Beyond belongingness: Rethinking innate behavioral predispositions,learning constraints,and cognitive capacities

Abstract

Learning is a major determinant of behavioral change for some organisms through their lifecycles. From an associative perspective, learning is assumed to occur whenever organisms experience particular statistical regularities in their environment; specifically, meaningful outcomes that follow certain cues or actions chiefly contribute to behavioral change. However, numerous empirical reports reveal that not all cue–outcome and action–outcome combinations are learned equally well, a phenomenon that is termed belongingness. Those reports are valuable as descriptive-level knowledge, but beg further considerations, like what is the origin, adaptive value of, and underlying mechanisms associated with the predisposition to couple particular events. Contrary to what is often assumed, the mere observation of learning predispositions says little as to whether they arise from genetics, are constrained by hardwired neural circuitries, or have been ecologically advantageous in an evolutionary timescale. The present paper aims to present a number of notions from different research fields outside the hard core of associative learning and, in so doing, provides elements for careful study and conceptualization of this issue. Thereafter, these notions are pooled to understand behavioral variation in a wide array of phenomena, thus, bringing a more informed approach to the nature versus nurture debate.

Behavior is an aspect of an individual’s phenotype; as such, its expression is a unique product of the interaction between genes and the environment. Whenever variation in behavior is observed, further inspection is required to fully understand the mechanisms by which it came to be or about what is making it happen. Some species have nervous systems that enable complex forms of behavior and, in many cases, reorganization of networks within these systems allows for dynamic adjustment of behavior. Learning can be defined as a potential change in behavior due to experience (Mackintosh, 2015). In animals, neural circuitry reorganization mediates learning, meaning that this process is constrained by the particulars of individual and/or species-wide neural anatomy and physiology. Select laboratory animal models are routinely used to investigate learning mechanisms given the assumption of cross-species continuity, as well as based on convenience. Some authors have tried to explain learning phenomena observed in the laboratory by appealing to a more or less finite set of simple associative principles. The main tenet of this associative learning framework is that a change in behavior takes place whenever the relationship between cues or actions and meaningful outcomes changes. There are, nevertheless, certain findings that complicate this generally applicable account.

In his seminal studies, Thorndike (1898) observed that cats more readily learned to perform some actions than others to escape from a puzzle box. He introduced the term “belongingness” to account for this learning bias within the context of instrumental conditioning (i.e., learning action-outcome relationships). What Thorndike meant by this term is that some pairs of events are more readily associated than others, as if they had already belonged with one another beforehand. Therefore, the term “selective association” was later coined to refer to the phenomenon more broadly. This reasoning implies that some pairs of events are less likely to be associated; instead, they are constrained in their potential to be paired for giving rise to new forms of behavior. This continuum from belongingness to constraint is remarkably useful as a taxonomic scheme to account for variations in learning rates and boundaries.

After Thorndike’s studies, belongingness and constraints on learning did not receive much attention until the publication of a report from Breland and Breland (1961) and an experiment from Garcia and Koelling (1966). Breland and Breland revealed that explicitly rewarding animals’ actions may initially strengthen a target behavior but could progressively degenerate it into maladaptive and unyielding behavioral stereotypes. The term “misbehavior” was introduced to denominate learned behavioral routines that are overly rigid and presumably resistant to adaptive change. Likewise, García and Koelling found that certain combinations of conditioned stimuli (CS) and unconditioned stimuli (US) more readily support Pavlovian conditioning (learning stimulus-stimulus relationships); specifically, gustatory–visceroceptive and audiovisual–haptic stimulus combinations were easier to learn than crossover combinations (i.e., gustatory–haptic and audiovisual–visceroceptive). These findings fostered a paradigm shift from general-process associative learning theories to theories that acknowledge the possibility of putatively innate constraints and predispositions when it comes to learning. However, some argue that in an eagerness to reject the old doctrine these findings were sometimes uncritically interpreted (e.g., Domjan, 2015, 2016).

One concern is that to unequivocally certify that a certain behavioral pattern is a product of belongingness (or lack thereof) there are some methodological requirements that are often not met (see Bevins, 1992). In a similar vein, a concern arises regarding the implications of belongingness or lack thereof. Specifically, there is a discussion on whether belongingness and related phenomena actually represent violations of general-process associative learning theories (Overmier & Meyers-Manor, 2015) or rather just a challenge to the refinement of those theories (e.g., incorporating new factors and parameters). Along the same lines, belongingness/learning constraints have been seen as not necessarily implying a biological predisposition. An observed behavioral outcome can indeed be biologically predisposed and hardwired in brain circuitries (either established prenatally or as a product of a fixed postnatal developmental program), but it can also be a product of prior experiences (Damianopoulus, 1989). An uncritical leaning toward the former assumption may have been in part fueled the belief that cognitive capacities are modular and implemented by discrete brain structures, rather than emergent behavioral patterns in response to environmental challenges (see Cisek, 2021; Pessoa et al., 2022). Cautionary notes couched as “friendly reminders” about the interpretation of belongingness and related phenomena have been periodically published for decades (Burgos, 2015); however, it seems that the notions in these published critiques either did not fully transcended the field of associative learning or failed to appeal to conceptual resources outside of this field.

In this paper, I make the case that evidence for behavioral and learning predispositions should be carefully studied and conceptualized. To date, there is compelling evidence of innate behavior biases (e.g., Catania, 2010), learning constraints (e.g., Amichai & Yovel, 2021), and among-species dissimilarities in cognitive capacities (e.g., Kret & Tomonaga, 2016). However, it is also common for studies inquiring into these issues to uncritically extrapolate, overhype, or poorly justify conclusions drawn from data showing behavioral variation. While the term belongingness has fallen into disuse, unwarranted conclusions regarding biological constraints on learning and behavior that fail to consider historical and ecological factors are still prevalent. Herein, I address this very issue. Sometimes there is a tendency to invoke genetic, evolutionary, neurocomputational, or stochastic noise explanations without considering reasonable alternative explanations. I argue that some tensions between general-process learning and nativist/modularist approaches could be tempered by accounting for certain emerging empirical findings and theoretical notions outside of the field of associative learning. These include nonlinear interactions between genes and environmental factors, the view of behavior as an implementation of closed negative feedback loops, accounting for the influence of non-obvious experiences, the mediating role of biomechanics, and the phylogenetic refinement of behavioral taxonomies. In the last part of the paper, these notions are pooled to elucidate select cases across a wide range of phenomena with varying degrees of proximity to the hard core of associative learning literature. This, I argue, aids in synthesizing apparently unrelated phenomena in associative learning, behavioral neuroscience, and comparative cognition.

1. Genotype–environmental interactions revisited with a focus on learning and behavior

The nature versus nurture controversy has been around for several decades in fields such as genetics, computational neuroscience, ecology, comparative cognition, anthropology, developmental psychology, personality psychology, linguistics, psychiatry, and intelligence studies. In some of these fields, this matter is still hotly debated (e.g., Graves, 2015; Zador, 2019); advocates on both sides of the divide charge their opponents with reductionist biases (e.g., Buss, 2020; DiFrisco & Jaeger, 2020). Learning theory has its own version of the nature versus nurture debate in the issue of belongingness, because of the prevalent assumption that evidence of belongingness implies support for biological constraints on behavior. This is illustrated by the fact that findings on belongingness and related phenomena are often invoked to support the nativist side of the debate, even outside the field of associative learning (e.g., Pinker, 2002). Therefore, conceptual advances and better methodological tools for studying belongingness and related phenomena through the lens of refined associative learning theories might help other fields gain insight on this issue.

Nature and nurture positions can be stretched to such extremes that they disprove one another and end up on the verge of absurdity. On the one hand, nativists might say that without genes the organism would have never developed from the fertilized egg. On the other hand, advocates of the nurture position could argue that an organism cannot develop in a void; for genes to direct protein synthesis, exogenous sources and signals are required. Extreme propositions like these overlook the fact that both genes and environment are necessary conditions for building a phenotype (e.g., Narvaez et al., 2021). One way to respond to such statements is found in the misleading proposition that phenotypes are ∼50% due to genetics and ∼50% due to environmental factors. In a similar vein, claiming that the composition of phenotypic variance is simply intractable given current methodological hindrances may discourage further inquiry.

Despite representing more moderate views, both of these latter propositions overlook the fact that heritability, the canonical proxy for genetic contribution to phenotype, is a relational—rather than fixed—phenomenon (Bateson, 2001; Hinde, 1968). In short, the heritability (i.e., correlation between genotype and phenotype) of a given trait depends on the specific environment and population from which data are sampled (Rose, 2006; Sauce & Matzel, 2013). One common working example of this (although certainly not as clever as Sauce and Matzel’s case for scurvy) explains that the heritability of having two eyes in humans is close to zero. This is because genes rarely influence variation in this attribute given our typical environment; it is more likely to lose one or two eyes through environmental factors than to be born with none, one, or three eyes and then to transmit this distinctiveness to progeny. However, that is not to say that genes have no influence on having two eyes. In a similar vein, offspring often speak their parents’ language (i.e., high parental–offspring correlation), which is not to say that the language one speaks is genetically determined. The nature versus nurture debate often emerges whenever two factions do not agree on the relative contribution of genetic determinants to observed phenotypes under reasonably equivalent environmental conditions. This disagreement arises regarding studies comparing different species, subpopulations within a species, individuals within a population, stimulus modalities (like in Garcia and Koelling’s study), or response topographies (as in Breland and Breland’s paper).

