Herb Simon (Reference Simon1990) proposed more than twenty years ago that the problem of identifying invariants of human behavior is complicated by the fact that people are adaptive systems with highly flexible behavior. Invariants must therefore be sought not in the behaviors themselves, but in the mechanisms that guide the processes of individual adaptation: the mechanisms of intelligence. This stance supported the search for mechanistic models of cognition, in the form of computational models (Baars & Franklin Reference Baars and Franklin2009; French Reference French1995; Holland et al. Reference Holland, Holyoak, Nisbett and Thagard1986; Newell Reference Newell1990; Rumelhart et al. Reference Rumelhart and McClelland1986; Simon Reference Simon1979).

Researchers have become increasingly aware that emotions cannot be considered separable from cognition (Damasio Reference Damasio2000; Dolan Reference Dolan2002; LeDoux Reference LeDoux1996; Phelps Reference Phelps2006). Emotional states are the lens through which perceptions become thoughts, and thoughts become actions. A more thorough understanding of human behavior requires the incorporation of emotion into our mechanistic models. We adapt Simon's proposition thus: The problem with identifying invariant emotional categories is complicated by the fact that people are adaptive systems, whose emotions are highly flexible. Invariants must therefore be sought in the mechanisms that allow them to produce these feelings – the mechanisms of affective responses.

Lindquist et al. show that locationism, and a strict adherence to a natural kinds view of emotions, is likely a theoretical dead end. More importantly, they seek to replace a modular view of emotion states with a mechanistic model of how those states come to be generated. The model they have proposed is vague, perhaps necessarily so, as emotional categories are emergent phenomena of a highly complex system. There may be flaws in their model, and it may eventually be discarded in favor of a better model. It is clear, however, that the incorporation of emotions into a mechanistic understanding of cognition requires a generative model of emotional states. The invariants of emotions are to be found not in the discrete emotional categories, but in mechanisms that generate emotional responses.

A functionalist perspective of emotions as natural kinds assumes, at minimum, that each emotion (or “family of emotions”; Ekman Reference Ekman1992) is controlled by an independent cognitive system, even if those systems are not locationally isolated. This is an organizational blueprint which is attractive in its clarity and simplicity, but which strikes us as unlikely. The brain's cognitive architecture is a system of vast complexity shaped by millions of years of evolution. It is unlikely that the emergence of each emotion involved a separate evolutionary trajectory, each with its own selection pressures and phylogenetic path. Instead, we consider far more plausible the evolution of a generative, multipurpose emotion system. This is supported by the principle that it is much easier to evolve a self-organizing complex system that can adapt to varied environments than to specify a unique module for every contingency (Kauffman Reference Kauffman1995; Schank Reference Schank2002).

Although emotions may not represent natural kinds in terms of brain organization, we caution against the complete dismissal of the idea of basic affective categories; emotions may exist as quasi-natural perceptual kinds without necessitating functional locationism. There is cross-cultural recognition of the meaning of emotional expressions (Ekman Reference Ekman, Wagner and Manstead1989), even if these expressions must be exaggerated to be identified (Barrett Reference Barrett2006a). This may reflect more about the perception of emotions than about their expression. Humans are natural pattern classifiers, and the ability to readily detect broad emotion categories from facial expressions makes sense from an adaptationist perspective (Daly & Wilson Reference Daly and Wilson1995).

Social animals are well served by the quick assessment of meaning behind others' emotional expressions, even if fast mapping means a loss of nuanced information. The sight of a confederate running towards you with a look of panicked horror yields an altogether different reaction on your part than a look of joyful excitement. We think it unlikely that emotional categories are largely driven by language, as proposed by Lindquist et al., as well as by Barrett and colleagues previously (Barrett et al. Reference Barrett, Lindquist and Gendron2007b). Rapid identification of emotional cues is clearly adaptive, possibly even more for individuals without language, and moreover is present in nonhuman primates (Tate et al. Reference Tate, Fischer, Leigh and Kendrick2006). If recurring families of cues become strongly associated with certain types of responses, evolution may select for enhanced recognition of those cues on the part of the observer, and, if successful transmission is adaptive, more precise production on the part of the transmitter (Bateson Reference Bateson2004; Maynard Smith & Harper Reference Maynard Smith and Harper2003). Therefore, rapid identification of important emotional cues may embody quasi-natural kinds as emergent properties of fast-acting response systems, and it seems to us that this hypothesis has not been damaged by Lindquist et al.'s analysis.

Finally, we hope that this target article represents the first step toward rigorously tested generative models of emotion. Although we can say little about this here, understanding the nature of emotions is critically important for producing explicit models of human and animal social behavior, which is in turn crucial for gaining insight into and predictions about our world.