests, even though declining a test resulted in a smaller food reward than answering correctly. Moreover, the overall accuracy improved when rats were allowed to opt out of difficult tests. Foote and Crystal (2007) concluded that the rats were aware of their own uncertainty. However, Crystal and Foote (2009) argue that because the behaviors can be explained in terms of associative processes, their previous metacognitive explanation was unwarranted. They write:

Importantly, it is necessary to rule out simpler, alternative explanations. … For example, if principles of associative learning or habit formation operating on a primary representation may account for putative metacognition data, then it would be inappropriate to explain such data based on metacognition…the burden of proof favors primary representations, by application of Morgan’s canon. (2009; emphases added)

In this case, “primary” representations are representations of features in the world, while “secondary” representations are representations of other representations.2 Crystal and Foote (2009) assume a cognitive hierarchy on which association is simpler than representation, and “primary” representations are simpler than “secondary” representations. However, as is common practice, Crystal and Foote do not explain what such complexity entails. Nor do they define “representation.” It is therefore unclear whether these researchers have in mind what David Pitt calls the broad sense of representation common to philosophy, on which representations are understood to be “mental objects with semantic properties” or the computational sense common to cognitive science, on which mental representations are “information-bearing structures” (Pitt 2012). On neither interpretation is it clear, nor is it explained, why a representation of a representation should be more complex than a representation of a state in the environment, and why this should be epistemically relevant. This example illustrates how a vague notion of cognitive complexity occasionally drives experimental research by presuming that, e.g., associative processes are simpler than metacognitive ones and then favoring associative explanation.3

Problem-solving and insight among New Caledonian crows

New Caledonian crows exhibit what appears to be the ability to plan. In one set of studies, they successfully chose a series of sticks of varying lengths in the order required to obtain a trapped food reward. A crow may be presented with three cages, one containing a food reward that could only be reached with a long stick, one containing a long stick that could only be reached with a short stick, and one containing a short stick that could be obtained with a beak. If the crow was capable of planning, it would be expected to observe the setting and then proceed by first reaching into the cage with the short stick, using that to reach the long stick, and finally using the long stick to reach the food. Researchers agree that, among available options, the planning hypothesis is the most extravagant, even if it turns out to be true. For example, Wimpenny et al. (2009), write:

While the ability of [some New Caledonian crows] to use three tools in sequence reveals a competence beyond that observed in any other species, our study also emphasizes the importance of parsimony in comparative cognitive science: seemingly intelligent behaviour can be achieved without the involvement of high-level mental faculties, and detailed analyses are necessary before accepting claims for complex cognitive abilities. (Wimpenny et al. 2009, 1; emphasis added)

Once again, the assumption that cognitive “faculties” and “abilities” can be ordered by complexity drives the researchers to shift the burden of proof to the planning explanation. Because planning is assigned to the category of “high-level” and “complex” mental faculties, Wimpenny et al. are cautious to exclude all other explanations first. This caution continues throughout the discussion of their experiments, such as in the following conclusion:

In our opinion, claims for analogical reasoning based upon sequential tool use remain unjustified…, and using sequential tool use as a benchmark of this ability is inappropriate. Reasoning (let alone analogical reasoning) is not the only cognitive mechanism to account for sequential tool use: simpler processes such as chaining may be sufficient. (Wimpenny et al. 2009, 11–12; emphasis added)

Similarly, in their study of New Caledonian crows, Taylor et al. (2012) address the availability of alternative explanations for the crows’ facility with a different puzzle—the string-pulling puzzle. For this task, crows were presented with a reward that was tied to a dangling string. The correct solution required the crows to pull the string with its beak, step on the coiled string with its foot, and repeat until the reward was attained. Taylor et al. (2012) conclude that the apparently planned string-pulling behaviors can—and therefore ought to—be explained by (associative) perceptual feedback rather than what they call “insight,” which presumably requires representational abilities.

Simulations of complex behaviors using simple rules

Finally, PoCS provides some evidentiary weight for computer simulations of animal behavior. Consider van der Vaart et al.’s (2012) simulation of theft-aversion behavior in Western Scrub jays, a species of corvid known for its extensive food caching in the wild. Studies have shown that these birds are able to recall a surprising amount of information about their caches, such as location of specific caches, time of caching, and rate of decay of food types (Clayton and Dickinson 1998). Follow-up studies revealed that the jays re-cached more when they were conspicuously observed by other jays (Dally et al. 2006). Not only do jays react to being watched, but the study found that their responses were mediated by personal experience with pilfering: jays who had experience as pilferers were more likely to re-cache in the presence of others (ibid). These findings suggested that the jays may be attributing the intention to pilfer to the watchers, which, in turn, suggests that the jays may have been predicting the behavior of others based on personal experience. In other words, they appeared to have some features of a theory of mind.