The nurture position mainly objects to conclusions that are drawn when the reasonably equivalent-environmental-conditions assumption is suspected to be violated; that is, in cases in which internal validity is threatened. On the other hand, some nativists blame nurturists’ claims of “blank slatism” (e.g., Carl, 2018). The “blank slate” is used as a rhetorical weapon, a boogeyman based on John Locke’s radical empiricist doctrine, which current academics rarely (if ever) adopt. It could even be said in an a priori sense that blank slate assumptions cannot hold empirically, as they are based on a null hypothesis statement; that is, a proposition in which differences between two conditions are entirely absent. From a traditional statistical perspective, in the real world any null hypothesis can be rejected if the sample of observations is sufficiently large (McElreath, 2020). In this regard, one could declare the case closed in favor of alternative hypotheses that support nativist views, irrespective of the meaningfulness of the data. In a similar vein, any detectable alteration to behavior as a result of experience could be used to support the phenotypic plasticity thesis (i.e., the nurture position). But, again, how meaningful the effect is matters for characterizing the impact that plasticity has on actual organisms at every stage of life. Even when observing substantial behavioral alterations in the laboratory, some dispute whether those alterations can be extrapolated to real-world settings; that is, they express doubts concerning external validity.

Studying genes directly provides a powerful means to assess nativist assumptions in organic systems. In particular, combining correlational and experimental approaches has proved to be very productive (as argued in Sauce & Matzel, 2013). For example, recent studies have revealed that individuals diagnosed with idiopathic autistic spectrum disorder (ASD) exhibit differences involving genes that control differentiation, axon guidance, migration, and regional patterning in neural cells during early development (DeRosa et al., 2018). This is in line with prior reports of an overabundance of cortical neurons and excessive brain weight in individuals with ASD (Courchesne et al., 2011). Based on these findings, animal models were used to simulate the presumed neural development anomaly in idiopathic ASD. For instance, experimentally induced abnormal cell proliferation demonstrated to cause brain overgrowth and patterns of behavior judged as analogous to those of idiopathic ASD (Fang et al., 2014). These converging lines of evidence lead to the proposal of a meaningful genetic etiology for ASD (Courchesne et al., 2020).

Remarkably, research endeavors like the one just alluded are not always as successful in providing clear-cut evidence for specific genetic influences on behavior. Hereditary mechanisms, such as genetic control of neuronal migration, have also been hypothesized as causal factors for developmental dyslexia, given numerous reports of parent–offspring correlations (e.g., Wolff & Melngailis, 1994). However, researchers in this field have recently expressed concerns about the strength of the evidence in favor of this view. Guidi et al. (2018), for example, questioned how relevant the contribution of this etiology is to the phenotype of developmental dyslexia, given that genetic studies have failed to replicate key findings. Hence, caution must be exercised when drawing conclusions from family patterns in psychiatric conditions; parent–offspring correlations are not always indicative of clear-cut genetic transmission of maladaptive traits.

Nurturist claims also benefit from combining analysis of experimental and correlational evidence. Any systematic demonstration of phenotypical plasticity could be regarded as evidence in favor of the nurturist position. Such sources of evidence often come from studies in which randomized samples receive different experimental treatments and display differentiated expression of some attribute (morphological, functional, etc.). In the case of the learning literature examples abound, such as habituation, sensitization, and simple Pavlovian conditioning phenomena. While these examples are well known, even outside the field of learning theory, more impressive cases involving two-phase learning phenomena (see Byrom & Murphy, 2018) are rarely acknowledged. Some examples of these phenomena include latent inhibition, perceptual learning, overshadowing, blocking, sensory preconditioning, second-order conditioning, conditioned reinforcement, conditioned inhibition, Pavlovian-to-instrumental transfer, and partial reinforcement extinction effects. These phenomena demonstrate that environmental regularities can shape behavior not only in situ, but also that they in turn can influence the way in which organisms engage with new environmental conditions. An important implication is that genetic predispositions in learning rates or maximum capacity may give rise to mediated genetic–environment interactions (Byrom & Murphy, 2018). This follows from the assumption that the influence of the environment (behavior after the second stage of learning) on the phenotype (behavioral state after the first stage of learning) depends on the genotype (predisposed rate or upper bound of learning). According to Byrom and Murphy (2018), this brings about a complex, multicausal web of effects, as genes also depend on the environment (first stage of learning) to influence phenotypic expression (behavior after the second stage of learning) in the first place.

As suggested above, canonical learning phenomena demonstrate the plausibility of principled mechanisms for phenotypic plasticity in the domain of behavior. These demonstrations beg inquiry into how these mechanisms are implemented by organic systems and their implications for adaptation in concrete ecological niches. However, assessing how these learning phenomena interact with other factors is fairly difficult since they sometimes require multiple control conditions to be indisputably demonstrated (e.g., Papini & Bitterman, 1993; Rescorla, 1967). Numerous control conditions require increased sample sizes or else lessen the statistical power of a study; in turn, adding biological manipulations (ablation, drug, nutritional, etc.) to classical learning protocols requires further controls that multiply the existing ones. One increasingly common research practice involves assessing correlations between performance in learning tasks and real-life outcomes, such as health-related traits. However, low or null correlations have been routinely reported (e.g., He et al., 2013). This may be explained by poor psychometric properties associated with behavioral measures (Hedge et al. 2018). Specifically, these measures are well suited to detect the effects of experimental manipulations or situational factors but are often not sensitive enough to account for spontaneous inter-individual variability. In addition, real-life outcomes are often measured via self-report scales, which intend to capture behavioral patterns in a variety of unstructured situations; however, these measures risk involving biased subjective judgments (Dang et al., 2020) and a mismatch between the reference population and the study sample. For these reasons, correlations between performance on learning tasks and real-life outcomes are usually noisy and hard to ascertain. Thus, behavioral markers, while theoretically valuable (Byrom & Murphy, 2018) have been deemed unsuitable for diagnosis or prognosis (Dang et al., 2020).

One alternative path to studying the plasticity of behavior or the preparedness of complex behavioral patterns can be found in the use of simulated artificial agents. Such an approach employs rationally discoverable computation principles to achieve specific functions or on which to perform controllable interventions (Bongard & Levin, 2021). Unlike organic systems, artificial agents allow for expeditious exploration of the boundaries of nativist and nurturist assumptions that is virtually free of ethical conflicts. For example, one might ask whether complex forms of adaptive behavior can arise from a neural network with random connection weights (which is arguably tantamount to a blank slate). A study from Najarro and Risi (2020) found that this may be the case, reporting that artificial agents with just that feature were able to solve a two-dimensional car racing task and quadruped locomotion using Hebbian learning rules ¹ . However, possibility does not imply feasibility; in this sense, Zador (2019) has argued that training such networks requires enormous numbers of trials, which starkly contrasts with how organic systems learn. This riddle instantiates the challenges for external validity associated with nurturist appeals mentioned above.

As we can see in this section, important developments have boosted the study of genetic and environmental contributions to behavior. However, both sides of the coin endure their own biases and limitations. While they are difficult to overcome directly, being mindful of the diversity of factors that determine and underpin behavior can contribute to a nuanced understanding of this matter.

2. Typical environment and non-obvious experiences

Behavioral taxonomies frequently distinguish between learned and unlearned behaviors; the former can be defined as those that crucially depend on experiences and the latter as those emerging regardless of any experience. In traditional associative learning studies, the experimentally induced experiences affecting subsequent behavior are often well structured and share multiple aspects with the testing situation. However, in real life, experiences impacting a particular behavior do not need to be very similar to those in which a new behavior pattern is expressed. Turvey and Sheya (2017) assert that cases in which experiences affect behavior in a non-obvious manner often led to the misclassification of behavior as unlearned (or innate). A remarkable example is a study conducted by Masataka (1993). This author reported that fear of snakes is expressed in squirrel monkey specimens which ate freely moving insects but not in those that based their diet on fruit. Another example is provided by a study conducted by Moore (1992); the main finding in this study was that sexual behavior in adult rats was determined by differential interactions between males and females with their mother during infancy.

These findings provide evidence that experiences that occur during development in natural environments may be sufficient to determine behavioral traits that are typically exhibited by adults. Therefore, overlooking key features in the species-typical environment determining these traits may lead to incorrect conclusions. Importantly, crucial non-obvious environmental factors that determine phenotypes in general (Anlas & Trivedi, 2021)—and behavior in particular (see Kuo, 1976)—could be traced back as far as the gestational or pre-hatching stages. Note that, given this possibility, “unlearned” (not relying in experience) and “innate” (present shortly after birth) are not equivalent terms because innate behavior could be influenced by prenatal experiences.

