1.1.1. The structure of cognition

The structure of human cognition continues to be debated (e.g., Ortiz Reference Ortiz, Goldstein, Princiotta and Naglieri2015). Nonetheless, the presence of g is now widely accepted due to the pervasive evidence from Carroll's (Reference Carroll1993) seminal meta-analysis of over 460 carefully selected data sets on human cognitive ability. An influential account is Horn and Cattell's fluid-crystallized gf-gc model (see also Major et al. Reference Major, Johnson and Deary2012). Fluid intelligence gf refers to the capacity to think logically and solve problems in novel situations independently of previously acquired knowledge, and to identify patterns and relationships, whereas crystallized intelligence gc refers to the ability to use skills, knowledge, and experience and crucially relies on accessing information from long-term memory. An explicit causal link from gf to gc is provided by investment theory (Cattell Reference Cattell1987), which is the developmental version of the gf-gc model and finds considerable empirical support (Thorsen et al. Reference Thorsen, Gustafsson and Cliffordson2014).

An integrated version, the so-called CHC (Cattell-Horn-Carroll) theory, has been supported by several studies and is a widely accepted consensus model (McGrew Reference McGrew2009). The CHC model is hierarchical, placing a general factor g at the top, which affects both gf and gc. Most current models involve some hierarchical structure involving a general factor, g, and fluid intelligence, gf (but see, for instance, Bartholomew et al. Reference Bartholomew, Deary and Lawn2009; Major et al. Reference Major, Johnson and Deary2012; van der Maas et al. Reference van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006). In fact, some have argued that gf and g represent the same entity (Kan et al. Reference Kan, Kievit, Dolan and van der Maas2011), and the previously mentioned definition of intelligence in a broad sense in fact emphasizes elements of both constructs.

Some models of general intelligence that do not involve g are also still being considered. Van der Maas et al. (2006), for instance, have presented a dynamic model of general intelligence that assumes independent cognitive processes early in ontogeny. Over the course of development, the positive manifold emerges because of mutually beneficial interactions between these initially independent processes. To the extent that one agrees to equate general intelligence with the positive manifold, the mutualism model may be viewed as a model of general intelligence for human and nonhuman animals in which variation between species would reflect the extent to which mutually beneficial interactions between cognitive processes arise during development. Because, across species, bigger brains require more time to mature than smaller brains (Schuppli et al. Reference Schuppli, Isler and van Schaik2012), and thus have more opportunities to develop such mutually beneficial interactions, such a scenario is compatible with an evolutionary perspective.

1.1.2. Executive functions and intelligence

Closely related to general intelligence are executive functions, or EFs (Barbey et al. Reference Barbey, Colom, Solomon, Krueger, Forbes and Grafman2012; Blair Reference Blair2006). EFs refer to “general-purpose control mechanisms that modulate the operation of various cognitive subprocesses and thereby regulate the dynamics of human cognition” (Miyake et al. Reference Miyake, Friedman, Emerson, Witzki, Howerter and Wager2000, p. 50). In other words, they are “a family of top-down mental processes needed when you have to concentrate and pay attention, when going on automatic or relying on instinct or intuition would be ill-advised, insufficient, or impossible” (Diamond Reference Diamond2013, p. 136). Three core EFs can be distinguished, namely inhibitory control (behavioral inhibition, cognitive inhibition, and selective attention), working memory (Baddeley Reference Baddeley2010), and cognitive flexibility.

Various measures of EFs have shown strong correlations with g/gf. Whereas the average correlation between working memory and g is 0.72, in some studies using latent variable analysis, it even reached identity (Colom et al. Reference Colom, Abad, Rebollo and Shih2005; Nisbett et al. Reference Nisbett, Aronson, Blair, Dickens, Flynn, Halpern and Turkheimer2012), leading some authors to suggest that the two cannot be distinguished from each other (Royall & Palmer Reference Royall and Palmer2014). That g and EF are closely related is consistent with two further lines of evidence. First, working memory can be trained, and these training gains can translate into gains in general intelligence even though not all procedures are effective, and it is not always clear whether the training affects working memory per se or instead improves learning strategies (reviewed in Klingberg Reference Klingberg2010; Morrison & Chein Reference Morrison and Chein2011; Nisbett et al. Reference Nisbett, Aronson, Blair, Dickens, Flynn, Halpern and Turkheimer2012; Shipstead et al. Reference Shipstead, Redick and Engle2012). Second, growing up bilingually, which makes high demands on a variety of EFs on a routine basis, is associated with stronger EFs in non-linguistic contexts, and thus with g (Abutalebi & Clahsen Reference Abutalebi and Clahsen2015; Bialystok et al. Reference Bialystok, Craik and Luk2012; Rabipour & Raz Reference Rabipour and Raz2012). Nonetheless, because EFs do not provide the logical problem-solving functions and learning that are the hallmark of general intelligence (Embretson Reference Embretson1995), some aspects of general intelligence are independent of EFs.

In sum, evidence for domain-general intelligence in humans, estimated by the first factor derived in psychometric, factor-analytical approaches, is pervasive, and is backed up by neurobiological evidence and various correlates of life-outcome measures. The psychometrically derived g factor is thus consistent with the broad notion of general intelligence, which stresses reasoning ability and behavioral flexibility and invokes cognitive processes such as learning and remembering, planning, and executive functions. This conclusion raises the question of the evolutionary origin of general intelligence in humans, which we will address by reviewing recent developments in the nonhuman literature. To do so, we will review evidence for g in animals, and whether it is warranted to assume that g in animals is also consistent with a broader notion of general intelligence.

