The prevailing view is that the first neurons and nervous systems emerged in the common ancestor of eumetazoans (“real metazoans”) during the Ediacaran period around 600 million years ago (Peterson and Butterfield, 2005). Animals before Eumetazoa are believed to be devoid of nervous systems, akin to extant sea sponges (Bucher and Anderson, 2015). In these early Eumetazoans, nervous systems were not organized into any recognizable “brain,” but rather a diffuse “nerve net” (Holland, 2003; Lowe et al., 2003; Galliot and Quiquand, 2011; Arendt et al., 2015). Evidence for this comes from the nerve nets of cnidarians, which are some of the earliest diverging eumetazoans.

Despite a lack of a brain, nonbilaterian eumetazoans such as cnidarians have a surprisingly rich set of neural features, strongly suggesting that early eumetazoans already contained many of the building blocks of brains. For example, neurons in cnidarians communicate using “all-or-nothing” action potentials and form chemical synapses with each other (Satterlie, 2015). Cnidarian neurons contain many modern-day neurotransmitters, including glutamate, GABA, and acetylcholine (Kass-Simon and Pierobon, 2007; Marlow et al., 2009; Delgado et al., 2010; Pierobon, 2012). Cnidarians have also been shown to contain many of the same ionotropic and metabotropic receptors for glutamate (both AMPA and NMDA), GABA, acetylcholine, and even some monoamines (Anctil, 2009; Collin et al., 2013; Bosch et al., 2017).

Neural circuits within cnidarians, and hence also likely early eumetazoans, also came with several well-known features of neurons across taxa such as adaptation and interoception. Adaptation, whereby animals respond less to repeated stimuli, has been shown in various responses such as the tentacular responses in hydras and sea anemones (Parker, 1916; Pantin and Pantin, 1943; Batham and Pantin, 1950). Interoception, whereby animals modulate their responses based on internal cues for various need states, has also been shown in cnidaria. For example, hydras and sea anemones seem to be less sensitive to food cues (less likely to trigger the feeding response and less responsive to mechanical stimulation) when full than when hungry (Parker, 1916; Pantin, 1935; Batham and Pantin, 1950; Lenhoff and Loomis, 1963; Han et al., 2018).

In contrast, even very early diverging bilaterians, such as C. elegans and planarians, have global neural integration centers that can be thought of as a “brain” (Garrity et al., 2010). Although there is still controversy regarding whether the first bilaterians had a neural net or an actual brain (Hejnol and Martindale, 2008; Arendt et al., 2015), both interpretations agree that the first brain emerged within the bilaterian lineage.

Before the first vertebrates, brains are believed to have been made up of a homolog of a hypothalamus, midbrain, and hindbrain, as in early diverging chordates such as amphioxus (Nieuwenhuys, 1998; Gorbman et al., 1999; Uchida et al., 2003; Murakami et al., 2005). There is a general consensus that with vertebrates, at least five additional structures emerged: the pallium, the BG, the tectum, the cerebellum, and the thalamus (Sugahara et al., 2017). Evidence mostly comes from the fact that these structures, their microcircuits, and the canonical connectivity between them are highly shared amongst extant vertebrates. Even the lamprey, considered one of the earliest diverging vertebrates and considered a model organism for the early vertebrates, shares exactly this template (Grillner and Robertson, 2016). Evidence suggests the amygdala, hippocampus, and neocortex all evolved from the pallium: specific subregions of the pallium in early diverging non-mammal vertebrates such as fish seems to be a proto-hippocampus (Macphail, 1982; López et al., 2003), and the other subregions of the pallium a proto-amygdala or proto-olfactory cortex, performing similar functions and containing similar microcircuits.

