Is “the presence of general intelligence” the “major evolutionary puzzle” that Burkart et al. claim? Like much of the literature on general intelligence in animals, the target article draws inferences about the nature and evolution of cognitive traits from the correlations among measures of performance, both within and between species. The “positive manifold” (sect. 1.1, para. 1) is thus taken to be a nontrivial finding, and g is treated as being – or reflecting – a trait with causal effects (a mechanism). g, however, is of course a statistical construct: When the authors refer to “the structure of cognition” (sect. 1.1.1, para. 1), what they actually describe is the statistical structure of variance in performance on behavioural tests. What can this statistical structure tell us about cognitive traits? We suggest that it tells us very little, or possibly nothing, because of the multiple plausible ways in which it might arise. Moreover, the analysis of g fails to provide a clear framework for empirical research, because the putative underlying mechanism, general intelligence, cannot be meaningfully defined in the absence of the correlations that are used as evidence for its existence.
More specifically, the reification of g involves a conflation of the proposed domain-generality of cognitive processes with the statistical pattern of variance in the behavioural output of those processes. Thus, “Massive modularity would appear to be irreconcilable with general intelligence” (sect. 1.2.1, para. 4) – well, only in the sense that apples are irreconcilable with oranges. Burkart et al. follow many in assuming that the positive manifold can be explained “by positing a dominant latent variable, the g factor, associated with a single cognitive or biological process or capacity” (van der Maas et al. Reference van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006, p. 842). As pointed out by the latter authors, other explanations, which account for not only the presence of g but also its heritability and neuro-anatomical correlates, are not only possible, but also plausible. In citing van der Maas et al. (Reference van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006), Burkart et al. explicitly “equate general intelligence with the positive manifold” (sect. 1.1.1, para. 3), implying that their position and that of van der Maas et al. are in harmony. The point emphasised by van der Maas et al., however, and the point we also emphasise, is that the positive manifold provides little or no constraint on the possible architectures of cognition.
To labour the point, correlated variance does not imply any particular kind of cognitive process. That said, we might still want an explanation for why performance or behaviours are correlated across domains. Here, in brief, are some possibilities.
-
(1) They are not really different domains. For example, Reader et al. (Reference Reader, Hager and Laland2011) and Fernandes et al. (Reference Fernandes, Woodley and te Nijenhuis2014) found positive correlations among the rates of social deception, social learning, innovation, extractive foraging, dietary breadth, percentage of fruit in the diet, and tool use across primate species, leading both sets of authors to conclusions about the domain-generality of cognitive processes. Neither these authors nor Burkart et al. explain how a domain is to be identified, and therefore how these behavioural measures can, in principle, be used to test for domain-generality. We can envisage plausible arguments to the effect that at least some of these behaviours draw on the same domain-specific processes. It is a question of natural ontologies: How do we carve nature at her joints? The only way that makes sense to us is in an evolutionary context where we identify a domain with a selection pressure. Deciding that “social” and “non-social” are distinct domains is, therefore, a hypothesis about what selection pressures have operated, not necessarily a fact about the world. Burkart et al. recognise this problem (“The issue of task selection is thus closely linked to the identification of domains in animal cognition” [sect. 2.4.2, para. 5]) but do not offer a convincing solution.
-
(2) Related to (1), it may well be that the behaviours measured are influenced by cognitive processes they share in common, but this does not mean it is helpful to describe those processes as “general processes,” or that together they comprise “general intelligence.” For example, primate species vary in their sensory-motor adaptations – in particular, in their stereo visual acuity and manual manipulative abilities – and these differences correlate with the evolution of binocular convergence supporting stereo vision, the size of visuomotor structures in the brain, and consequently overall brain size (Barton Reference Barton2012; Heldstab et al. Reference Heldstab, Kosonen, Koski, Burkart, van Schaik and Isler2016). Clearly, such sensory-motor specializations may influence performance of a range of behaviours and/or experimental test procedures. Yet, describing them as “domain general” tells us nothing about how they work or how they evolved. We also do not share the optimism of Burkart et al. that reversal learning is free of such problems.
-
(3) Niche dimensions tend to be correlated (Clutton-Brock & Harvey Reference Clutton-Brock and Harvey1977). For example, folivorous primates generally live in smaller social groups, have smaller home ranges, and engage less in extractive foraging and tool use than more omnivorous primates. Cognitive adaptations for specific niche dimensions could therefore theoretically be completely informationally encapsulated and yet performance across domains would still be correlated.
-
(4) The rates of naturally occurring behaviours in the wild (Reader et al. Reference Reader, Hager and Laland2011; Fernandes et al. Reference Fernandes, Woodley and te Nijenhuis2014, cited by Burkart et al.), may be systematically biased, leading to spurious correlations. Although these studies attempt to control for observation effort, they don't control for the number of individuals under observation. Rates of all behaviours will, other things being equal, correlate positively with group size and therefore with each other, because more individuals are under observation per unit time in larger groups. Variation in observability due to habitat will only exacerbate the problem. The implications are obvious.
For a theory to be useful, it has to be well defined in such a way as to generate testable predictions that differentiate it from other theories. Burkart et al., along with the wider literature on general intelligence and g, fail to achieve this. If we are to make progress in our efforts to understand the evolution and structure of cognition, we need to stop confusing the map for the territory.