Intelligent behavior needs to be distinguished from behavior that may appear intelligent but lacks flexibility (Shettleworth Reference Shettleworth, Vonk and Shackelford2012a). Intelligent behavior in animals is often referred to as behavior that shows some degree of flexibility and emanates from some kind of mental representation rather than immediate perception only (Tomasello & Call Reference Tomasello and Call1997). For instance, when digging wasps are interrupted anywhere in the sequence of actions involved in measuring the size of a hole to place a larva together with a prey item into it, they must start again at the very beginning of the behavioral sequence (Wooldridge Reference Wooldridge1968). Thus, many behaviors that at first sight look like they are the product of reasoning or learning turn out to be inflexible adaptations or modules (Sherry Reference Sherry2006). A collection of such dedicated adaptations presumably represents the ancestral state (e.g., Shettleworth Reference Shettleworth, Vonk and Shackelford2012a; Reference Shettleworth2012b), and thus the null model against which the hypothesis of general intelligence has to be tested. Before turning to nonhuman animals, we will therefore provide an overview of domain-specific, modular conceptions of the mind that have been put forward particularly, but not exclusively, by evolutionary psychologists.

1.2.1. Modularity and general intelligence

A modular organization of mind is particularly appealing to evolutionary thinking because modular systems allow parts to be removed, added, or modified without affecting the function of the structure as a whole. Therefore, modular systems may be more evolvable or even the only evolvable systems (Clune et al. Reference Clune, Mouret and Lipson2013; Pavlicev & Wagner Reference Pavlicev and Wagner2012; Ploeger & Galis Reference Ploeger and Galis2011; Schlosser & Wagner Reference Schlosser and Wagner2004; Shettleworth Reference Shettleworth2012b). Thus, whenever conditions are sufficiently stable or at least predictable across generations, natural selection should favor solving recurrent fitness problems via modules rather than via general cognitive processes, because the former solve these problems on average quickly, effortlessly, and efficiently (Cosmides et al. Reference Cosmides, Barrett and Tooby2010) and can presumably evolve more readily. General intelligence, in contrast, is thus expected to evolve under conditions of social or environmental unpredictability. Solutions to these evolutionarily novel problems have to be acquired effortfully, via slow learning (e.g., Geary Reference Geary2005; Geary & Huffman Reference Geary and Huffman2002).

The advantages of a modular solution to recurrent fitness problems, however, are not necessarily as straightforward. First, the fundamental assumption that a modular solution is indeed more evolvable can be questioned on both empirical and conceptual grounds (e.g., Anderson & Finlay Reference Anderson and Finlay2014; Bolhuis et al. Reference Bolhuis, Brown, Richardson and Laland2011; d'Souza & Karmiloff-Smith Reference d'Souza and Karmiloff-Smith2011; Lefebvre Reference Lefebvre2014). Empirical evidence for a direct mapping of specialized adaptive behavioral functions to specific modular neural units is actually rare, even for neural systems as simple as those of invertebrates. Novel adaptive functions seem mostly to be achieved via massive re-use of neural tissue rather than via the addition of encapsulated neuronal pools. Conceptually, the evolvability argument seems largely incompatible with what is known about short-term neuromodulation, brain plasticity over the life span, response to damage, and ontogenetic principles of brain development. The a priori evolvability argument, therefore, does not lead to an unambiguous conclusion as to the superiority of domain-specific over domain-general organization.

Second, the other advantage of modularity – fast, effortless, and ultimately efficient solving of evolutionarily recurrent fitness problems – may hold only for particular notions of modularity, such as Fodorian modules (Fodor Reference Fodor1983). These are thought to be domain-specific functional units that process distinctive input stimuli using distinctive mechanisms. In particular, a module is thought to exclusively process information from a specific domain and to produce a correspondingly specific output in the form of representations and/or a behavioral response. Fodor listed criteria that must – at least to “some interesting extent” (Fodor Reference Fodor1983, p. 37) – be fulfilled by a functional unit to qualify as modular. These criteria include domain specificity, mandatory processing, high speed, production of shallow outputs (i.e., not requiring extensive processing), limited accessibility, a characteristic ontogeny (reliable emergence without explicit learning), a fixed neural architecture, and informational encapsulation (meaning it is not affected by other cognitive processes, a criterion thought to be particularly important). Paradigmatic examples of Fodorian modules are optical illusions. Accordingly, the presence of modules involving the processing of sensory information is widely accepted, and that their speed and efficiency are beneficial is obvious. However, a modular organization has also been proposed for more higher-level cognitive processes including ones related to folk psychology (e.g., processing of faces and facial expressions, theory of mind, cheater detection), folk biology (e.g., animate-inanimate distinction, flora-fauna), or folk physics (e.g., movement trajectories, gravity biases, representation of space, solidity, and causality; summarized in Geary Reference Geary2005). Indeed, massive modularity accounts hold that the mind is exclusively made up of modules (Barrett Reference Barrett, Zeigler-Hill, Welling and Schackelford2015; Carruthers Reference Carruthers and Stainton2005; Sperber Reference Sperber and Dupoux2001).