Critiquing this theory of mind hypothesis, another group, van der Vaart et al. (2012) write that one “reason to favor [a] ‘prior [associative] learning hypothesis’ is that it seems cognitively simpler than ‘theory of mind’; it does not require corvids to be capable of mental state attribution” (van der Vaart et al. 2012, 1). However, they quickly note that the prior associative learning strategy may carry a heavy cognitive load, as “birds would have to remember distances between cache sites and onlookers, and also relate these distances to pilfering rates” (van der Vaart et al. 2012, 2). They concede that this would require “cognitive complexity” as well. As an alternative, they propose a model of jay caching behavior based on one rule: cache more when stressed. Van der Vaart et al. simulated jay re-caching behaviors using this single-variable stress model, and were able to duplicate jay re-caching behavior. Although the researchers admitted that such simulations were only “how-possibly” explanations of behavior, they nevertheless treat the positive results of their simulation study as evidence against allegedly complex cognition.

Similarly, M.E. Le Pelley (2012) offers an associative model for rat behavior that had previously been considered evidence of metacognition:

… suppose that a suitably well-specified psychological theory of higherlevelmetacognition were developed; should we then prefer the associative or the higher level account? The available behavioral evidence cannot decide between the two, and hence, the conservative answer would be to fall back on C. Lloyd Morgan’s (1903) canon … (Le Pelley 2012; emphasis added)

Although Le Pelley uses the term “conservative” rather than “parsimonious,” his reference to MC reveals a commitment to PoCS. As in the other examples, the burden of proof is shifted to the metacognitive account, because the default hypothesis is the “low level,” putatively simpler account. In this case, that simple account is associative.

In addition to simulation-studies, cognitive models can exhibit PoCS-like commitments. Cruse and Wehner (2011), for example, propose a cognitive model that they claim can account for the navigational behaviors of ants and bees using association-formations alone, without relying on a cognitive map. “Cognitive maps” are representations of an organism’s environment, and serve as one type of explanation for the navigational abilities of insects: the insect identifies features of its environment and compares it against a cognitive map of the known terrain, which allows it to generate a flight or walking path. On their model, information from disparate systems within a given insect is processed independently to produce apparently coordinated behavior. On the cognitive map model, information from disparate systems is integrated into a central representation, or a “cognitive map,” which produces coordinated behavior. Cruse and Wehner conclude that, since both models capture the behavioral phenomena equally well, but the cognitive map model contains an additional component (the central processing system), there is “no need for a cognitive map.” Their title reflects this conclusion: “No Need for a Cognitive Map: Decentralized Memory for Insect Navigation.” The phrase, “no need,” is an expression of the principle that cognitive complexity should only be posited when simpler mechanisms cannot account for the behavior—i.e., an expression of PoCS.

As these case studies suggest, PoCS is so common that researchers appear to take for granted that it can be justified. Meanwhile, its impact is broad: PoCS restricts inference from experimental data and gives the results of simple how-possibly computer models the status of evidence against allegedly complex alternatives. A justification for PoCS is, therefore, required.

Defenses of PoCS

In this section I consider three defenses of PoCS, which I call (1) the Taxonomic Ubiquity Argument, (2) the Systems Argument, and (3) the Energetic Cost Argument. All appeal to evolutionary theory to ground PoCS, and each is problematically speculative.

The taxonomic ubiquity argument

The Taxonomic Ubiquity Argument holds that simple cognitive systems, such as associative systems, are the most likely to underlie behavior because they are the most taxonomically widespread. Their ubiquity suggests that they are ancient and that cognitive simplicity is evolutionarily conserved. This view is so pervasive that it can be found in textbooks introducing comparative cognition. For example in her 2009 textbook, Shettleworth writes that:

A reasonable modern interpretation of the Canon … is that a bias in favor of simple associative explanations is justified because basic conditioning mechanisms are widespread in the animal kingdom (Shettleworth 2009, 18; emphasis added)

In other words, since earlier studies have explained a wide range of behaviors in a wide range of organisms in terms of associative mechanisms, we may conclude that associative mechanisms are responsible for whatever new behavior we encounter in the future. Note that Shettleworth takes associative systems to be among the simplest cognitive systems—an assumption that makes her argument a version of PoCS.