Is “the presence of general intelligence” the “major evolutionary puzzle” that Burkart et al. claim? Like much of the literature on general intelligence in animals, the target article draws inferences about the nature and evolution of cognitive traits from the correlations among measures of performance, both within and between species. The “positive manifold” (sect. 1.1, para. 1) is thus taken to be a nontrivial finding, and g is treated as being – or reflecting – a trait with causal effects (a mechanism). g, however, is of course a statistical construct: When the authors refer to “the structure of cognition” (sect. 1.1.1, para. 1), what they actually describe is the statistical structure of variance in performance on behavioural tests. What can this statistical structure tell us about cognitive traits? We suggest that it tells us very little, or possibly nothing, because of the multiple plausible ways in which it might arise. Moreover, the analysis of g fails to provide a clear framework for empirical research, because the putative underlying mechanism, general intelligence, cannot be meaningfully defined in the absence of the correlations that are used as evidence for its existence.
More specifically, the reification of g involves a conflation of the proposed domain-generality of cognitive processes with the statistical pattern of variance in the behavioural output of those processes. Thus, “Massive modularity would appear to be irreconcilable with general intelligence” (sect. 1.2.1, para. 4) – well, only in the sense that apples are irreconcilable with oranges. Burkart et al. follow many in assuming that the positive manifold can be explained “by positing a dominant latent variable, the g factor, associated with a single cognitive or biological process or capacity” (van der Maas et al. Reference van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006, p. 842). As pointed out by the latter authors, other explanations, which account for not only the presence of g but also its heritability and neuro-anatomical correlates, are not only possible, but also plausible. In citing van der Maas et al. (Reference van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006), Burkart et al. explicitly “equate general intelligence with the positive manifold” (sect. 1.1.1, para. 3), implying that their position and that of van der Maas et al. are in harmony. The point emphasised by van der Maas et al., however, and the point we also emphasise, is that the positive manifold provides little or no constraint on the possible architectures of cognition.
To labour the point, correlated variance does not imply any particular kind of cognitive process. That said, we might still want an explanation for why performance or behaviours are correlated across domains. Here, in brief, are some possibilities.
(1) They are not really different domains. For example, Reader et al. (Reference Reader, Hager and Laland2011) and Fernandes et al. (Reference Fernandes, Woodley and te Nijenhuis2014) found positive correlations among the rates of social deception, social learning, innovation, extractive foraging, dietary breadth, percentage of fruit in the diet, and tool use across primate species, leading both sets of authors to conclusions about the domain-generality of cognitive processes. Neither these authors nor Burkart et al. explain how a domain is to be identified, and therefore how these behavioural measures can, in principle, be used to test for domain-generality. We can envisage plausible arguments to the effect that at least some of these behaviours draw on the same domain-specific processes. It is a question of natural ontologies: How do we carve nature at her joints? The only way that makes sense to us is in an evolutionary context where we identify a domain with a selection pressure. Deciding that “social” and “non-social” are distinct domains is, therefore, a hypothesis about what selection pressures have operated, not necessarily a fact about the world. Burkart et al. recognise this problem (“The issue of task selection is thus closely linked to the identification of domains in animal cognition” [sect. 2.4.2, para. 5]) but do not offer a convincing solution.
(2) Related to (1), it may well be that the behaviours measured are influenced by cognitive processes they share in common, but this does not mean it is helpful to describe those processes as “general processes,” or that together they comprise “general intelligence.” For example, primate species vary in their sensory-motor adaptations – in particular, in their stereo visual acuity and manual manipulative abilities – and these differences correlate with the evolution of binocular convergence supporting stereo vision, the size of visuomotor structures in the brain, and consequently overall brain size (Barton Reference Barton2012; Heldstab et al. Reference Heldstab, Kosonen, Koski, Burkart, van Schaik and Isler2016). Clearly, such sensory-motor specializations may influence performance of a range of behaviours and/or experimental test procedures. Yet, describing them as “domain general” tells us nothing about how they work or how they evolved. We also do not share the optimism of Burkart et al. that reversal learning is free of such problems.
(3) Niche dimensions tend to be correlated (Clutton-Brock & Harvey Reference Clutton-Brock and Harvey1977). For example, folivorous primates generally live in smaller social groups, have smaller home ranges, and engage less in extractive foraging and tool use than more omnivorous primates. Cognitive adaptations for specific niche dimensions could therefore theoretically be completely informationally encapsulated and yet performance across domains would still be correlated.
(4) The rates of naturally occurring behaviours in the wild (Reader et al. Reference Reader, Hager and Laland2011; Fernandes et al. Reference Fernandes, Woodley and te Nijenhuis2014, cited by Burkart et al.), may be systematically biased, leading to spurious correlations. Although these studies attempt to control for observation effort, they don't control for the number of individuals under observation. Rates of all behaviours will, other things being equal, correlate positively with group size and therefore with each other, because more individuals are under observation per unit time in larger groups. Variation in observability due to habitat will only exacerbate the problem. The implications are obvious.
For a theory to be useful, it has to be well defined in such a way as to generate testable predictions that differentiate it from other theories. Burkart et al., along with the wider literature on general intelligence and g, fail to achieve this. If we are to make progress in our efforts to understand the evolution and structure of cognition, we need to stop confusing the map for the territory.