Observation of shared phenotypic features within a taxa or populations can lead to the erroneous conclusion that they are genetically determined. Aspects of the environment that are taken for granted (non-obvious) may play an important role in influencing behavior and cognitive capacities. Some of them may be so ubiquitous in the species-typical environment that might preclude the emergence of alternative phenotypes. Another hallmark example of this possibility is the classic experiment conducted by Held and Hein (1963). In this study, light-deprived kittens failed to pass visual perception tests unless they were exposed to a condition involving visual feedback dependent on self-displacement. Importantly, the kittens exposed to equivalent displacement-contingent visual stimulation but in a fashion that did not depend on their own movements failed the visual perception tests. This finding indicates that an environmental aspect that is present during typical development, such as the non-obvious self-motion dependent visual experience, plays an important role in perceptual abilities in later life.

Pointing at a set of behavioral phenotypes that is characteristic of all members of a species is surely of descriptive value; however, comparative cognition researchers should inquire into what would have happened if those subjects were removed from their typical environment (at least, in a purely counterfactual thought exercise). At the whole-organism level of analysis, the typical environment could either be a species’ natural niche or artificial laboratory housing (see DiFrisco & Jaeger, 2020 for other levels of analysis; e.g., cellular, tissular). Importantly, key aspects may be present in both natural and artificial settings. This is the case of displacement-dependent visual stimulation for cats, which is (almost) inevitably present both in natural and laboratory settings. Another, perhaps more controverted example, is language development in humans. At a certain age, all human beings exhibit language skills unless they are severely impaired. This might lead some to conclude that language is an unlearned or innate faculty. However, to survive to that age, humans must be nurtured by other humans. There are no examples of humans surviving from birth to maturity away from a cultural environment so that we can unequivocally assess the contribution of nurture to language competence. The unavailability of this counterfactual condition (i.e., what would have happened holding all else constant) prods researchers to consider possibilities beyond the available data.

3. The environment from the perspective of the subject

Held and Hein’s (1963) findings mentioned in the section above demonstrate that the environment is more than a static arrangement of stimuli surrounding subjects. Aspects that are immediately and continuously contingent (i.e., statistically dependent) upon subjects’ actions can be considered part of the subjective environment (or unwelt; see Gomez-Marin & Ghazanfar, 2019) and have meaningful and lasting effects on behavior. An account proposed by Powers (1973) and taken up recently by others (Gomez-Marin & Ghazanfar, 2019; Yin, 2020) states that behaving agents can be characterized as closed-loop negative feedback control systems. Yin (2020) asserts that behavior should not be conceived of as a unidirectional input → output process; instead, behavior is composed of inputs guiding actions and actions guiding inputs in a closed feedback loop. The input → output sequence is just a part of this whole interaction, while the output → input unit is equally crucial but is frequently disregarded and must be addressed as well (see also Brembs, 2020). Thus, a minimal model of behavior (or else cognition; see Brancazio et al. 2020) should consider it an emergent phenomenon composed of a collection of closed loops. This rationale has been assumed to hold for single-celled organisms (e.g., Bich & Bechtel, 2022) all the way to multicellular organisms with complex nervous systems (e.g., Sosa & Alcalá, 2022). Previous emphasis on the unidirectional input → output component may, in part, be explained by the relative ease with which researchers are able to manipulate “what happens to the subject” (Yin, 2020). However, actual behavior is likely more complex than often assumed.

This account faces the challenge of being rather counterintuitive. When presenting a stimulus in a laboratory setting (or elsewhere), researchers often assume that it remains invariant throughout its duration. However, from the subject’s perspective, the stimulus may vary substantially before, during, and after an action has been performed in response to it. For a behaving agent, an action that occurs at a given moment, t _i , will modify the ensemble of sensory inputs at a subsequent moment, t _i+1 . This occurs because motor systems are coupled with proprioceptive receptors or can indirectly alter the signals that other sensory modalities provide. Such output-dependent input variation has meaningful implications whenever sensory inputs systematically trigger further actions or possess system-disturbing properties, which in practice is often the case. Researchers may have some control over what happens to the subject at t _i , but they will know little of what happens thereafter from the subject’s perspective. In short, a manipulated stimulus event cannot be equated with an actual input (Yin, 2020). Therefore, it is best to assume that a stimulus function or stimulus gradient that depends on the subject’s action exists from the onset of the stimulus event. This would be a particularly fertile framework for comparative cognition and behavioral neuroscience. Such research programs often aim to disentangle or even integrate cognitive processes and associative learning to account for seemingly analog behavioral patterns in different taxa (Zentall, 2013).

As a working example, we can consider a counterintuitive finding in a recent cognitive neuroscience study conducted by Basile et al. (2020). These authors reported that bilateral lesions in the hippocampus leave macaque monkeys’ performance virtually intact in several memory tasks. This finding controverts the dominant notion that the hippocampus plays a role as a “memory storing” module. However, Basile et al. (2020) attempted to conciliate their results with previous findings by taking into account their subjects’ perspective. Lesion studies in rodents and clinical reports in humans have indisputably related the hippocampus with memory performance. Nevertheless, tasks for assessing memory in rats and the way in which humans learn from their natural environment (e.g., Capaldi & Neath, 1995) may differ in a crucial way from the conditions in Basile et al.‘s study.

Importantly, in this study all the tasks for assessing memory capacity were programmed on a touchscreen interface. Basile et al. suggested that the hippocampus plays a major role in spatial allocentric memory used for navigation (see Robinson et al., 2020), rather than in general associative memory. Given that interacting with a touchscreen does not require displacement, which presumably would have required substantial recruitment of the hippocampus, performance was not disrupted by the lesions. The authors underscored that most paradigms used to assess memory in rodents involve displacement between different locations; even if such a feature of the task is irrelevant from the perspective of the researcher, it inevitably impacts performance and, thus, the interpretation of data. Spatial navigation, a variable that is dependent on the subjects’ actions, may act as a confounding factor in canonical studies on general memory.

Altogether, Basile et al. (2020)’s findings illustrate that anchoring cognitive capacities on intuitive taxonomies detached from careful study of behavior is problematic (e.g., Cisek, 2021; Pessoa et el., 2022). Rather than regarding brain structures as modules that perform certain computations, considering the unique way in which organisms experience and act upon their environment has been proposed as an upgraded approach to make sense of the neurobiological basis of behavior (e.g., Krakauer et al. 2017). This perspective highlights the dangers of taking for granted that the nominal task requirements are mere triggers for the capacities that the brain possesses. Viewing the brain as a biological container of cognitive capacities lends to the assumption that specific genetic programs directly lead to the development of and thus cause said capacities. Therefore, modeling behavior as an emergent complex of action–perception loops help prevent unwarranted claims for belongingness and similar biological predispositions.

4. Uses and abuses of twin studies as natural experiments and the animal model solution

One of the most powerful methodological tools available for inquiring into the contribution of genes to behavior (or any phenotypic trait) is the comparison of monozygotic twins, dizygotic twins, fraternal twins, and unrelated individuals. Studies of monozygotic twins reared together are assumed to abstract the contribution of the unique environment from within-twin variation. The discrepancy in the similarity of traits between monozygotic and dizygotic twins is usually invoked to estimate the contribution of genetic factors. The discrepancy in the similarity of traits between pairs of twins (either monozygotic or dizygotic) reared together and reared apart is used as a proxy for the contribution of the shared environment. For example, in a recent study, Baxi et al. (2020) scanned the brains of a large sample of monozygotic and dizygotic twins. The main findings were that the thickness of the gray matter showed a greater genetic contribution, while the microstructure of the cortex (e.g., the cytoarchitecture and gyrification) exhibited a relatively greater unique environment contribution. Although this study did not use an experimental design, the authors suggest that the unique environment may play an important role in differences in the personality traits and mental health of monozygotic twins. It is worth recalling that the estimates for the contribution of genetic and environmental factors are crucially dependent on the surroundings in which the population develops (Sauce & Matzel, 2013). Therefore, these estimates may at least partly be constrained by the geographical and cultural circumstances from which the sample was extracted.

Despite these sorts of studies being deemed as “natural experiments,” their interpretation presents some difficulties, which arise from a lack of control over several relevant factors. For example, we might consider whether and to what extent twins who are reared apart develop in meaningfully different environments (see Moore & Shenk, 2017). In addition, most twin studies are blind to complex genotype–environment interactions (Jaffee & Price, 2007). For example, one’s physical appearance is influenced by genes, but could also determine how others treat us (a non-obvious environmental feature), which in turn might shape our behavior (Scarr & McCartney, 1983; but see Conley et al., 2013). Furthermore, giving birth to twins in humans is not random (Bhalotra & Clarke, 2019), bringing into question the validity of twin studies as well-controlled natural experiments. Yet another problem is that twin studies often rely on the assumption of random mating between parents, for which there is evidence that it is not met (Yengo et al., 2018). This later issue could impact estimates of genetic contributions to a nontrivial degree (Vinkhuyzen et al. 2012).