Massive modularity would appear to be irreconcilable with general intelligence (and therefore with the ability to solve evolutionarily novel problems), but much of the longstanding controversy about the massive modularity hypothesis of the human mind comes down to the use of different notions of modularity (see also Barrett & Kurzban Reference Barrett and Kurzban2006). Indeed, a variety of highly divergent notions have developed (Barrett Reference Barrett, Zeigler-Hill, Welling and Schackelford2015; Barrett & Kurzban Reference Barrett and Kurzban2012; Chiappe & Gardner Reference Chiappe and Gardner2012; Coltheart Reference Coltheart2011; Grossi Reference Grossi2014; Mahon & Cantlon Reference Mahon and Cantlon2011), and many of these are much broader than the Fodorian one (e.g., Sternberg Reference Sternberg2011). Because they also encompass the possibility of overarching, central control processes (Carruthers Reference Carruthers2011), they are entirely compatible with the coexistence of domain-general processes and general intelligence (Barrett Reference Barrett, Zeigler-Hill, Welling and Schackelford2015; Carruthers Reference Carruthers2011). In fact, Carruthers (Reference Carruthers2011) argued that most modules are specialized learning systems. Such broad notions of modularity, however, arguably no longer support the original idea of automatically providing fast and frugal solutions to recurrent fitness problems.

Unlike many proponents of massive modularity in humans, comparative behavioral biologists and comparative psychologists typically refer to notions of modularity that hew closely to the classical Fodorian modules, that is, dedicated, inflexible cognitive adaptations that have evolved in response to specific recurrent fitness-relevant problems (e.g., Fernandes et al. Reference Fernandes, Woodley and te Nijenhuis2014; Shettleworth Reference Shettleworth, Vonk and Shackelford2012a; Reference Shettleworth2012b). Functional specialization here is mostly used in the biological, ultimate sense – that is, referring to the specific adaptive pressures that gave rise to the evolution of specific dedicated modules. This perspective is grounded in research traditions such as neuroecology (Sherry Reference Sherry2006) that have provided empirical evidence for the occurrence among animals of dedicated cognitive adaptations, such as spatio-temporal memory abilities in food-caching species, birds in particular (Brodin Reference Brodin2010; Pravosudov & Roth Reference Pravosudov and Roth2013). These cognitive adaptations typically do not generalize to problems for which they did not evolve.

A mind composed of such dedicated adaptations represents a plausible null model, and indeed a plausible ancestral state of vertebrate cognition. Dedicated adaptations and general intelligence can obviously coexist (e.g., Cosmides et al. Reference Cosmides, Barrett and Tooby2010; Geary Reference Geary2005) – for instance, when the output of modules serve as inputs for intelligent reasoning, which may be responsible for the fact that in humans general intelligence predicts reasoning ability even in evolutionarily familiar contexts (Kaufman et al. Reference Kaufman, DeYoung, Reis and Gray2011). The key questions with respect to the evolution of general intelligence, therefore, are how central, domain-general processes could evolve on top of domain-specific adaptations, whether and to what extent they also exist in nonhuman animals, and what adaptive benefits drove their evolution.

1.2.2. Adaptive canalization beyond modularity

Strictly domain-general approaches that construe the mind as a general-purpose computer face several well-known problems (Cosmides & Tooby Reference Cosmides, Tooby, Hirschfeld and Gelman1994; Cosmides et al. Reference Cosmides, Barrett and Tooby2010; Frankenhuis & Ploeger Reference Frankenhuis and Ploeger2007; Heyes Reference Heyes2003; Kolodny et al. Reference Kolodny, Edelman and Lotem2015; see also Table 1). First, an agent has to efficiently identify relevant information and filter out irrelevant information in the process of problem solving, a challenge known as the frame problem. Second, once the relevant information has been identified, the agent has to decide what to do with it. To do so, she has to solve the problem of how to pick and combine correct, adaptive behavioral options or cognitive processes out of an exponentially growing number of possibilities (the problem of the combinatorial explosion) or to learn important associations and skills in a limited period of time despite dealing with relevant stimuli that occur at a low rate (the poverty of the stimulus problem). Third, correct responses have to be made quickly and efficiently (the urgency problem). And fourth, while doing so, the agent has to find general, rather than only locally successful, solutions (the functionality problem). It is thus beyond doubt that some canalization of cognitive processes is necessary.

Table 1. Overview of some specific problems that a domain-general cognitive apparatus has to overcome in order to produce ultimately adaptive behavior, as well as potential solutions – that is, adaptive canalization mechanisms. Note that these solutions may be very general themselves, such as a preference for social learning. See text for references.

¹ That is, evolved cognitive domains that are fleshed out with experience; for example, Gelman (Reference Gelman1990).

Evolved Fodorian modules (referred to as “cognitive adaptations” by behavioral biologists and neuroecologists) are clearly one way of solving the problems highlighted previously, in particular when they define the entire sequence from the acquisition of information to the adaptive behavioral response. However, they are not necessarily the only possible way, and natural selection may also overcome these problems in a different way that would allow domain-general abilities to evolve. A straightforward solution to this problem would be that domain-general abilities coevolve together with adaptive canalizing mechanisms that guide how general abilities are applied. Canalizing mechanisms can have a phylogenetic origin, such as a genetically predetermined preference for a certain category of stimuli: for example, the preference for faces in human infants (Shah et al. Reference Shah, Happé, Sowden, Cook and Bird2015). Alternatively, they can have an ontogenetic origin, such as the propensity of chimpanzees from tool-using communities to automatically perceive a stick as a potential tool, compared to genetically indistinguishable chimpanzees from non-tool-using communities who do not recognize this affordance (e.g., Gruber et al. Reference Gruber, Muller, Reynolds, Wrangham and Zuberbühler2011).