However, the apparent ubiquity of associative systems cannot ground PoCS, or the blanket presumption in favor of simple systems. First of all, even if associative systems are more widespread than other cognitive systems, this fact would not ground a general preference for simple systems.4 This is because the conclusion that (putatively simple) associative systems are widespread is itself built on a mass of data from experiments that have taken association as the default hypothesis. These experiments are both behavioral and neuroanatomical.5 While it is infelicitous to imagine that all the experiments that Shettleworth has in mind were biased, it is at least likely that their results are open to precisely the kind of alternative interpretations that Shettleworth’s suggestion aims to forestall. Furthermore, this line is question-begging if what we wish to know is whether behaviors that can be explained using, e.g., associative concepts, are in fact produced by, e.g., associative systems. Lastly, while it is true that associative systems are very common, they are also varied, and many are not simple in any straight-forward way. Moreover, as Buckner (2011) argues, it remains to be seen whether associative systems are incompatible with complex cognition. So, the ubiquity of associative systems does not support PoCS.

The systems argument

Another argument, which I call the Systems Argument (SA), offers a reason for the supposed taxonomic ubiquity of putatively simple cognitive systems and the relative taxonomic rarity of more complex systems. The Systems Argument holds that animal behavior is controlled by multiple cognitive systems, and that systems producing complex cognition, such as reasoning and representing, are relatively recent—and thus rare—evolutionary developments (Karin-D’Arcy 2005; Shettleworth 2012).6 The SA relies on a story about the evolution of cognitive systems, best illustrated by Karin-D’Arcy (2005). On her view, complex biological systems evolve by means of adding new systems on top of older systems in a hierarchical yet integrated fashion. To use one of her examples, the marine mammal’s respiratory system includes both the ancient air-breathing system and a new and specialized system for maintaining oxygen levels under water (2005, 189). The older systems do not get replaced as newer systems are added, but are, instead, integrated into the new “hierarchically structured integrated system” (ibid). Cognitive systems, being biological systems like any other, are, on her view, likely to follow the same evolutionary trajectory. If this story is true, then some ancient components of cognitive systems should be shared across a wide range of taxa, with each taxon containing additional, derived specializations. Shettleworth (2013) recommends modifying MC to fit this conclusion, writing:

A reasonable modern interpretation of Morgan’s Canon is that explanations in terms of general processes of learning along with species-typical perceptual and response biases should always be sought before invoking more complex or specialized cognitive processes. This stance is justified by the fact that the simple forms of learning such as habituation and classical conditioning…are very widespread in the animal kingdom. (Shettleworth 2013, 12; emphasis added).7

She concludes that, “the burden of proof is on anyone proposing that some novel, additional, cognitive mechanism has arisen on a particular branch of the evolutionary tree” (Shettleworth 2013, 12–13).

Similarly, Karin-D’Arcy (2005) suggests her own alternative modernization of MC, which aims to bring it up to date with contemporary biology by “replace[ing] the notion of a psychological scale with the concepts of ancestral and derived psychological processes” (Karin-D’Arcy 2005, 196). It should be noted that the concepts of “derived” and “ancestral” are independent of the notion of complexity. However, Karin-D’Arcy seems to suggest that her revised MC would issue in the same complexity-reducing result as the original parsimony-based reading of the MC. Although she disavows the parsimony-based reading of MC as well as the cognitive hierarchy implicit in the original MC, she nevertheless concludes that the revised “canon serves to balance the natural human tendency to interpret observed behavior in terms of complex psychological processes” (Karin-D’Arcy 2005, 190; emphasis added).8

On her view, the revised MC combines with her view of hierarchical cognitive evolution to explain Daniel Povinelli’s Reinterpretation Hypothesis (RH), on which both humans and other primates use the same (homologous) systems to produce most behaviors, but humans use a specialized, derived, module to reinterpret the behaviors of other humans in intentional terms. Povinelli and Barth (2005) characterize RH as follows:

The reinterpretation hypothesis posits that the ancestor of the ape/human group possessed a suite of systems dedicated to representing and reasoning about behavior (detailed in Povinelli &Vonk 2004), but not intentions or other mental states. Further the model posits that, at some point in the evolution of the human lineage (probably coincident with evolution of natural language), a new system for encoding the behavior of self and other in terms of mental states was grafted into these ancestral systems for representing and reasoning about behavior. (Povinelli and Barth 2005, 713)

On the RH view, humans are unique among primates (and all other animals) because they utilize a special language-based module for decoding the behaviors of others in terms of mental states. Karin-D’Arcy builds her revised MC on this view, writing that,

In the case of social cognition, it is likely that the same ancestral system produces social behavior in humans and other primate species, but observed behaviors are interpreted by different systems, with humans primarily using the specialized intention attribution system to make these interpretations (Karin-D’Arcy 2005, 189; original emphasis).