In light of the problems listed above, some authors have called for studying the relative contributions of genetic and environmental factors to behavior by embracing a true experimental approach. This can be reached using randomized interventions in human twins (Burt et al. 2019) or in animal models of twinship (Briley et al., 2019; Byrom & Murphy, 2018) to test the boundaries of phenotypic plasticity holding genetic factors constant. The nine-branded armadillo (Dasypus novencintus) can help to circumvent some of the problems with human twin studies. One relevant peculiarity of this species is that females give birth to monozygotic quadruplets; this potentially allows for a more powerful and controlled assessment of the relative contribution of genes, shared environments, and unique environments to phenotypes. For example, Ballouz et al. (2019) compared genetic variability (allelic expression imbalance) in armadillo quadruplet litters living in captivity starting from the gestation period with that reported in human twins. This study’s main finding points to substantially higher epigenetic variation in humans when compared to captive armadillos. The authors suggest that such a degree of variation may arise because human twins can accumulate sufficiently unique lifestyle experiences given the non-controlled features of the environment in which they typically develop. Although the main dependent variable in this study was allelic expression at a nucleotide level, the authors hypothesized that said factor may have key phenotypic effects downstream. For instance, the authors incidentally reported that variability in allelic expression was correlated with weight variability within the armadillo litters.

5. Unique environment and stochasticity

Researchers from different groups employ different terms for conceptualizing the source giving rise to traits that are neither related to genes nor to known variations in the environment. Some researchers use the term “unique environment,” and others prefer “stochastic variation” (e.g., Mitchell, 2021). While both terms usually refer to the same portion of variance in different studies, they seem to denote different underlying assumptions. On the one hand, the term “unique environment” seems to imply a set of experiences that are not controlled for in the study design. This factor would influence behavior (or any other phenotype) via an equivalent mechanism as that through which the shared environment does; the crucial difference is the fact that one of these environmental factors is actually traced, while the other is not. The term “stochastic variation,” on the other hand, implies that there is a third—qualitatively different (see Mitchell, 2021)—factor in addition to genes and environment for determining behavior (or any phenotype).

Honegger and de Bivort (2018) suggest that when genes and the environment are matched and phenotypic differences are still observed it is likely due to stochastic influences. The stochastic factors that they appealed to are not unknown quantum operations (see Bell, 2014) but rather phenomena that can be understood within Newtonian physical rules. These authors argue that most of the stochastic processes that give rise to individual behavioral differences come from chaotic molecular (e.g., gene transcription) and cellular (e.g., axonal guidance) events. Specifically, small differences in initial conditions could lead to relatively large effects that can be further amplified by positive feedback mechanisms at higher levels. Honegger and de Bivort (2018) list a number of sound adaptive advantages of a built-in (or endogenous) susceptibility to stochastic variation. Furthermore, they provide examples of brain mechanisms presumed to generate behavioral variability in non-vertebrate species. The authors recognize the possibility that other factors may play a role in individual differences, such as phenotypic plasticity. This factor would involve adjusting some traits of the biological agent in response to the sensed environment, ideally in order to optimize some resource-exploitation process. However, Honegger and de Bivort (2018) discard this factor as a relevant influence on individual differences in behavior because, according to them, embodied morphing rules as such would be costly to evolve. Therefore, the authors chose stochasticity-generating genetic mechanisms to account for most within-species individual differences, as they represent a solution that is readily attainable by evolution.

The rationale advanced by Honegger and de Bivort (2018) requires, at least, some nuance. The evidence and arguments they provide in support of the plausibility of genetic-based mechanisms to give rise to stochasticity are compelling. However, this kind of mechanism’s contribution to individual differences may vary from one taxon to another. Some animal lineages have evolved particularly malleable nervous systems for implementing morphing rules to adjust their behavior in response to environmental regularities (Fanselow & Wassum, 2016). According to Honegger and de Bivort (2018), this is especially complex and costly; therefore, it may not apply to species with relatively simple nervous systems that are more bound to stereotypical behavioral routines throughout their lifecycles. In addition, the notion of a chaotic flow from equivalent states to individually differentiated paths could also be applied considering learning theory. Initial experience-mediated behavioral changes would lead organisms to engage differently with their environment, further shaping the behavioral phenotype in a given direction (Byrom & Murphy, 2018). Such a process can be distinguished from purely biological stochastic events in that the former involves the environment’s impact on receptors in the nervous system as proximal causes. In short, stochastic noise in the environment may indeed shape individuality; if so, it would have to be via (known and yet to be known) principled rules of behavior change as well as by other means.

Some researchers have already highlighted the need to identify the mechanisms through which behavioral traits are acquired. For example, Zietsch and Sidari (2019) have asserted that the mechanisms that lead genetic variation to give rise to behavioral traits (or phenotypes in general; see Jaeger et al., 2012; Uller et al., 2020) are, in the best case, underspecified. Herein lies an opportunity for the field of learning theory. As mentioned above, the literature on learning is full of studies that show how structured experiences lead behavioral changes in situ as well as in transference situations. In a given situation, behavior is determined by the configuration of the nervous system, such that (all other things being equal) changes in behavior require a disruption of such a configuration.

Unfortunately, documenting specific plastic changes in the nervous system that arise from particular experiences is logistically difficult. However, there is increasingly compelling evidence that changes in the composition of certain nervous system sites can emerge from typical learning protocols and how those changes are disrupted by adverse early experiences. One example can be found in the literature on latent inhibition. This phenomenon consists in learning to ignore repeatedly presented stimuli, which is considered an adaptive behavior because of the importance of disregarding uninformative events in our surroundings (Costa et al., 2021). Moreover, deficits in this process have been associated with psychotic proneness (Lubow, 2005) revealing its impact on health. Recently, Jacob et al. (2021) found that stimulus pre-exposure lead to particular activity patterns in certain dopamine sites. In turn, activation of those sites was required to learn from pre-exposure experiences. Complementarily, Han et al. (2012) showed that peri-adolescent social isolation led to deficits in learning from stimulus pre-exposure and that this disruptive effect was mediated by the expression of certain dopamine receptors. Then again, physical appearance (arguably, in part, genetically determined) is associated with peer acceptance and thus with varying degrees of social withdrawal (Vannatta et al. 2009). Altogether, these sources of evidence reveal the intricate causal webs in which genetic and environmental factors can interact to give rise to behavioral variation. Accounting for learning principles informed by neurobiological mechanisms, in addition to genes and genuinely stochastic processes, would provide a better principled understanding of individual differences in behavior.

6. Neural computation versus biomechanics

As described in the previous section, behaviors that are peculiar to an individual, a species, or a developmental stage are constrained by neural architecture (either genetically determined or acquired). However, caution must be exercised when proposing this kind of mediation, at least in the absence of reasonable evidence for the specific neural underpinnings involved. Gomez-Marin and Ghazanfar (2019) provided two great examples of how overlooked interactions of biomechanistic and environmental factors could dictate, in some cases, peculiar behavioral unfolding. First, Zhang and Ghazanfar (2018) placed adult marmoset monkeys in an artificial gaseous environment lighter than air and reported that they produced vocalizations typical of young marmosets. Such a finding led these authors to hypothesize that the transition from infant to adult vocalization is influenced by the mechanical properties of their lungs, rather than solely by neural maturation. Second, Thelen (1984) showed that human infants around 2 months of age also regress to behaviors that are typical of a prior developmental stage under appropriate environmental conditions. She decreased the gravitational force through buoyancy by partly submerging infants in water. As a result, infants performed motor patterns that are typical to the previous stage of development. This finding is remarkable because changes in motoric patterns at early stages are often attributed solely to mechanisms of neural development. In some fields of developmental science, a common interpretation of behavioral changes across the life span is that a dormant genetic program is unpackaged in the brain at certain ages, perhaps with the help of certain trigger signals. Both these examples challenge explanations inclined to claim neural computational differences whenever distinctive behavioral patterns are observed.

I will provide one more example to supplement those by Gomez-Marin and Ghazanfar (2019). Early reports featured evidence that a specific gene mutation in Drosophila flies, making them yellow in color instead of the typical gray, showed a reduced fertilization rate for males (Sturtevant, 1915). Observation thereof led researchers to suggest that an alteration in the neural circuitry controlling courtship behavior contributed to lessen mating success (e.g., Bastock,1956). ² However, Massey et al. (2019) found that for intact individuals, the so-called yellow gene determined a morphological feature in males’ legs that helped with clutching females for copulation, thus, bolstering reproduction. Therefore, males with the mutation had difficulty mating due to mechanical properties associated with their surface anatomy, rather than due to alterations in neural functioning. While there are some compelling examples of genes affecting courtship behaviors through neural means (see Yamamoto & Koganezawa, 2013), Massey et al. (2019) urged abandonment of the idea that all so-called instincts are somehow “hardwired” into the nervous system (e.g., Stockinger et al. 2005). Instead, these authors argued that the mechanical properties of the anatomy must be considered in order to fully understand behavior.

7. Belongingness in the light of phylogenetic refinement of behavior

Cisek (2019) suggested that phylogenetic refinement may help to build strong theories of behavior to inform research on comparative cognition. The basic idea is that every behavioral mechanism is an extension of an ancestral one, which appeals to the principle of functional continuity. According to Cisek (2019), this framework might be an alternative to cognitive modules (e.g., attention, memory, executive function) as a framework for comparing animal behavior across species. Cisek (2019) argues that the phylogenetic refinement approach will lead to the development of conceptual taxonomies that more readily reflect the nature of biological systems that give rise to complex animal behavior (e.g., nervous systems). Within this framework, history begins with a unified function. This author suggests that behavior is a specialized closed feedback loop like the approaches described in previous sections. This emergent feature is thought to have aided metabolism in the beginning to maintain a relatively stable, self-perpetuating homeostatic state. The nervous system function underlying this ancestral form of behavior evolved to take over interactions with the external environment. This process endowed some organisms with increased control over the external environment, which required key changes in their neurobiological organization. In short, metabolism is an archaic control mechanism in which behavior is embedded, and the latter has further developed increasingly complex nested features in select species. This could lead to characterizing animal behavior as an assembly of parcellated sensory–motor loops varying in sophistication. Interestingly, according to Yin (2020), advanced neuroscience techniques would help to elucidate the pathways and nodes in neural networks that underpin this hierarchical organization of behavior.