Table 1 summarizes the phylogenetic and ontogenetic canalizing mechanisms that ensure that domain-general cognition produces adaptive behavior despite the problems highlighted previously. Unlike Fodorian modules, these mechanisms do not define the entire sequence from signal detection to behavioral output, but may be deployed at different stages during information processing. We will now examine the evidence for such domain-general canalization processes.

The first problem an individual faces is what to attend to in the continuous stream of stimuli coming in from different sensory modalities. This can be solved by innate dispositions or data acquisition mechanisms (also referred to as phylogenetic inflection: Heyes Reference Heyes2003). Importantly here, innateness is not equivalent to inflexibility because innate dispositions to pay attention to one stimulus over another can be conditional. For instance, an animal foraging for berries may have an attentional bias to perceive small red entities, but the same animal when exposed to a raptor will be biased to perceive only potential hideouts. Alternatively, animals can learn ontogenetically which targets are particularly worth attending to (ontogenetic inflection). Here, social guidance of attention may play a particularly important role. Ontogenetic inflection automatically arises whenever immatures follow the mother and later other conspecifics, and is even more powerful in species that follow gaze (Shepherd Reference Shepherd2010). In many species, including humans, immatures are particularly attracted to everything conspecifics are interacting with, and immatures of some species, such as aye-ayes (Krakauer Reference Krakauer2005), marmoset monkeys (Voelkl et al. Reference Voelkl, Schrauf and Huber2006), or orangutans (Forss et al. Reference Forss, Schuppli, Haiden, Zweifel and van Schaik2015) are highly neophobic toward stimuli they have not witnessed their mother or other familiar conspecifics interact with. Natural selection can, therefore, favor the disposition to preferentially use social information to decide which stimuli to attend to, and thus leave the specific target of attention largely unspecified.

In a second step, the individual has to “decide” what to do with the stimuli that have captured its attention, because input mechanisms filter incoming stimuli but do not produce behavior. Subsequent processes are therefore required to determine what to do with these stimuli without being stymied by the problems of poverty of stimulus and the combinatorial explosion. First, in the case of phylogenetic inflection, coevolution of input mechanisms and response tendencies is frequent (Lotem & Halpern Reference Lotem and Halpern2012), as when a moving stimulus in the sky automatically triggers a flight reaction, but also when individuals are more likely to associate a snake (but not a flower) with fear (Cook & Mineka Reference Cook and Mineka1989), or a taste (but not an auditory stimulus) with subsequent nausea (known as biologically prepared learning or the Garcia effect: Garcia & Koelling Reference Garcia and Koelling1966). Second, in the case of ontogenetic inflection, social learning can also affect how the individual processes a stimulus that has come to its attention. Third, the stimuli that have attracted an individual's attention may be integrated with innate bodies of knowledge, so-called core knowledge (Gelman Reference Gelman1990; Spelke & Kinzler Reference Spelke and Kinzler2007) or psychological primitives (Samuels Reference Samuels2004), and so give rise to more elaborate skills and conceptual systems (Carey Reference Carey2009).

A third problem for the individual is that decisions often have to be made under time pressure (the urgency problem). Evolved modules, heuristics, or direct and reflexive triggering of responses are particularly good at providing fast responses because they bypass central processes. But quick and efficient responses can also be achieved in evolutionarily novel contexts, such as solving algebraic equations or playing chess, if a learned heuristic approach becomes an automated subroutine and can be applied effortlessly (Bilalić et al. Reference Bilalić, Langner, Ulrich and Grodd2011; Chang Reference Chang2014). Such problem solving has similar surface properties to modular organization sensu Fodor. This fact has sometimes led to conceptual misunderstandings (see also section 1.2.3), and is relevant for approaches that try to identify domain-general processes in nonhuman animals (see also section 2.4.3).

A final potential problem is that developmentally acquired response tendencies may be successful in solving local problems, but nevertheless may not ultimately help an individual survive and reproduce (the functionality problem). Individuals, be they animals or humans, typically do not represent ultimate fitness goals in their everyday behavior. Rather, they pursue a set of innate psychological goals, which on average results in fitness-enhancing behaviors (Tinbergen Reference Tinbergen1963) but may become maladaptive in environments other than the one in which the goals evolved, as shown by our strong preferences for sweet, fatty, and salty foods. However, innate goals may be modified or supplemented by socially acquired end-state preferences. For immatures, who are most strongly affected by the canalization problems listed in Table 1, copying successful adult individuals is widespread and generally results in adaptive behavior because they are copying individuals who have survived until adulthood and managed to reproduce. Socially acquired end-state preferences and goals are particularly widespread in humans, who are highly susceptible to conformity and prestige biases (Dean et al. Reference Dean, Vale, Laland, Flynn and Kendal2014; Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Narr, Newson, Ross, Smaldino, Waring and Zefferman2016). Increasing evidence also suggests the existence of such biases in at least some nonhuman primates and birds (Aplin et al. Reference Aplin, Farine, Morand-Ferron, Cockburn, Thornton and Sheldon2015; Kendal et al. Reference Kendal, Hopper, Whiten, Brosnan, Lambeth, Schapiro and Hoppitt2015; Luncz & Boesch Reference Luncz and Boesch2014; van de Waal et al. Reference van de Waal, Borgeaud and Whiten2013).