Based on the RH, Karin-D’Arcy argues that the human ability to impute beliefs and desires to others is a derived human specialization, not shared with other primates. This ability, commonly known as theory of mind, is presumed to require significant cognitive complexity. Thus, the human behavior-analysis system is, on this view, more complex than the nonhuman behavior-analysis system.

One objection to Karin-D’Arcy’s view can be drawn from the work of Elliott Sober (2005). According to Sober, the principle of cladistic parsimony cautions against the conclusion of the revised MC.9 If we wish to know whether an existing animal species possesses a cognitive feature that humans are known to possess and which we know to be derived (relative to the animal under investigation), or whether it merely possesses the ancestral trait, and if no additional information regarding common ancestry is available, then using cladistic parsimony leads to the conclusion that both inferences are equally parsimonious (Sober 2005). Take the example of puzzle-solving in crows and humans: if all that is known is that humans solve the puzzle through planning, then it is as parsimonious to assume that crows also use planning to solve the puzzle as it is to assume that crows use some other mechanism to solve the puzzle. However, Sober argues, if it is known that both the target species—let us say, crows—and humans exhibit the same behavioral trait, B, and B is known to be a shared derived trait (i.e., the ancestor of humans and crows did not exhibit B), then parsimony recommends treating B as homologous10 rather than as homoplastic.11 But, under the rules of cladistic parsimony, if B is a homology then, ceteris paribus, the underlying (proximate) mechanisms in both humans and crows should be the same. Therefore, according to Sober, replacing “higher” and “lower” with “derived” and “ancestral,” respectively, would recommend attributing the same underlying mechanisms whenever two species display homologous behaviors—precisely the opposite of the conclusion Morgan intended and contrary to Karin-D’Arcy’s conclusion.

Sober’s objections have limited purchase, however, since cladistic parsimony, as Sober admits, is a tool of limited utility. For instance, it may turn out that homologous traits may be underwritten by homoplastic or even divergent mechanisms.12 However, Karin-D’Arcy’s modified MC and the Systems Argument faces another problem: it appears to neglect the role of convergence. She would be right if the derived system in, e.g., humans, is unlikely to have been duplicated (through convergent evolution) in other animals. This would suggest that elements of complex human cognition may be unique. There is no reason to suppose that, even under similar environmental constraints, two distantly related species cannot develop functionally homoplastic cognitive systems. In fact, functional homoplasies are exceedingly common (Conway 2003)—bat wings and butterfly wings being only one (popular) example. Why, then, should cognitive function be exempted? By Karin-D’Arcy’s own logic, it cannot: she argues that cognitive systems should develop just as any other complex system. In fact, there is strong behavioral evidence of convergence between the mammalian and avian cognitive mechanisms (i.e., the underlying processes that produce similar behaviors), despite vastly different brain tissue organization (Emery and Clayton 2004).

Moreover, using the Systems Argument threatens to paper over important complexities in the evolution of complex systems. To give just one example, consider the case of homologous and homoplastic features in two very distant taxa: birds and mammals, which diverged between 312.3 million and 330.4 million years ago. As these taxa diverged, the structural organization of their neuroanatomy diverged as well. Despite these differences, however, behavioral evidence from multiple species of corvid, such as crows and jays, and parrots, such as African Grey parrots, suggests that some birds are capable of remarkable cognitive sophistication. As Emery and Clayton (2004) write, “cognition in corvids and apes must have evolved through a process of divergent brain evolution with convergent mental evolution” (Emery and Clayton 2004, 1907). That is, while the neuroanatomies of birds and mammals evolved separately, the cognitive abilities, and the mechanisms underpinning those abilities, appear to have converged.