Some of the elements that constitute the behavioral repertoire in an organism could be recruited in a parallel or alternate fashion. This implies mutually exclusive systems competing at times to obtain or maintain control according to a hierarchical scheme. The performance of these systems is determined by the architecture of the nervous system as well as by transitory states informed by internal and external environments. The activity associated with the elements in charge of a behavioral function is tuned to respond to a particular class of trigger events and is abolished by another class of events. However, in complex systems, sensory–motor loops could become less fixed and gain “autonomy” by anticipating future outcomes from different cues via certain decoupling (learning) mechanisms. From the perspective of associative learning, experience with environmental statistical regularities is required for anticipation to happen. In addition, functional decoupling can be understood as changes in the degree to which certain environmental features affect (increase or decrease, in real time or prospectively; see Sosa & Alcalá, 2022) the activity of a given effector system.

Environmental pressures, such as conspecific competition, predation, scarcity, and unpredictability, may have fueled the evolution of this capacity for learning in some species. Hebb (1949) was a pioneer in suggesting that the internuncial portions of nervous systems (such as the associative cortex in mammals) are pivotal for learning and performance of complex behavioral patterns. He argued that animals with rigid behavioral repertoires have relatively large portions of primary sensory and motor nervous tissues compared to internuncial tissue. In contrast, a smaller ratio of primary to internuncial tissue leads to a capacity for finer discriminations to anticipate imminent outcomes, and thus to actions that exhibit less control from the apparent immediate environment. Relatively large internuncial regions enable organisms to display more seemingly spontaneous (decoupled) activity, but at the cost of slower initial learning rates. That is presumably one of the reasons for the atypically long period it takes human beings to develop proficiently sophisticated skills compared to other species.

Recent studies have captured the hierarchical organization that Hebb (1949) hypothesized. For example, Fox et al. (2020) found that the elicitation rate (i.e., the proportion of times that brain electrode activation reliably led to detectable responses in awake persons) of different regions in the human cortex correlates with its functional anatomical hierarchy. Specifically, stimulation of primary (unimodal sensory or motor) areas showed higher elicitation rates, and this measure decreased as the distance from these primary areas increased. In addition, the elicited effects became more heterogeneous, ascending the cortical hierarchy; that is, stimulating primary areas mainly elicited effects of a single type between participants, while stimulating areas connectionally distant from primary areas exhibited a wider variety of complex effects. Interestingly, a similar gradient of functional connectivity has been reported in other species. Huntenburg et al. (2021) identified several different gradients from primary sensory areas to more transmodal regions in the mouse cortex; importantly, gradients corresponded with sensory specialization in this species. For instance, the authors reported a gradient separating audiovisual from somatomotor areas and another separating the snout from the lower limb somatosensory regions.

In conclusion, from an evolutionary point of view, sensory and motor systems that are coupled and involve meaningful feedback loops benefit from efficient transmission (involving “hardwired” spinal or brainstem circuitries in chordates) and usually implement relatively stereotyped forms of behaviors, such as reflexes. On the other hand, sophisticated behavioral patterns that are not universally required for survival or reproduction can be learned. Their implementation demands neural tissue that is flexible enough (e.g., higher order association cortices) to couple sensory and motor assemblages that are not already linked, and extensive experience would be needed. In the middle of this spectrum lies the case of behaviors that are underpinned by cortical loops that are least separated by internuncial tissues and thus learned more readily. The latter would constitute a putative neural basis for so-called belongingness.

8. Miscellaneous examples for analysis

The phylogenetic refinement approach to behavior that Cisek (2019) proposed is quite compelling and promising as a conceptual tool for understanding belongingness, learning constraints, and differential cognitive capacities. However, in its current state, it does not commit to any particular characterization of behavioral change. Integrating Cisek’s (2019) approach and the notions described in previous sections with key findings from the fields of associative learning, comparative cognition, and behavioral neuroscience would be of great value. In the following subsections, I will attempt to do just that and will revisit the notions presented above to approach representative findings from these research domains. First, it is important to characterize one of the most elemental forms of behavioral change, reinforcement learning, and a possible mechanism for its pervasiveness.

8.1. Goal-directed actions and habits in reinforcement learning

Animals possess mechanisms to track resources in their environment, adjusting their actions to obtain (near) maximum benefit. This involves the capacity to select whether and when to enact specific behaviors. A reinforcing event or reward is one that makes a particular behavioral pattern more likely to occur in similar conditions when contingent (i.e., statistically associated) upon it. For this to happen, the detectable environmental features that are present when the reinforcer occurs are said to acquire some associative value. One implication is that those cues will come to control the behavioral pattern reinforced in their presence (i.e., discriminative control). This process also implies that said features will reinforce response patterns themselves; that is, they become conditioned reinforcers (Sosa et al., 2011; Sutton & Barto, 2018). In addition, features in the environment (objective and subjective) that are associated with rewards become attractants that behaving agents pursue in real-time (Karin & Alon, 2021).

Changes in the associative value of cues according to their co-occurrence with reinforcement have been formally accounted using a number of computational models. The basic rationale behind these models is that correcting prediction errors determines learning (potential behavioral change as a consequence of experience). In short, when a cue or an action is followed by an outcome (e.g., a reinforcer), the less an outcome is predicted by prior experience with environmental regularities the greater the change in the associative value of the preceding events. Thus, prediction errors (the discrepancy between actual and expected outcomes) are assumed to be an important driver of behavioral change (Iordanova et al., 2021). Recent evidence indicates that there are dissociable Pavlovian (stimulus–outcome) and instrumental (response–outcome) error correction mechanisms (Bouton et al., 2020).

In the initial stage of instrumental learning, actions are said to be goal directed; that is, the performance of these actions is sensitive to several forms of outcome revaluation. As these encounters of the action and the reinforcer accumulate, however, the former becomes increasingly insensitive to outcome revaluation. This process is known as “goal-directed to habitual control transition” (Perez & Dickinson, 2020). Habits have been documented to involve rapidly recruited and topographically invariant actions, as well as insensitivity to consequences and contextual changes (Balleine & Dezfouli, 2019). Recent models, however, have shifted the characterization of goal-directed and habitual control as differing in terms of the source of feedback upon which the subject relies while executing motor actions. In particular, some authors (e.g., Balleine & Dezfouli, 2019; Cleaveland et al., 2018; Miller et al., 2019; Sosa & Alcalá, 2022) claim that habits—rather than being “automatic” or insensitive to consequences—heavily rely on response-produced feedback, such as proprioceptive cues. This implies that such self-produced cues could function as discriminative stimuli, conditioned reinforcers, and attractants. This conceptual shift dismantles (or redefines) the apparent distinction between volitional (spontaneous) and automatic (decoupled) behavior in favor of a more parsimonious account; namely, one in which behavior invariably entails some kind of feedback, and the resulting perceptual-control loop could transition from less reliable to more reliable feedback sources (for the organism, importantly). An obvious methodological complication to test this assumption is that many sources of response-dependent feedback cannot be easily observed or controlled by experimenters in purely behavioral studies. However, recent advances in the field of behavioral neuroscience have shed some light on this issue.

Firing by canonical midbrain dopaminergic neurons is believed to underly positive reward prediction errors (Lee et al., 2020)—that is, the mechanism driving changes in behavior when the organism encounters unexpected rewards. In other words, these dopamine signals are, in part, the cause for the shift in the reinforcing function from the actual reward to the antecedent stimuli upon positive reward prediction errors. Correspondingly, acute depressions (dips) in dopamine release facilitate inhibitory learning (e.g., Chang et al., 2018) when expected rewards are omitted (i.e., negative reward prediction error). The terminals of midbrain dopamine cells innervate different locations of the striatum; neurons from the ventral tegmental area project their axons to the ventral and mediodorsal portions of the striatum, whereas those from the substantia nigra project mainly to the dorsolateral portion. Tsutsui-Kimura et al. (2020) reported that dopaminergic axons showed distinctive activity patterns in response to environmental events across different striatal regions in mice. Specifically, in a well-trained instrumental task, the mediodorsal portion of the striatum was sensitive to reward-anticipating exteroceptive stimuli, unexpected presentation or omission of reward, and variations in reward magnitude. In contrast, dopamine activity in the dorsolateral portion was insensitive to reward omission and showed excitation rather than inhibition to smaller than expected rewards.