Despite being incomplete, Table 1 serves to highlight that adaptive canalization of cognition not involving Fodorian modules is possible, indeed potentially quite frequent. It also highlights the prominent reliance on social inputs to overcome the canalization problems inherent to domain-general mechanisms. Social learning is broadly defined in the animal literature – that is, learning influenced by observation of, or interaction with, another animal or its products (Heyes Reference Heyes1994; see also Box Reference Box1984). It is widespread in the animal kingdom, both in vertebrates and invertebrates, and ranges from processes as simple as social facilitation and enhancement learning to observational forms of social learning such as true imitation (e.g., Hoppitt & Laland Reference Hoppitt and Laland2013). Interestingly, it is increasingly assumed that many of the cognitive mechanisms involved in social learning are of a general nature rather than specialized, and are thus not specific to social learning (Behrens et al. Reference Behrens, Hunt, Woolrich and Rushworth2008; Heyes Reference Heyes2012; Reference Heyes2016). Indeed, all forms of social learning also include a major element of individual learning. This is most evident in forms such as stimulus enhancement, where the attention of a naïve individual is drawn to stimuli other individuals are interacting with, which then releases individual exploration, play, and trial-and-error learning with this stimulus. Individual learning and practice, however, are also involved in the acquisition of skills through imitation learning, whereby it is typical that, after observation, a phase of individual practice is required (Galef Reference Galef2015; Jaeggi et al. Reference Jaeggi, Dunkel, Van Noordwijk, Wich, Sura and van Schaik2010; Schuppli et al. Reference Schuppli, Meulman, Forss, van Noordwijk and van Schaik2016). Thus, natural selection for social learning seems to automatically trigger selection on individual learning and general cognitive ability, suggesting that ontogenetic canalization through social learning may have contributed to enabling the evolution of domain-general cognition, an issue to which we return in section 3.3.

1.2.3. Primary and secondary modularization, and implications for general intelligence in nonhuman animals

Evolved Fodorian modules have specific surface properties: they work fast, effortlessly, and automatically, and they do not require significant amounts of executive control and working memory. Nevertheless, identifying modules in animals based on these properties is problematic because skills, capabilities, and solutions to problems that are acquired through effortful problem solving and learning based on general cognitive processes may become automatized over time, a process we refer to as secondary modularization. After such secondary modularization, or automatization, these skills have many of the surface properties in common with primary, evolved Fodorian modules. Note that this distinction in primary and secondary modularization is analogous to the distinction in primary and secondary cognitive abilities by Geary (Reference Geary1995), but whereas the latter has been developed specifically for humans, the former is thought to apply to a broad array of animal species.

Despite the similarities in surface properties, primary and secondary modules differ fundamentally with regard to their origin (see Table 2): Primary modules are evolved adaptations with canalized, buffered development, whereas secondary modules represent ontogenetically acquired skills that were automatized during ontogeny. In fact, secondary modularization is particularly common during the immature period (d'Souza & Karmiloff-Smith Reference d'Souza and Karmiloff-Smith2011). A consequence of the different etiology of primary and secondary modules is that the latter are more variable in their content and distribution across individuals or populations of the same species. Because little is known about the ontogeny of many of the specialized cognitive modules postulated for humans (Geary Reference Geary2005), we should also acknowledge the possibility that some or all of these are secondary rather than primary (Anderson & Finlay Reference Anderson and Finlay2014) or at least subject to experiential influences. For instance, even some prototypical modules such as those involved in face perception depend on experience (Dahl et al. Reference Dahl, Chen and Rasch2014).

Table 2. Primary and secondary modules differ with regard to their etiology and development, which has implications for their content and distribution within a species or population

¹ Greenough et al. (Reference Greenough, Black and Wallace1987).

The implication for the question of general intelligence in nonhuman animals is that it is no longer possible to uniquely rely on surface properties such as speed, effort, efficiency, and reliability to infer the presence of evolved domain-specific modules, because secondary modules have similar properties. Instead, a better diagnostic tool for the presence of general cognitive abilities is the presence of variable skill profiles across individuals and genetically similar populations due to secondary modularization (see section 2.4.3).

We have shown that human cognition involves elements of domain-specific and domain-general processes, but that the same can potentially be true for animals as well. Hence, animal minds need not be bundles of specialized cognitive adaptations. Having thus leveled the playing field, we first, in section 2, review recent evidence for whether a positive manifold (g) is present in nonhuman animals at all, and if so, how such a g factor is best explained. In particular, we will focus on the question whether such psychometric intelligence shows any of the features usually referred to as general intelligence. Even if we can be confident that this is the case in humans, whether the same applies to animals must be an empirical question (Galsworthy et al. Reference Galsworthy, Arden, Chabris, Finkel and Reynolds2014), and we highlight different research strategies that may prove to be fruitful in the future. In section 3, we then use this pattern of results to examine the ultimate evolutionary question of why general intelligence evolved, and which selection pressures may have favored it.

2.4.1. Statistical issues

The use of PCAs or related procedures involves a suite of decisions, including whether exploratory or confirmatory analyses are applied, whether non-rotated or rotated factors are considered, and whether oblique or orthogonal rotations are used. A detailed discussion of factor-analytical procedures is far beyond the scope of this review, and we refer readers to the specialized literature (e.g., Barney et al. Reference Barney, Amici, Aureli, Call and Johnson2015; Garson Reference Garson2013; Stevens Reference Stevens2012). However, because these decisions may critically affect the conclusions of animal studies, we must highlight some issues that appear relevant to the empirical results summarized previously.