Some protostomes, such as arthropods, have brain regions that are structurally and functionally similar to the vertebrate pallium and basal ganglia, such as the mushroom body (similar association properties as the pallium) and central complex (similar properties as the basal ganglia). Some have argued that these structures share homology with vertebrate structures, suggesting that precursors to them emerged early in Bilateria (Strausfeld and Hirth, 2013). However, others argue it is simply an example of convergent evolution (Northcutt, 2012; Farries, 2013). Given that these structures are entirely absent from most protostome taxa, including nematodes, flatworms, mollusks, and annelids, the principle of parsimony makes convergent evolution seem more likely. Further, modern views of the urbilaterian ancestor suggest that its brain was devoid of such structures and instead was made up of a simple sensory region and a motor region (Arendt et al., 2015). Either way, the protostome versions of these structures have many more differences from the vertebrate versions than the vertebrate versions have from each other. This implies that the modern form of these structures first emerged in early vertebrates.

There is broad consensus that the six layered neocortex first evolved in early mammals, and was derived from the three layered dorsal pallium in earlier amniotes (Tosches et al., 2018). Along with this modification, the surrounding pallial structures are also believed to have become recognizable in their mammalian forms—the medial pallium became the mammalian form of the hippocampus, and the lateral pallium became the mammalian olfactory cortex and/or the amygdala (Kaas, 2009; Tosches et al., 2018). The basal ganglia, tectum, thalamus, hypothalamus, and other midbrain and hindbrain structures are remarkably similar across non-mammal vertebrates and mammals alike, strongly suggestive that they were left relatively unchanged in early mammals.

Although there is little controversy regarding the evolution of the neocortex in early mammals, the degree to which it is “different” from the structures that emerged earlier is somewhat controversial. For example, some evidence suggests that the dorsal ventricular ridge (DVR) of birds (containing the “nidopallium” and “mesopallium”) and the neocortex of mammals both derive from the pallium of their shared amniote ancestor (Karten, 1969, 1997; Reiner et al., 2004; Dugas-Ford et al., 2012). It has been shown that the DVR and neocortex share many features, including molecular properties and the subcortical structures they interact with. Hence, this might be used to call into Question the uniqueness of the mammalian neocortex, and perhaps suggests that the amniote common ancestor had neocortex-like structures. However, this interpretation is unconvincing for two reasons.

First, the brains of birds are a poor model organism for the amniote last common ancestor. It is not even clear that the DVR itself is homologous with the neocortex and instead might share homology with the mammalian amygdaloid complex (Jarvis et al., 2005; Striedter, 2005). Further, even if the DVR does share homology with the neocortex, the microcircuitry of the DVR is very different from that of the neocortex. The neocortex is organized into six layers, while the DVR is organized into clustered nuclei (Ulinski, 1983). The ontogeny of the DVR and the mammalian neocortex is different (Jones and Levi-Montalcini, 1958; Striedter and Keefer, 2000; Dugas-Ford et al., 2012). Additionally, the pallial homologs in non-bird and non-mammal extant amniotes such as reptiles, also have unique ontogeny and microcircuitry from both birds and mammals (Goffinet et al., 1986; Cheung et al., 2007). For example, turtles have a three layered pallium, instead of the clustered nuclei of the DVR or the six layered neocortex. The turtle cortex is more like the three layered pallium of other non-amniote vertebrates, such as fish (Mueller and Wullimann, 2009), than it is to the DVR of birds or the neocortex of mammals. This is more consistent with the proposal that early vertebrates had a three layered pallium, much like that of extant reptiles and teleosts, and that the pallial homologs in birds and mammals each underwent substantial independent modification since the amniote last common ancestor.

Although there is considerable debate regarding which neocortical regions emerged first, it is generally accepted that by the time the early placental mammals emerged around 65 million years ago the following neocortical areas existed: V1, S1, A1, M1, cingulate cortex, insular cortex, orbital frontal cortex (Kaas, 2012). Even in earlier diverging mammals such as marsupials, which diverged from our lineage over 150 million years ago, there are numerous neocortical areas, such as a cingulate, V1, S1, and A1 (Karlen and Krubitzer, 2007; Wong and Kaas, 2009). This is consistent with the model herein which proposes that the very early neocortex contained both frontal regions and sensory regions. I use the label ACC to refer to the entirety of the agranular frontal cortex in early diverging mammals, inclusive of the areas sometimes called prelimbic or infralimbic cortex. Consistent with this, evidence suggests the entire prefrontal cortex of rodents is homologous to the anterior cingulate of primates (Laubach et al., 2018; van Heukelum et al., 2020).