Tsutsui-Kimura et al. (2020) suggested that reduced sensitivity to negative prediction errors in the dorsolateral striatum may be critical to the maintenance of habitual actions. The main non-midbrain inputs to dorsolateral striatum come from the primary motor and somatosensory cortices, whereas the dorsomedial striatum receives relatively more inputs from prefrontal, associative, and other primary cortices (Khibnik et al., 2014; Voorn et al., 2004). Thus, the dorsolateral striatum may be predominantly activated by sensory information that is often not directly manipulated by researchers in freely moving paradigms—that is, from tactile and proprioceptive sources. Exteroceptive sensory pathways may have a temporal deadline for “detecting” the absence of a reward’s material features for producing the dopamine dips needed for negative prediction errors (see Mollick et al., 2020). In contrast, unless physically blocked, many of the actions associated with reward tracking (along with their coupled feedback) may be present, even if the reward is ultimately not encountered. This may explain the absence of negative prediction errors in the recorded activity of the dorsolateral striatum in Tsutsui-Kimura et al.‘s study. Proprioceptive prediction errors might not have occurred simply because researchers were able to prevent reward delivery, but not response performance. Interestingly, the authors also speculated that the associative strength of the “motor states” related to action performance might serve to chain responses to produce sequences of movement, and to do so smoothly and in a seemingly automated fashion (see also Enquist et al., 2016; Sosa & Alcalá, 2022). It seems plausible that each instance of proprioceptive feedback could serve as an attractant and conditioned reinforcer to that specific action if it eventually leads to primary reinforcement; in addition, those cues could serve as a discriminative stimulus for properly setting the conditions for generating the next action and so on.

8.2. Sign-tracking, goal-tracking, and misbehavior

Aside from the García–Koelling effect, which was observed in aversive conditioning settings, studies on Pavlovian appetitive (i.e., reward-related) conditioning have also reported remarkable interaction effects; in particular, between the modality of the conditioned stimulus (CS) and the topography of the conditioned response (CR). Pairing a reward unconditioned stimulus (US) with a compound CS consisting of a focalized light and the availability of a manipulable object (such as a retractable lever), leads some rats to approach and manipulate such object. Importantly, the reward is delivered and can be consumed by the subjects, regardless of whether the approach behavior occurs. Subjects presenting a relatively high proportion of approach behaviors are labeled as “sign-trackers ³ .” These behaviors have been assumed to arise from incentive salience imbued to the visually conspicuous stimulus. Evidence supporting this contention can be instantiated by the observation that sign-trackers respond relatively more for conditioned rewards than so-called “goal-trackers” (i.e., subjects that pursue the food source without approaching the focalized CS; Robinson & Flagel, 2009). In contrast, other types of CS do not elicit approach behaviors despite being paired repeatedly with a reward. For instance, an auditory CS paired with reward delivery does not elicit approach behaviors toward the source of the stimulus; however, it serves to maintain responding as conditioned rewards (Meyer et al., 2014). This finding has perplexed researchers because rats are well capable of detecting the source of an auditory CS (Cleland & Davey, 1983) but still do not approach it when paired with reward.

This pattern of results has been interpreted as indicating that sign-tracking behavior and conditioned reinforcement are underlaid by dissociable mechanisms but share the involvement of incentive salience (Meyer et al., 2014). However, from the subject’s perspective, both phenomena can be regarded as functionally equivalent but occurring at different timescales. For both sign-tracking and conditioned reinforcement, an increase in some physical parameter associated with the stimulus is contingent upon the subject’s specific actions. While some aspects support homology between the two conditions, a crucial distinction is also evident. The most obvious commonality, again, is the involvement of a cue associated with reward. When it comes to sign-tracking, the onset of this cue depends on the researcher. On the other hand, in conditioned reinforcement, the onset of this cue is programmed to occur iff the subject performs an arbitrary response. Therefore, for sign-tracking, the subject only controls the stimulus’s physical properties while the stimulus is switched on. Importantly, in both paradigms, the physical properties of the reward-related cue are under the subject’s control, albeit to different degrees.

These notions could serve to formulate a working hypothesis to account for the interaction between the sensory modality of the CS and the response topography of the CR in appetitive Pavlovian conditioning. In short, the gradient by which the subjects are able to control the stimulus while it is present may differ crucially from one modality to another. On the one hand, some stimuli, such as a focalized light or a retractable lever, may produce relatively abrupt (thus, more salient) gradients of change in the perceptual experience contingent upon actions performed by the subject. On the other hand, the acoustic experience may not change sharply upon approaching behaviors performed by the subject for as long as the tone is on. In contrast, in a conditioned reinforcement preparation (see Meyer et al., 2014), the tone goes from off to on, concomitant to an action of the subject. That is, the transition from the off to the on state is maximally abrupt. This may explain why rats readily perform actions that turn the reward-associated tone on, but do not approach the source of this tone when the researcher has already turned it on.

Although this explanation may not be taken as conclusive, it at least prevents the formulation of assumptions based on differing organization of neural circuitries involved in each stimulus modality. As Yin (2020) outlines, regarding behavior as a mere function of manipulated stimuli could be misleading. In such a limited framework, the correlation between stimulus and response could be wrongly assumed to be a function of the nervous system’s configuration. Rather, researchers must consider the possibility that inconspicuous properties in the subjective environment play a nontrivial role in controlling behavior. Such an approach is very useful in terms of its integrative potential. For example, taking into account stimuli from the subject’s perspective allows us to consider findings from other fields under the framework of sign-tracking. A good example of this is oculomotor capture by visual stimuli paired with reward in humans, even when it is counterproductive (Le Pelley et al., 2015). This finding is functionally equivalent to sign-tracking but, to the best of my knowledge, this similitude has yet not been recognized and elaborated.

The view of sign-tracking as behavior maintained by feedback that is inevitably contingent on the approaching response is not new (see Burgos, 2007; Enquist et al., 2016). For instance, Enquist et al. (2016) put forth a model in which behavioral sequences could arise efficiently by chaining actions via associative mechanisms, although with some contribution from innate biases. According to them, conditioned reinforcement and discriminative functions from cues in the final links backpropagate to those in the initial links of the behavioral chain. Efficient learning of behavioral sequences could emerge if the cues in the final links readily accrue conditioned reinforcement properties, then backpropagating their value to cues occurying reliably ahead of them and so on. However, maladaptive behavioral tendencies arise if this happens in cues associated with the first links of a stipulated chain of actions needed for a reinforcer to occur. In that case, the animals engage in spurious self-reinforcing loops, occasionally stumbling upon reinforcement but at a low rate and spending energy unnecessarily.

The above-mentioned rationale allows Enquist et al. (2016) to elegantly account for adaptive behaviors, such as tool use, and maladaptive behaviors, such as classic sign-tracking and misbehavior (see Breland & Breland, 1961). In their model, a set of parameters may be varied so that a given behavioral sequence could be characterized as a multidimensional continuum; namely, according to the ease of certain sensory–motor sets to associate (i.e., belongingness) and their position in the chain. Accordingly, a particular behavior sequence could range from being deeply hardwired (i.e., readily unfolded without extensive training) to malleable, but could also be unlearnable (i.e., learning constraint) or unyieldingly maladaptive (i.e., misbehavior). The strength of this model can be found in its ability to bridge the fields of behavioral ecology and computational psychiatry via the principles of associative conditioning. It would be interesting to explore whether it is capable of incorporating the notion of feedback (attractant) gradients presented in the paragraphs above to account with more detail for the conditions in which sign-tracking is not unfolded.

8.3. Human speech: characterization, acquisition, and performance

Another working example for applying the notions presented in this paper comes from a seemingly disparate field, namely, that of human speech. Humans have evolved a remarkable device for symbolic communication that requires coordination of vocal, respiratory, and orofacial movements to produce coherent strings of sound. It is generally accepted that this capacity depends to a large degree on certain unique features in neural organization. However, the degree to which experience is needed for human speech remains a matter of debate and thus qualifies as a belongingness-related issue. For example, Chomsky (2006) asserts that language is an innate scheme in human minds, while Skinner (1957) believed that language is socially shaped via explicit and tacit contingencies of reinforcement. As with any other forms of behavior, human speech may be conceptualized as a closed feedback loop. Under this assumption, some of the human brain’s unique features should be understood as adaptations contributing to the sensory–motor coordination that the implementation of speech requires. The brain and other parts of the human body may be organized to make this form of symbolic communication not only possible, but also efficient. This rationale applies for both the acoustic structure and the pragmatics of spoken language. Here, I will focus on the former.

Humans and songbirds possess similar neural circuits that are not found in other species, which could be considered a case of convergent evolution (Nieder & Mooney, 2020). Vocalization in birds and mammals recruits respiratory and upper vocal tract motor neurons. Species in these taxa vary in their involvement of the forebrain for providing input to the motor nuclei in which these neurons are located, in the spinal cord and brainstem. Both songbirds and humans represent the extreme end of this spectrum, presenting numerous forebrain regions that provide input to downstream respiratory and vocal circuits. These latter ancestral circuits serve as a scaffold for complex acquired vocal control. Looking further upstream, forebrain motor control regions receive inputs from premotor regions, which in turn receive inputs from associative sensory areas and so on. In primates, two main (independent, yet interconnected) cortical sources for vocal control have been identified: the anterior cingulate and the laryngeal/orofacial motor cortex (Nieder & Mooney, 2020). On the one hand, the anterior cingulate is primarily involved in emotional species-specific vocalizations and emotional intonation of “spontaneous” vocalizations (Jürgens, 2002). On the other hand, the laryngeal/orofacial motor cortex has been associated with learned vocal production (Nieder & Mooney, 2020). The human motor laryngeal cortex exhibits higher connectivity with sensory cortices than that found in macaques (Kumar et al., 2016). This is reflective of the fact that human vocalization is relatively more susceptible to be brought under the control of multiple subtle and complex environmental cues; among the most important of such sources of information are conspecifics’ and their own vocalizations.