First, the use of confirmatory analyses requires an a priori decision of what a domain is, and which tasks are associated with the respective domains (this also applies to Bayesian approaches that likewise categorize tasks a priori to hypothesized domains: Amici et al. Reference Amici, Barney, Johnson, Call and Aureli2012; Barney et al. Reference Barney, Amici, Aureli, Call and Johnson2015). The identification of domains of animal cognition, however, is not straightforward. For instance, some classify spatial reversal learning tasks as spatial cognition (e.g., Locurto & Scanlon Reference Locurto and Scanlon1998) whereas others stress their inhibition component (Tapp et al. Reference Tapp, Siwak, Estrada, Head, Muggenburg, Cotman and Milgram2003). In reality, of course, subjects may recruit several specific abilities to solve a particular task, and in fact different subjects may even recruit a different mix. Accordingly, Hopkins et al. (Reference Hopkins, Russel and Schaeffer2014) found that their exploratory PCA findings were not entirely consistent with the a priori structure of the PCTB originally proposed by Herrmann et al. (Reference Herrmann, Call, Hernández-Lloreda, Hare and Tomasello2007; Reference Herrmann, Hernandez-Lloreda, Call, Hare and Tomasello2010b). An a priori allocation of tasks to domains is thus not straightforward; in fact, the structure of a species' cognition is an empirical question (see also sect. 1.1.1 for corresponding efforts in human intelligence research). Accordingly, the use of confirmatory techniques may lead to diverging results compared to analytical approaches that are a priori agnostic with regard to factor structure.

Second, studies vary with regard to whether they present rotated or non-rotated solutions. Because rotations are designed to make the pattern of factor loadings more pronounced, it is generally recommended to use non-rotated solutions in g studies (Galsworthy et al. Reference Galsworthy, Arden, Chabris, Finkel and Reynolds2014; Jensen & Weng Reference Jensen and Weng1994; Locurto et al. Reference Locurto, Fortin and Sullivan2003; Plomin Reference Plomin2001; Woodley of Menie et al. Reference Woodley of Menie, Fernandes and Hopkins2015). Rotated and non-rotated solutions from the same data set are presented in Hopkins et al. (Reference Hopkins, Russel and Schaeffer2014) and Woodley of Menie et al. (Reference Woodley of Menie, Fernandes and Hopkins2015). Whereas the varimax-rotated solution (Hopkins et al. Reference Hopkins, Russel and Schaeffer2014, Table 1) appeared to suggest that a general factor g was lacking, the results of non-rotated solutions, verified by parallel analysis, demonstrated it was in fact present.

Third, a common intuition in general intelligence studies on animals is to compare the amount of variance explained by a first factor, and to conclude that the higher the amount of explained variance, the stronger the evidence for g. In human studies, the first non-rotated factor typically accounts for about 40% of variance (Plomin Reference Plomin2001), which is in fact similar to what has been reported for mice (see Table 3). However, an exclusive focus on the amount of explained variance is problematic for empirical and conceptual reasons. Empirically, the proportion of explained variance not only depends on the statistical issues discussed previously, but also on the heterogeneity of the subjects in the sample: the more heterogeneous, the higher the proportion of variance explained. In interspecific investigations, for instance, this means that studies that involve species that vary widely in general intelligence and brain size (e.g., 20 species of primates ranging from great apes to prosimians) will find higher proportions of explained variance than studies with a similar sample size, but where the species are all relatively similar (e.g., 20 different species from the same genus or taxonomic family). Conceptually, to the extent that the mind is a combination of both specialized cognitive adaptations and domain-general processes (see also sect. 4.1), very small proportions of explained variance may still be indicative of a real g. Likewise, a first factor with high loadings of some tasks but not others may reflect the absence of general intelligence, but may also reflect the co-occurrence of a general factor and one or several additional, more specialized domains (e.g., for spatial orientation, see Herrmann et al. Reference Herrmann, Call, Hernández-Lloreda, Hare and Tomasello2007; see also first PCA factor in Hopkins et al. Reference Hopkins, Russel and Schaeffer2014).

Last but not least, the most severe statistical restriction of nonhuman psychometric studies is that they critically lack power due to their small sample sizes. Reaching a near-consensus about the structure of human intelligence required meta-analyses involving thousands of subjects (Carroll Reference Carroll1993). Obtaining sample sizes comparable to human studies is unrealistic for most nonhuman animal species, in particular for nonhuman primates (albeit less so for rodents). However, replicating studies is feasible, and if this reveals the same factorial solution in a different set of subjects, and if combining such data sets also increases the fit of the solution, we can be increasingly confident that we are not dealing with statistical artifacts. Unfortunately, although this approach minimizes Type I errors, it suffers from very limited power to avoid Type II errors. In other words, if successful, we can be confident that we have obtained a real result, but if it fails, this may reflect either the absence of a general factor or too low a number of subjects. This shortcoming highlights the need to use external validation for psychometric g/G studies, as discussed below in section 2.5.

2.4.2. Methodological issues

We now turn to the possibility that a g/G factor may arise as a methodological artifact, because the results reflect variation in underlying variables other than general intelligence (see also Macphail's [1982] contextual variables) or because the tasks mainly tap into problems of the same domain.