Neuroscientists believe this cognitive sophistication to be the product of the brain regions called the nidopallium and mesopallium, comprising the dorsal ventricular ridge (DVR). The DVR is colloquially referred to as the “avian prefrontal cortex,” indicating its functional analogy with the mammalian prefrontal cortex. This functional analogy, combined with the structural divergence, suggests that the DVR and the mammalian prefrontal cortex are homoplasies (Güntürkün 2012). However, the full story is more complicated. Recent research has shown that cells in the DVR are homologues of the cells in the mammalian neocortex—that is, that the cells in avian DVRs and mammalian neocortices share a common ancestor (Dugas-Ford et al. 2012). Although the cells are organized into very different structures in the two taxa, they appear to carry out similar functions—that of supporting complex cognition. This raises the question of whether the DVR is a homologue or analogue (homoplasy) of the mammalian prefrontal cortex. It is also possible that while the overall structure is homoplastic, some of the active components of the two structures are homologous.

Returning to Karin-D’Arcy’s Systems Argument, we can see that the advice of the revised MC would be misleading: Given similar behaviors by, e.g., humans and chimps when it comes to puzzle solving, the revised MC recommends erring on the side of explaining the chimps’ behavior in terms of ancestral features common to humans and chimps—i.e., homologous features. Therefore, not only does the Systems Argument ignore homoplasies, but it glosses over the fact that the same structure or ability can be homoplastic in one respect and homologous in another. For instance, the avian DVR is homologous with the mammalian neocortex at the cellular level, but homoplastic at the functional level. In summary, the SA for PoCS relies on an oversimplified view of the evolution of complex systems, neglecting both convergence and the variety of ways that systems can be homologous or homoplastic. It is no wonder that, as de Waal and Ferrari (2010) write, “Scala Naturae assumptions remain prevalent enough [in comparative cognition] that cognitive similarities between distant taxa, such as birds and primates, are sometimes viewed as antithetical to evolutionary theory” (de Waal and Ferrari 2010, 201).

The energetic cost argument

The final argument for PoCS holds that there is a fitness tradeoffs between the costs associated with metabolically expensive cognitive machinery and the benefits that accrue to cleverer organisms. All else being equal, selection should be expected to favor less expensive cognitive machinery over more expensive machinery for the production of a given behavioral trait. Most organisms are, on this view, expected to use simple, metabolically inexpensive, cognitive machinery. Because “natural selection often results in the evolution of less complex cognitive structures,” it is rational to expect that the cognitive systems responsible for the observed behaviors are simple (Mameli and Bortolotti 2006, 87).

This argument proceeds as follows: Sophisticated cognitive abilities require sophisticated cognitive machinery, i.e., brain structures. However, increases in brain sophistication—understood as increases in size, density, or activation—require increases in energy consumption. More sophisticated brains are not only more metabolically more expensive than smaller brains, but pound-for-pound more expensive than other organs. Any increase in brain size will therefore be, ceteris paribus, the more metabolically costly evolutionary option. Since increases in metabolic needs under conditions of competition for limited resources, and barring compensating benefits, tend to reduce fitness, an increase in brain sophistication must be compensated for with some fitness-increasing ability. However, there are many less energetically costly means of increasing fitness, such as modifications made to more energy-efficient organs, or the development of less costly cognitive mechanisms capable of generating fitness-enhancing behaviors. Therefore, increases in brain sophistication will tend to be the least likely adaptive strategy. We should therefore expect very few instances of extant organisms with highly developed brains and, by extension with highly developed cognitive functions.

This argument runs into conceptual problems at nearly every step. First, it problematically assumes that cognitive sophistication correlates positively with increased energetic (metabolic) costs. Certainly, larger, denser, and more active brains carry larger metabolic costs. However, not all evolutionary increases in behavioral sophistication may require changes to the physical structure of the brain. Some increases in behavioral flexibility, for example, may be exaptations in which already existing neural machinery is co-opted for a new function; others may be evolutionary by-products. For example, the human ability to drive did not require the development of new brain regions; instead, it required the repurposing of parts of an existing system. The same may be true of, e.g., metacognition in rats or planning in crows. Second, the argument fails to consider the possibility of developmental constraints on the kinds of cognitive architecture that is allowed to develop. PoCS recommends choosing the simplest cognitive system, but even if selection favored cognitive simplicity, extant cognitive systems may not be the simplest possible systems, but merely the simplest among those available to selection. Further, although one may object that the associative system was clearly available to selection, there is reason to doubt that associative cognitive architecture has lower built-in energetic costs. The view that associative systems are metabolically inexpensive has been challenged by a number of recent thinkers. For example, Gallistel (2008) argues that associative mechanisms would require far greater energy expenditures than alternative mechanisms. Gallistel uses the honeybee navigation system to argue that the honeybee brain does not have enough computing power to process information through associations alone, and must require a representational system of mental maps (ibid). Finally, this argument fails to account for PoCS in the case of the many animals whose neuroanatomy satisfies the conditions for complex cognition, unless it can be shown that the same brain uses less energy to perform, e.g., associative functions than it does to perform representational functions.