Human vocal learning seems to critically depend on the capacity to imitate vocal expressions from conspecifics. Importantly, the ability to imitate others’ vocalizations is strengthened by hearing both the imitée and the subject’s own vocalizations. Such a faculty provides an efficient means for perceiving the discrepancy between someone else’s and our own acoustic production to engage in error correction. The perception of phonetic patterns has been hypotesized to rely on error prediction mechanisms (Divjak et al., 2020; Sohoglu & Davis, 2020) like those studied in classic associative learning experiments (Polack & Miller, 2018). This would allow for the acquisition of perceptual learning rules that are particular to our own language and it occurs in early infancy, long before effective vocalization (Mueller et al., 2012; for a homologous process in songbirds see Goldstein & Schwade, 2010). At a later stage, infants begin to babble, which helps them to associate the enactment of motor command patterns with specific acoustic patterns. Then, as they continue to learn others’ acoustic patterns, they gradually perform sounds that are increasingly similar to the ones they hear others making (Locke, 1986).

The transition from babbling to speech is documented in relatively naturalistic observations. Once again, experimental data from non-human animals in the laboratory demonstrate the plausibility of the mechanisms presumably involved. For example, St Claire-Smith and McLaren (1983) conducted an experiment that illustrates a possible mechanism for the transition from babbling to speech through response feedback. In this study, rats produced a neutral stimulus by means of an arbitrary response during an initial training phase. Subsequently, they were more likely to perform that response if the associated stimulus was later paired with food. This finding suggests an acquired associative value by feedback itself, resulting in what might be called “backward conditioned reinforcement ⁴ ” (for a similar finding in humans, see Eder et al., 2015). da Silva and Williams (2020) claim that stimulus–stimulus pairings may be crucial for infants to enact language-appropriate vocalizations in a way that is analogous to sign-tracking (see subsection above) ⁵ . Importantly, this suggests that social reinforcement (see Goldstein & Schwade, 2008; Gros-Louis et al., 2006) is not the only factor that contributes to language development in children; yet it does suggest that reinforcement learning (Silver et al., 2021), a broader mechanism (including both social reinforcement and “self-reinforcement”) makes a major contribution to this process.

Infants with hearing loss show delayed babbling onset (Oller & Eilers, 1988) and are at risk of language learning problems in absence of early intervention (Roberts & Hampton, 2018). This happens despite having intact phonatory and articulatory control systems. Such a phenomenon illustrates the importance of hearing others or hearing oneself for language acquisition; however, it does not shed any light on whether they are separately critical. A study conducted by Sasisekaran (2012) provides insight into the role of immediate auditory feedback in the vocalization of fluent adult speakers. In this study, participants were asked to repeat non-word sequences of phonemes while receiving either synchronous or delayed auditory feedback of their own voice. The main finding was that delayed auditory feedback significantly disrupted performance, supporting the idea that acoustic production is guided by self-generated auditory feedback. However, vocalization was not completely impaired. This outcome suggests that humans rely on proprioceptive feedback for speech production, which was not controlled for in Sasisekaran’s (2012) experiment. Orofacial, vocal, mandibular, and diaphragmatic proprioception (among other sensory sources) may play an important role as feedback in human speech. Proprioceptive feedback is probably even more informative than that which proceeds from hearing because it is likely more consistent across contexts; for example, one might not hear one’s own voice in a very noisy environment but one always receives feedback from proprioceptive sensory systems. However, for practical and ethical reasons, it is generally difficult to entirely disrupt the transmission of this feedback modality in humans.

An interesting study by Méndez et al. (2010) showed that unilaterally ablating vagal afferent innervation in zebra finches severely (although transitorily) disrupted song production. The unilateral vagotomy produced modest changes in the air sac pressure at the acoustic segment of syllable production. However, some subjects presented meaningful disturbances in acoustic harmony and terminated syllables or songs prematurely. This finding suggests that feedback from variation in lung expansion, along with auditory feedback, is important for singing in birds. A similar mechanism may account for the non-total disruption of speech by altering auditory feedback in humans that was reported by Sasisekaran (2012).

8.4. Human speech: neural underpinnings

The empirical examples above provide behavioral-level evidence of speech (and birdsong) production as a closed feedback loop; namely, an interaction episode in which the behaving agent’s actions modify the (subjective) environment, which, in turn, influences its next action and so on to articulate words. On the side of neural-level models, some approaches characterize the biological substrate of human speech as a mechanism that underpins a closed feedback loop. For instance, Matchin and Hickok (2020) propose that this neural network consists of multiple layers of nested closed feedback loops. In this scheme, the lowest of the layers is the phonetic system, which is embedded in the morpheme system, which is further embedded in a lexical system, which then is embedded in a syntactic system, and which itself is embedded in a semantic system. Matchin and Hickok (2020) asserted that the architecture for implementing this function is a parcellated array of cortical regions hierarchically interconnected with one another. Despite being at different levels of the hierarchy, these systems are conceived of as computationally homologous; specifically, they function as an error correction feedback mechanism involving sensory and motor nodes (see Hickok, 2014). Like other perspectives on perceptual control (e.g., Seth, 2015; Yin, 2013), this predictive processing account assumes that feedback can derive from either external or internal (i.e., covert) sources.

According to Matchin and Hickok (2020), the arcuate fasciculus, a pathway for carrying internal feedback, is key in implementing human speech. This bundle mainly connects the pars triangularis in the inferior portion of the frontal lobe (premotor) and the posterior middle temporal gyrus (associative sensory) in the left hemisphere. Researchers have paid considerable attention to this tract in an attempt to understand the phylogenetic boundaries of human speech. For instance, a recent study by Eichert et al. (2019) compared the extension of the cortex covered by the arcuate fasciculus in macaque monkeys and humans. They found that the macaque’s arcuate fasciculus is relatively biased to the dorsal premotor regions in the frontal lobe, whereas in humans it mainly occupies the ventral regions. This suggests the major role that the human arcuate fasciculus plays as a pathway for vocal–auditory feedback.

Electrical stimulation of the ventral portion of the premotor cortex has effects in speech production (Fox et al., 2020). In contrast, electrical stimulation of the dorsal portion of the premotor cortex is associated with arm and leg movements (Rizzolatti & Fabbri-Destro, 2009). Correspondingly, the ventral premotor cortex is nearer to the laryngeal motor cortex, while the dorsal premotor cortex is nearer to the motor cortices associated with the limbs. Therefore, monkeys (and other mammals) may be equally as likely to learn limb–acoustic and vocal–acoustic associations. On the other hand, one of the human brain’s unique architectural adaptations for language may be the ease with which it couples vocal and respiratory effectors that generate acoustic feedback with the auditory system. Such a feature can be regarded as an instance of belongingness. However, it is crucially complemented by the capacity to learn flexibly and achieve higher levels of abstraction. The cortical architecture of primates—characterized by extended higher-order association regions (Cisek, 2021)—may have served as a scaffold for the capacity to flexibly generate and discern complex acoustic patterns.

People with congenital hearing impairment are usually taught to communicate by means of a sign language that mainly consists of arm and hand gestures. Since the performance of such gestures relies on a different motor system, one could expect that the brain areas underlying sign language would diverge from those that underlie vocal language. However, this does not entirely appear to be the case. Emmorey, Mehta, and Grabowsky (2007) recorded brain activity during gestural and vocal speech production using functional magnetic resonance imaging in participants with congenital hearing impairment and hearing participants, respectively. The brain areas activated during language production were strikingly similar in both groups, regardless of the response modality; those were mostly, the left inferior frontal and posterior temporal regions. Still, two regions in the dorsal left parietal cortex were more active in the hearing-impaired group during language production. These cortical areas may be involved in processing haptic and/or proprioceptive inputs from the arms. Although slight, the involvement of arm proprioceptive feedback in sign production was expected; however, the similarities between the groups merit special consideration. The inferior premotor cortex is closer to the primary motor cortex associated with the motility of the lips, tongue, larynx, jaw, and diaphragm than to that associated with arm movement, whereas the posterior temporal lobe specializes in processing complex acoustic information, especially hierarchically ordered sensory streams. Thus, the fact that these brain regions are activated when people born with hearing impairments produce sign gestures is perplexing.

Some considerations prevent the conclusion that the frontotemporal pathway is an innate device for language production based on Emmorey et al.(2007)’s findings. First, there is indeed a left-side bias in the cortex thickness of the superior temporal lobe from birth that is distinctive of humans (Leroy et al., 2015). This observation rules out the possibility that such cortical specialization is tuned either via complex social interactions or extensive vocal-acoustic feedback experience, hinting at a species-specific adaptation for prompt language learning. However, in contrast, the arcuate fasciculus is relatively immature at early stages in humans when compared to other intercortical tracts (Dubois et al., 2008). This feature is probably related to the high degree of plasticity needed for the potential to learn any spoken language at an early stage. However, since people with hearing impairment have little or no meaningful vocal–auditory feedback, this pathway may remain “vacant” until it is occupied by a more active sensory–motor process (see Flor et al., 1995). It seems that the circuitry involved in typical speech production follows distance minimization principles concerning the trajectories of the intercortical pathways recruited (Matchin & Hickok, 2020). This neural organization, in conjunction with the enhanced plasticity of said preexisting tracts, is probably a major factor enabling efficient human speech performance and learning. Namely, the coupling of the main sensory–motor components involved in different modalities of linguistic expression has topographical constraints, but also feeds on particular patterns of experience.