Some individuals, or some species, may systematically outperform others not because they are more intelligent, but because they are less fearful and better habituated to testing, are more motivated to participate in tasks, have sharper senses, or are simply more active than others (Macphail Reference Macphail1982). Ideally, such confounds are directly quantified, as for instance in Matzel et al. (Reference Matzel, Townsend, Grossman, Han, Hale, Zappulla, Light and Kolata2006). In a sample of 43 mice individuals, they examined to what extent the general learning ability g extracted via PCA from a test battery of six cognitive tasks was correlated with 21 measures of exploratory behavior, sensory/motor function (e.g., running and swimming speed, balance tasks), activity, or fear/stress sensitivity. They found that g was not explained by general activity, sensory/motor function, physical characteristics, or direct measures of fear, but was correlated with several exploratory behaviors. Follow-up studies suggested that this link is caused by variation in habituation rates when exposed to potentially stressful situations (Light et al. Reference Light, Grossman, Kolata, Wass and Matzel2011) rather than by fearfulness influencing both exploration and task performance: Treatment with anxiolytic drugs did increase exploratory behaviors but did not improve performance in individual tasks or g (Grossman et al. Reference Grossman, Hale, Light, Kolata, Townsend, Goldfarb, Kusnecov and Matzel2007). Likewise, temporary environmental enrichment resulting in increased exploration tendency did not improve performance on the cognitive test battery (Light et al. Reference Light, Kolata, Hale, Grossman and Matzel2008). Thus, exploration and g may covary because more exploratory individuals are more likely to encounter contingencies in the environment that promote learning and problem solving, which over time leads to greater experience. The correlation between exploration and g may thus reflect a long-term, cumulative effect of experience on g. This is in line with investment theory (Cattell Reference Cattell1987), and with findings in human infants, where the preference for novelty and habituation is positively correlated with later performance in IQ tests (Teubert et al. Reference Teubert, Vierhaus and Lohaus2011), but also with apes, where individuals more likely to approach novel objects and a human stranger performed better in physical-cognition tasks (Herrmann et al. Reference Herrmann, Call, Hernández-Lloreda, Hare and Tomasello2007). Thus, the rodent studies support the idea that g is not an artifact of confounding factors.

Another non-cognitive factor that may explain variation in cognitive performance is motivation to participate. Female callitrichid monkeys have been reported to outperform males in problem-solving tasks (Brown et al. Reference Brown, Kaplan, Rogers and Vallortigara2010; Yamamoto et al. Reference Yamamoto, Domeniconi and Box2004). However, female callitrichids are typically also more food-motivated, whereas males are more vigilant than females (Koenig Reference Koenig1998). Accordingly, males are less interested in participating in experimental tasks and more easily emotionally aroused during testing. But if male performance is controlled for the presence of attention to the test stimuli, their performance is no longer inferior to that of females (Schubiger et al. Reference Schubiger, Wüstholz, Wunder and Burkart2015). The sexes thus do not differ in cognitive ability, but in their motivation to participate in experimental tasks.

The problem that we may never be sure if species differences in cognitive performance are the result of differences in cognitive ability or differences in contextual variables (Macphail Reference Macphail1982) remains an ongoing challenge for any species comparison. Nevertheless, not all tasks are affected by this problem to the same extent. Reversal learning tasks, for instance, are arguably less affected, because individuals first have to reach a criterion of making an initial discrimination. Differences in sensory-motor abilities and so on may well influence how difficult it is for a species to learn a particular discrimination. However, the crucial test is applied only once a specific criterion has been reached, and at least in marmosets, the time needed to achieve this criterion does not predict performance in the reversal trials (Strasser & Burkart Reference Strasser and Burkart2012). Furthermore, it is reassuring that the strongest association between a specific task and G in Deaner et al. (Reference Deaner, van Schaik and Johnson2006) was the one between reversal learning and G.

A second fundamental methodological issue refers to the task selection and battery development. With respect to task design, it is increasingly recognized that small differences in methodological details can strongly influence task performance, which has to be taken into account when performing species comparisons. For instance, memory performance strongly depends on task format in both marmosets and squirrel monkeys. Tests of memory often rely on a two-option choice task (e.g., Banerjee et al. Reference Banerjee, Chabris, Johnson, Lee, Tsao and Hauser2009), but many individuals are then happy to follow a random choice, which yields a 50% reward rate. When the choice involves many more options, subjects will be more motivated to remember the location of the food items and provide more-accurate estimates of their ability to memorize the location of the food item (Schubiger et al. Reference Schubiger, Kissling and Burkart2016). Regarding battery design, if all tasks in the test battery are drawn from the same domain (i.e., a lower-order group factor), rather than from a variety of domains, the positive correlations will reflect a domain-specific ability rather than a more general underlying cognitive factor (g/G). For instance, a positive manifold across a number of maze tasks is consistent with a spatial factor, but not informative with regard to g. The issue of task selection is thus closely linked to the identification of domains in animal cognition, which in fact is part of the empirical question that needs to be addressed in intelligence research in animals in general, by using batteries as diverse as possible and statistical procedures that are a priori agnostic to the underlying factor structure.

2.4.3. False negatives as a result of secondary modularization

Task selection may also bias the result and potentially produce false negatives if tasks prone to secondary modularization are included. Secondary modularization refers to the process that during ontogeny, individuals may specialize on a specific set of problems in a particular domain (Table 2). Problem solving in this domain becomes automatized and thus acquires many features commonly associated with modules rather than domain-general reasoning, particularly fast and frugal information processing, which is independent of reasoning. Thus, despite the presence of g in a given species, performance among individuals across domains need not be correlated whenever heterogeneous developmental inputs prevail that lead different individuals to specialize in different tasks (see Fig. 1c). This applies in particular to the small samples typical for nonhuman primate studies.