Primate brains are bigger in size than earlier diverging mammals relative to the overall body, but for the most part, contain the same neural structures. Some relative differences within primate brains include a shrinking of the olfactory bulbs and a substantial expansion of the visual cortex, somatosensory cortex, and posterior parietal cortex. However, the four substantial differences in the brains of extant primates from earlier diverging mammals are: the addition of (1) granular prefrontal cortex (gPFC; premotor areas, dorsolateral prefrontal cortex, and frontopolar cortex; Semendeferi et al., 2001); (2) PSC (including the superior temporal sulcus and the temporoparietal junction); (3) the dorsal pulvinar (Preuss, 2006); and (4) a unique cortico-motoneuronal system, where corticospinal projections bypass older circuits and make direct connections with spinal motorneurons (Murabe et al., 2018; Lemon, 2019). Although some suggest such direct corticospinal projections also occur in rats (Elger et al., 1977; Carlin et al., 2000; Yang and Lemon, 2003; Alstermark et al., 2004; Maeda et al., 2015; Gu et al., 2017), new evidence shows that these direct projections disappear in adulthood (Murabe et al., 2018), unlike in primates (Armand et al., 1997; Eyre, 2007).

The predominant view is that these structures and modifications emerged in early primates as they do not seem to exist in non-primate mammals, while they do exist in most extant primates (Kaas, 2009). Further consistent with the proposed model, and the idea that these new structures served a single purpose, the dlPFC, temporal cortex, and parietal areas all make up their own interconnected network, through the uniquely primate dorsal pulvinar (also called medial pulvinar; Goldman-Rakic, 1988; Gutierrez et al., 2000).

There are two main differences in the connectivity of human brains relative to those of non-human primates. First, the arcuate fasciculus (AF), which is a network of cortico-cortical connections between areas in the prefrontal cortex (“broca’s area”) and areas in the posterior cortex (“Wernick’s area”), is massively expanded in humans (Aboitiz and Garciía V, 1997; Aboitiz et al., 2006, 2010; Rilling et al., 2008, 2012; Aboitiz, 2012; Catani and Bambini, 2014; Petrides, 2014; Rilling, 2014; Stout and Hecht, 2017) and contains unique connectivity (Petrides and Pandya, 1984; Catani et al., 2004; Schmahmann et al., 2007; Rilling et al., 2008; Thiebaut de Schotten et al., 2012). Second, in humans, there is a direct projection from the motor cortex to laryngeal motor neurons, which is not found in nonhuman primates (Fitch, 2018; Jarvis, 2019). Although Broca’s area and Wernicke’s area are highly implicated in unique human abilities such as language, homologous regions of both have been found in nonhuman primates (Cantalupo and Hopkins, 2001; Schenker et al., 2009; Spocter et al., 2010), and therefore do not seem to have uniquely emerged in early humans.

To clarify, the hypothesis here is that taxis navigation, implemented in neurons and muscles, first emerged in bilaterians, and not that taxis navigation in general first emerged in bilaterians. Because this article is interested in brain evolution, the functions and features of the first brains are the focus of this article. Therefore, taxis navigation that existed in non-neural substrates is not explored in detail. C. elegans and flatworms, generally considered model organisms for urbilateria, demonstrate clear taxis-based navigation (Pearl, 1903; Larsch et al., 2015). These model organisms are also able to perform cross-modal decision-making. They can integrate conflicting input from various modalities (thermosensation, mechanosensation, photosensation, and chemosensation) in order to make a single integrated navigational decision. For example, C. elegans will make different decisions about whether to cross a copper barrier (which is aversive) to get to the food on the other side depending on the strength of the food smell relative to the concentration of copper (Ishihara et al., 2002). Similarly, flatworms will navigate towards food despite an aversive light source, but as the light becomes brighter, they will not go as far towards the food. This type of integrated decision-making occurs across numerous modalities including mechanosensation and thermosensation (Inoue et al., 2015). The taxis navigation of these model organism also demonstrate persistent navigational states, which would have been useful for persisting navigational decisions even after sensory stimuli have faded. For example, the C. elegans demonstrates at least three behavioral states: roaming, dwelling, and sleep (Fujiwara et al., 2002). Roaming is categorized as primarily straight-line swimming with infrequent turns—enabling an animal to relocate. In contrast, dwelling is categorized by slow swimming and frequent turning, enabling an animal to “locally search” its general area.