Ultimately, to ascertain the affinity of human brain organization and language capacity beyond the involvement of particular sensory–motor pathways for speech production, a crucial control condition is lacking. To date, there is no available information on how the linguistic function would organize the human brain if frontotemporal loops were already used to implement non-linguistic behavioral patterns. Consider a hearing person who early on learned to exclusively use their vocal–auditory system to accomplish non-linguistic goals and then learned to communicate exclusively through sign-language. Would the brain patterns generated while performing that hypothetical form of non-linguistic vocal–auditory behavior resemble those involved in speech? From a distance minimization point of view, we should expect so since the frontotemporal pathway is the shorter means for coupling vocal effectors with auditory inputs in the human cortex. Alternatively, if that pathway were exclusively reserved for language, other neural circuitries might implement non-linguistic vocal–auditory behavioral patterns.

An affirmative answer to the above question brings up a further question. If the frontotemporal pathway were already occupied by circuitries implementing non-linguistic vocal–auditory behavioral patterns, would gesturing resort to that same pathway or would it find another route for coupling the sensory–motor elements involved in sign communication? The former possibility would constitute virtually unequivocal evidence that language is hardwired in the frontotemporal pathway. The latter would indicate that language’s relationship with this pathway is mostly incidental and rather has to do with the main sensory–motor modalities involved in the typical way in which hearing people first learn language. Although conducting a study to answer such questions goes beyond the realm of ethical science, we should entertain its logical possibility and consider our knowledge gaps before jumping to premature conclusions.

8.5. More on the need for nuanced interpretation of inconclusive data

For mammals, both gustatory and viscerosceptive modalities of stimuli are processed within the insular cortex and separated by a relatively small stretch of tissue (Chen et al. 2021; Gehrlach et al. 2019; Stephani et al. 2011). It should not be surprising that in the study by Garcia and Koelling (1966) aversive conditioning was promptly learned with a gustatory–viscerosceptive set of stimuli (taking into account Hebb’s now well-supported foresight). However, even if intuitively compelling, García and Koelling’s hallmark study had some design shortcomings that preclude establishing their findings as evidence of an innate bias in associative learning (Domjan, 2015).

Through a painstaking series of studies, Domjan and his collaborators tested the boundaries of García and Koelling’s effect, challenging the conclusions that other authors simply took for granted. First, Miller and Domjan (1981a) successfully replicated the García–Koelling effect using a one-trial conditioning protocol. This served to rule out a plausible non-associative alternative explanation, which they themselves (Miller & Domjan, 1981b) had observed in a previous study—namely, selective sensitization. This was taken as evidence that the García–Koelling effect was a genuinely associative phenomenon; however, it remained possible that it was not an innate tendency, given that their experiments were performed using adult rats. Numerous experiences with, for example, gustatory stimuli and visceral outcomes along the lifespan might predispose this particular coupling of stimulus modalities to be associated. To address this issue, Gemberling and Domjan (1982) aimed to reproduce the García–Koelling effect by adapting its rationale to 1-day-old rat pups. These authors successfully replicated the García–Koelling effect, finally providing convincing evidence that postnatal experience is not required for this phenomenon to occur. Quests for knowledge of this sort are laudable and should be taken as an example of how to further our understanding of any putative innate behavior bias or learning constraint. However, jumping to conclusions that tend toward genetic explanations sometimes seems to be the rule rather than the exception.

For example, Lemaire et al. (2021) found that 3-day-old chicks readily imprinted to a predisposed stimulus (a blue geometric figure). However, in contrast, the chicks showed idiosyncratic yet consistent tendencies (either approach or withdrawal) to stimuli with low improntability. Given the early stage at which these chicks unfolded such individual differences, the possible role for previous learning was obviated or trivialized. The authors concluded that differences might have been determined by either genetic differences or stochastic factors. Of course, the explanations offered by the authors for this idiosyncratic predisposition are plausible. However, imprinting (the process and the outcome, not the potential to be imprinted) can be considered a type of learning (Horn, 2004; Mackintosh, 2015), and it could occur at stages before 3 days in chicks (Jaynes, 1956). Moreover, as other instances of learning, imprinting could be generalized to situations that are different from those in which it appeared the first time (Jaynes, 1956). Therefore, it remains quite plausible that some experiences within the very first days of the chicks’ lives in Lemaire et al.‘s study were relevant for influencing subsequent approach or avoidance tendencies.

Another example comes from a well-designed study by Seitz et al. (2020) with humans. These authors asked college students to imagine situations involving a child and then performed recall tests for details about the items in that scenario. They found that varying the biological relatedness with the imaginary child (own vs. adopted) modulated recall, with situations involving high relatedness resulting in the ability to boost memory. The results of this study are compelling and intriguing; however, the discussion statements did not contemplate all the relevant mechanisms that could have yielded that tendency. Seitz el al. (2020) discussed their results in terms of the adaptive value of recalling information in scenarios involving biological offspring in evolutionary terms, which is not in dispute. Although they made no commitment to any explicit mechanism for how that may have occurred (e.g., prenatal brain wiring, social learning), they were disinclined to consider environmental explanations. Instead, the authors hinted that their results relate to those of belongingness in non-human animals; they even cited the study of Gemberling and Domjan (1982) with neonate rats. This seems to imply that they favored a genetic mechanism over a social-learning one, given that no mention of culture was made. Culture is a powerful determinant of human behavior. Even if cultural practices may come to be by a mechanism analogous to natural selection, the underpinnings of cultural transmission are different from those involved in genetic inheritance (Kronfeldner, 2021). Since it is virtually impossible to devoid cultural influence in human psychological experiments, it is important to acknowledge its pervasive impact on human group-typical behaviors (Heyes, 2020).

9. Closing remarks

Behavior can vary from one species to another, from one population to another, from one individual to another, and within the same individual given different environmental constraints. When differences are robust and reliable, scientists are challenged to further study the mechanisms that drive variation. In so doing, they would do well to acknowledge the diversity of potential environmental arrangements, neurobiological substrates, and ontogenetic pathways by which a particular behavior pattern comes into being. For some species, learning can be regarded as a major mechanism driving behavioral change and studying it from an associative perspective has provided formal rules and principles by which behavior varies as a result of fluctuations in environmental regularities. Accordingly, organisms engage with their environment in new ways, leading to similarities or differences in their phenotypes, depending on whether said regularities are shared. However, the capacity to learn is hardly ever equipotent, which complicates the matter for both theory building and attempts at intervention. Although biases in learning potential have typically been attributed to genetic, neurobiological, and evolutionary factors, there are a number of other potential sources that can influence and underlie such tendencies.

This paper aimed to offer researchers from the field of associative learning insights (whether well-established or emerging) from other disciplines on mechanisms that influence and underpin behavioral phenotypes. Such insights can help to inspire novel research questions, as well as new ways of inquiring into and making sense the enduring debate on nature versus nurture. Special emphasis has been laid on strengthening ties among fields and adopting a cautious approach to inconclusive findings. Being mindful of the variety of possible factors through which behavioral differences can arise is key to reveal our knowledge gaps. Primarily aiming to reappraise the role of environmental factors that are often overlooked, this paper supplemented, reframed, contested, but most of all added nuance to certain conventional interpretations of key findings across several domains. Some of the claims herein are novel, while others echo and buttress ideas already found in the literature with varying degrees of acceptance. Certainly, while there is room for further refinement and improvement, at minimum, these insights can promote constructive discussion. I discussed in depth phenomena that touch on the nature–nurture issue as it relates to associative learning as well as of behavioral neuroscience and comparative cognition. This kind of inquiry, however, might as well be relevant to other fields like behavioral genetics, computational psychiatry, artificial intelligence, evolutionary biology, and developmental science.

The concerns discussed herein may be considered as a part of a growing effort across fields—such as behavioral ecology, systems neuroscience, and evolutionary developmental biology—to carefully reinterpret data in consideration of factors that are subtle, complex, or difficult to control for. For example, evolutionary biologists are increasingly interested in building mechanistic theories of phenotypic plasticity. However, even when acknowledging the important role of environmental factors, researchers from these fields do not always appeal to the literature on associative learning theory. This might be because developmental dynamics determine phenotypic pathways at multiple scales, and biologists have mainly focused on mechanisms on a molecular scale. The principles of learning are a powerful tool when approaching the plasticity of behavior at the whole-individual level, one fundamental share of animal phenotypes. Here, I outlined possible avenues for more quickly bridging the schism across boundary scales. It seems that learning is among the links missing between genetics and behavioral development. Bringing these fields into dialogue with associative learning theory has the potential to yield fruitful collaboration and advancement.