In summary, all three arguments for PoCS are inadequate. The Taxonomic Ubiquity Argument begs the question by assuming that, since many organisms have been shown to employ associative processes, their behavior cannot also be driven by alternative cognitive processes. The Systems Argument takes a controversial position regarding the evolution of complex systems, giving too little weight to convergence of function. In addition, it appears to conflict with cladistic parsimony. Finally, the Energetic Cost Argument takes a simplistic view of cognitive function by presuming that metabolic rates track cognitive complexity, and that metabolic rates are visible to selection.


I have argued that PoCS is both pervasive and, currently, unjustified. Although a justification for the blanket preference for cognitive simplicity is possible, given that one is not currently on offer, PoCS-based inferences should be avoided unless and until such a justification is offered. In other words, the hypothesis in a given experiment should not be the one that presupposes the simplest cognitive ontology. No justification is available for the blanket preference for a simplicity-based default or starting hypothesis across all experimental contexts, suggesting that starting hypotheses should be chosen on a case-by-case basis.


  1. 1.

    In fairness to the researchers, one may argue that many research projects do not require an explicit commitment to a single definition of cognition. For example, in narrowly scoped experiments, such as asking whether a Western Scrubjay appears to possess episodic-like memory (autobiographical event memory in non-human animals), it is not necessary to specify whether episodic-like memory is meaningfully cognitive (Clayton et al. 2001). What is required is not a general definition of cognition, but a clearly articulated strategy for identifying the behavioral markers of episodic-like memory. To put this point more generally, whether the hypothesized trait counts as cognitive is presumed to be a separate question from how to set up and evaluate experimental tests of that trait. However, these questions are not entirely separate, since the former plays a role in establishing what comes to count as the default hypothesis. Therefore, while the science “works” just fine without broad agreement regarding concepts central to its inquiry—just as biology works without a philosophically unassailable definition of life—the experimental programs in comparative cognition are shaped by PoCS.

  2. 2.

    They credit Carruthers (2008) with their conclusion that lower-level representations can accommodate their data (Crystal and Foote 2009, 1).

  3. 3.

    For another example of the PoCS assumption, see Foote and Crystal’s (2012) rejection of the metacognitive explanation in response to the development of a non-metacognitive model by Smith et al. (2008). They write, “Clearly, putative evidence for metacognition in rats is critically undermined when a non-metacognition model can produce the observed pattern of behavior” (Foote and Crystal 2012, 188). However, it is important to note that comparative cognition researchers, such as Foote and Crystal, may have other rational reasons than simplicity for favoring associative explanations. For example, models of associative learning are better developed than, e.g., metacognitive models. Devising tests that further refine such models is, therefore, possible, while refining metacognitive models is not possible at present. The availability of these models, moreover, makes associative hypotheses more predictive than the alternatives. While this reason has its own shortcomings, it is perfectly rational for scientists to prefer hypotheses that have generated well-specified models over hypotheses that have yet to be fully fleshed out and which continue to make only rough predictions. I develop this idea in my dissertation ([present author] 2013). For a discussion of pre-empirical preferences in comparative cognition see also Buckner (2013).

  4. 4.

    Shettleworth is not, I should note, suggesting that her argument is grounds for a general simplicity preference.

  5. 5.

    The Taxonomic Ubiquity Argument is based on more than behavioral data, as an anonymous reviewer helpfully notes: In fact, as the review notes, the ubiquity of associative-learning has been established by non-behavioral observations, such as evidence of physical mechanisms that can support associative-learning in such invertebrates as the Aplysia sea slug. Behavioral data is, of course, supplemented with neuroanatomical data, which may be used to further the Taxonomic Ubiquity Argument. In fact, the Aplysia example supports my conclusion rather than imperil it: the mere presence of a physical realizer of associative learning does not establish that only associative learning is responsible for the observed behavior. Such an inference successfully shows that associative learning is possible across a broad range of taxa, but not that other cognitive abilities are therefore likely to be absent, as the Taxonomic Ubiquity Argument holds.