Figure 1. Performance across different cognitive domains (D₁-D₄). Each line represents the performance of an individual (or, in highly cultural species, of a population). (a) Performance is driven by domain-specific abilities; all individuals perform well in some domains but worse in others, but individual differences across domains are random; (b) performance is driven by domain-general abilities and individuals experience homogeneous developmental conditions, which leads to correlated performance between individuals across domains; (c) performance is driven by domain-general abilities but heterogeneous developmental conditions lead to specialization and secondary modularization of individuals in different domains. As a result, performance between individuals across domains is not correlated despite the presence of g.

Prima facie, this situation (Fig. 1c) may seem incompatible with the positive manifold, which is well documented in humans and perhaps other animals. It is important to keep in mind, however, that psychometric studies in humans are typically performed on subject pools with a rather uniform cultural background (the same is also true for the rodent studies performed on lab animals with virtually identical rearing conditions). If, in human studies, the cultural backgrounds of subjects were more diverse (e.g., ranging from Western-industrialized to a variety of hunter-gatherer societies), and only a small number of subjects tested, such an outcome (as in Fig. 1c) is quite likely (see also Reyes-García et al. Reference Reyes-García, Pyhälä, Díaz-Reviriego, Duda, Fernández-Llamazares, Gallois, Guèze and Napitupulu2016). The notorious difficulty of devising culture-free or at least culture-fair intelligence tests is a direct consequence of this problem (Saklofske et al. Reference Saklofske, van de Vijver, Oakland, Mpofu, Suzuki, Goldstein, Princiotta and Naglieri2014).

The prime example for secondary modularization in nonhuman primates is tool use, which is part of many test batteries typically used with nonhuman primates (e.g., Herrmann et al. Reference Herrmann, Call, Hernández-Lloreda, Hare and Tomasello2007; Reader et al. Reference Reader, Hager and Laland2011). Nonhuman primates vary considerably with regard to tool use, with great apes typically outperforming monkeys. But differences also occur within a species, both between wild and captive animals and among wild populations. Individuals of the same species show much higher propensities to use tools in captivity compared to their counterparts in the wild (Meulman et al. Reference Meulman, Sanz, Visalberghi and van Schaik2012; van Schaik et al. Reference van Schaik, Deaner and Merrill1999). Once proficient, individuals show tool use with high degrees of automatization and efficiency. Wild populations, too, vary significantly with regard to their propensity to use tools and solve tool-related problems (e.g., chimpanzees, Gruber et al. Reference Gruber, Muller, Reynolds, Wrangham and Zuberbühler2011; orangutans, van Schaik et al. Reference van Schaik, Ancrenaz, Borgen, Galdikas, Knott, Singleton, Suzuki, Suci and Merrill2003; or capuchin monkeys, Cardoso Reference Cardoso2013), arguably because they have ontogenetically acquired systematically different affordances of sticks or stones, which are perceived as potential tools in habitual tool users but not in non-tool users.

False negatives resulting from secondary modularization (see Fig. 1c) can be minimized if subjects with comparable rearing conditions are selected for the tests, but also if tasks prone to secondary modularization are excluded from test batteries. Thus, instead of naturalistic tasks that test for ontogenetically constructed skills that are likely to become automatized, such as tool use, or the ability to point and understand pointing, or even to use human language systems (Savage-Rumbaugh et al. Reference Savage-Rumbaugh, Fields, Segerdahl and Rumbaugh2005), it is preferable to include tasks testing for more elementary cognitive abilities, such as reversal learning, mental rotation, or quantity discrimination. Likewise, tests should avoid reliance on experience and knowledge of affordances that may differ among individuals depending on their biographies.

Although it is important to identify tasks and abilities prone to secondary modularization, it is not always easy to identify them. One way to do so is to examine the ontogeny of skills that are suspected to be the result of secondary modularization. Such skills should be acquired by developing immatures after a period of learning (perhaps following alternating series of instances of social learning and practice: Meulman et al. Reference Meulman, Seed and Mann2013; Schuppli et al. Reference Schuppli, Meulman, Forss, van Noordwijk and van Schaik2016), and could also potentially show high variation among adults. The increasing evidence for a major amount of skill learning by immature primates (e.g., Schuppli et al. Reference Schuppli, Meulman, Forss, van Noordwijk and van Schaik2016) and mammals and birds more generally (van Schaik et al. Reference van Schaik, Burkart, Damerius, Forss, Koops, van Noordwijk and Schuppli2016) suggests a greater prevalence of secondary modularization in nonhumans than revealed by the size of cultural repertoires (Whiten & van Schaik Reference Whiten and van Schaik2007). Because in wild populations, social and ecological problems tend to be very uniform for all individuals, variation of skill profiles (see Fig. 1) between populations (that live under similar, wild or captive, conditions), rather than among individuals of the same population, provides an additional heuristic tool to distinguish between genuine primary and secondary modularity. This criterion would work for primate tool use, for instance. Most powerful to disentangle primary from secondary modularity, finally, are cross-fostering experiments. When cross-fostered individuals exhibit species-typical behavior from the foster species rather than their own species, these behaviors clearly cannot result from primary modules. If the same procedure works within a species at the level of populations, it is similarly evidence for secondary, and thus learned, modules.