  6. 6.

    To see why recent origins suggest rarity, consider the following simple example. Imagine a very simple tree with one ancestral node, A, representing the ancestral species, and ten branches representing ten new species, S1–S10. If a trait, T, is ancient, then it will be found in A and is, therefore, more likely to be found in S1–S10 as well (depending on how conserved T happens to be). However, if T is a recent development, i.e., if it arose in S7, then it would have to be recreated in the other nine branches as well, which is an unlikely scenario. Therefore, the recent origins of T strongly suggest its relative rarity.

  7. 7.

    The reader may note that Shettleworth’s argument in the quoted passage draws on the TUA as well as on the SA. In keeping with the theme of drawing out the arguments in favor of PoCS, I have elected to distinguish between these two arguments for the purposes of this essay.

  8. 8.

    Also of note is Karin-D’Arcy’s implication that the canon is an anti-anthropomorphism principle (Karin-D’Arcy 2005).

  9. 9.

    Sober is not responding to Karin-D’Arcy. His aim is to show that no re-interpretation of the Canon in evolutionary terminology will issue in the parsimony reading of the Canon that has become so popular. Along the way, Sober offers what he considers to be the best interpretation of Morgan’s actual intent (Sober 2005).

  10. 10.

    A homology is “a similarity inherited from a common ancestor” (Sober 2005, 94).

  11. 11.

    A homoplasy is “a similarity that is the result of two or more independent derivations of the trait” (ibid).

  12. 12.

    It is now clear that we cannot reliably infer common developmental mechanisms from the fact that a trait is taxically homologous (Wagner 2007). Furthermore, as Powell and Shea (forthcoming) argue, behavioral homology relations can be preserved notwithstanding a complete turnover in the underlying developmental mechanisms of a trait in one or both lineages, and despite even a shift in the inheritance system (genetic, cultural, etc.) through which the trait is transmitted.


  1. Buckner C (2011) Two approaches to the distinction between cognition and ‘mere association’. Int J Comp Psychol 24:314–348Google Scholar
  2. Buckner C (2013) Morgan’s Canon, meet Hume’s Dictum: avoiding anthropofabulation in cross-species comparisons. Biol Philos 28:853–871Google Scholar
  3. Carruthers P (2008) Meta-cognition in animals: a skeptical look. Mind Lang 23:58–89CrossRefGoogle Scholar
  4. Clayton NS, Dickinson A (1998) Episodic-like memory during cache recovery by scrub jays. Nature 395:272–278CrossRefGoogle Scholar
  5. Clayton NS, Griffiths DP, Emery NJ, Dickinson A (2001) Elements of episodic-like memory in animals. Philos Trans R Soc Biol 356:1483–1491. doi:10.1098/rstb.2001.0947 CrossRefGoogle Scholar
  6. Conway MS (2003) Life’s solution: inevitable humans in a lonely universe. Cambridge University Press, CambridgeCrossRefGoogle Scholar
  7. Cruse H, Wehner R (2011) No need for a cognitive map: decentralized memory for insect navigation. Public Libr Sci Comput Biol 7:e1002009. doi:10.1371/journal.pcbi.1002009 Google Scholar
  8. Crystal JD, Foote AL (2009) Metacognition in animals. Comp Cogn Behav 4:1–16Google Scholar
  9. Dally JM, Emery NJ, Clayton NS (2006) Food-Caching western scrub-jays keep track of who was watching when. Science 16:1662–1665. doi:10.1126/science.1126539 CrossRefGoogle Scholar
  10. de Waal FBM, Ferrari PF (2010) Towards a bottom-up perspective on animal and human cognition. Trends Cogn Sci 14:201–207Google Scholar
  11. Dickinson A (2012) Associative learning and animal cognition. Philos Trans R Soc Biol Sci 367:2733–2743. doi:10.1098/rstb.2012.0220 CrossRefGoogle Scholar
  12. Dugas-Ford J, Rowell JJ, Ragsdale CW (2012) Cell-type homologies and the origins of the neocortex. PNAS 109:16974–16979CrossRefGoogle Scholar
  13. Emery NJ, Clayton NS (2004) The mentality of crows: convergent evolution of intelligence in corvids and apes. Science 306:1903–1907CrossRefGoogle Scholar
  14. Fitzpatrick S (2008) Doing away with Morgan’s Canon. Mind Lang 23:224–246Google Scholar
  15. Foote AL, Crystal JD (2007) Metacognition in the rat. Curr Biol 17:551–555CrossRefGoogle Scholar
  16. Foote AL, Crystal JD (2012) Play it again: a new method for testing metacognition in animals. Anim Cogn 15:187–199CrossRefGoogle Scholar
  17. Gallistel CR (2008) Learning and representation. In: Menzel R, Byrne J (eds) Learning and memory: a comprehensive reference. Elsevier, New York, pp 227–242Google Scholar
  18. Güntürkün O (2012) The convergent evolution of neural substrates for cognition. Psychol Resour 76:212–219. doi:10.1007/s00426-011-0377-9 CrossRefGoogle Scholar
  19. Heyes C (2012) Simple minds: a qualified defense of associative learning. Philos Trans R Soc Biol Sci 367:2695–2703CrossRefGoogle Scholar
  20. Karin-D’Arcy MR (2005) The modern role of Morgan’s canon in comparative psychology. Int J Comp Psychol 18:179–201Google Scholar
  21. Le Pelley ME (2012) Metacognitive monkeys or associative animals simple reinforcement learning explains uncertainty in nonhuman animals. J Exp Psychol Learn Mem Cogn 38:686–708. doi:10.1037.a0026478Google Scholar
  22. Mameli M, Bortolotti L (2006) Animal rights, animal minds, and human mindreading. J Med Ethics 32:84–89CrossRefGoogle Scholar
  23. Morgan LC (1894) An introduction to comparative psychology. Scribner’s, New YorkCrossRefGoogle Scholar
  24. Pitt D (2012) Mental representation. In: Edward N. Zalta (ed) The Stanford Encyclopedia of philosophy (Winter 2012 Edition).
  25. Povinelli DJ, Barth J (2005) Reinterpreting behavior: a human specialization? Behav Brain Sci 28:712–713CrossRefGoogle Scholar
  26. Powell R, Shea N (forthcoming) Homology across inheritance systems. Biol PhilosGoogle Scholar
  27. Shettleworth SJ (2009) Cognition, evolution, and behavior: edition II. Oxford University Press, OxfordGoogle Scholar
  28. Shettleworth SJ (2012) Modularity, comparative cognition and human uniqueness. Philos Trans R Soc Biol Sci 367:2794–2802. doi:10.1098/rstb.2012.0211 CrossRefGoogle Scholar
  29. Shettleworth SJ (2013) Fundamentals of comparative cognition. Oxford University Press, New YorkGoogle Scholar
  30. Smith JD, Beran MJ, Couchman JJ, Coutinho MVC (2008) The comparative study of metacognition: sharper paradigms, safer inferences. Psychon Bull Rev 15:679–691Google Scholar
  31. Sober E (2005) Comparative psychology meets evolutionary biology: Morgan’s canon and cladistic parsimony. In: Daston L, Mitman G (eds) Thinking with animals: new perspectives on anthropomorphism. Columbia University Press, New York, pp 85–99Google Scholar
  32. Taylor AH, Knaebe B, Gray RD (2012) An end to insight? New Caledonian crows can spontaneously solve problems without planning their actions. Proc Royal Soc B Biol Sci 279(1749):4977–4981Google Scholar
  33. Thomas RK (2001) Lloyd Morgan’s cannon: a history of misrepresentation.
  34. van der Vaart E, Verbrugge R, Hemelrijk CK (2012) Corvid re-caching without ‘theory of mind’: a model. Public Libr Sci Online 7(3):e32904. doi:10.1371/journal.pone.0032904 Google Scholar
  35. Wagner GP (2007) The developmental genetics of homology. Nat Rev Genet 8:473–479CrossRefGoogle Scholar
  36. Wimpenny JH, Weir AAS, Clayton L, Rutz C, Kacelnik A (2009) Cognitive processes associated with sequential tool use in new Caledonian crows. Public Libr Sci Online 4:6471Google Scholar
  37. Zentall TR (2001) The case for a cognitive approach to animal learning and behavior. Behav Process 54:65–78CrossRefGoogle Scholar

Copyright information

© Springer Science+Business Media Dordrecht 2014

Authors and Affiliations

  1. 1.Department of PhilosophyBoston UniversityBostonUSA

Personalised recommendations