On the other hand, evidence for taxis navigation in adult form non-bilaterian eumetazoans is sparse. The hunting strategy of sea anemones, believed to be a model organism for early cnidarians (Yuan et al., 2011), is to wait for food to come to them (Ruppert et al., 2004). Retraction reflexes in cnidarians don’t drive locomotion in a specific direction, and instead seem to simply globally increase arousal (Batham and Pantin, 1950). Even most medusae, a more complex adult cnidarian form that likely evolved after the cnidarian-bilaterian divergence, do not show taxis navigation towards food sources, and merely seem to orient themselves in direction of current (Fossette et al., 2015), and hunt by moving in a “levy walk” (Hays et al., 2011). There are admittedly some exceptions to this. Box jellyfish can use eye spots to avoid obstacles (Garm et al., 2007), but it is generally accepted that the eyes of box jellyfish evolved independently (Nilsson, 2013; Bosch et al., 2017). Sea anemones have been shown to move towards light sources (Parker, 1916), but this has been shown to be independent of their own visual apparatus and driven simply by nearby amoebae (Pearse, 1974; Foo et al., 2019).

While larvae of earlier diverging metazoans, such as sponges, show taxis navigation (Leys et al., 2002), their adult forms show no such behavior and the neural implementation is based on cilia and not neurons and muscles. Adult forms of ctenophores move through ciliated pumping which may be coordinated via neurons and therefore may represent an adult form metazoan with neuron-based taxis navigation. However, how well extant ctenophores represent early metazoans is unclear; much evidence favors the idea that ctenophores independently evolved many features of nervous systems (Ryan, 2014; Moroz, 2015; Moroz and Kohn, 2016; Liebeskind et al., 2017). This would suggest that ctenophore taxis navigation is not indicative of early metazoans before the cnidarian-bilaterian last common ancestor.

Most evidence also suggests that associative learning first emerged in early bilateria (Bennett, 2021; Ginsburg and Jablonka, 2021). Associative learning, including classical conditioning and instrumental conditioning, has been shown broadly across Bilateria, including even early diverging species such as aplysia (Hawkins et al., 1989), planarians (Prados et al., 2012), and C. elegans (Ardiel and Rankin, 2010). In contrast, attempts to find associative learning across cnidaria have shown primarily negative results (Rushforth, 1973; Torley, 2009). Admittedly there is a single report of associative learning in a sea anemone (Haralson, 1975)—however, this is inconsistent with most other studies. Similar to others, I conclude that cnidaria do not contain associative learning (Ginsburg and Jablonka, 2019), and at the very least, if they do, it evolved convergently.

It should be noted that there are different interpretations of associative learning. Some make a distinction between “alpha conditioning” and “beta conditioning” (Razran, 1971; Moore, 2004). “Alpha conditioning” is defined as that where a non-habituated CS elicits the same reflexive response as the US, but after pairing with the US the magnitude of the response elicited by the CS increases. In contrast, “beta conditioning” is considered “true associative learning” whereby the CS elicits no reflexive response before pairing. However, others disagree with this distinction (Hawkins and Kandel, 1984; Kandel, 2006), and instead argue that there is, in fact, no difference between alpha and beta conditioning—they argue that all wiring is pre-wired, the only difference is whether the wiring is strong enough to elicit a response before pairing. All these perspectives agree that there is a distinction between general sensitization, where the specific US globally sensitizes a suite of reflexive responses, and associative learning, whereby this sensitization is local and specific to the paired US.

In the context of map-based navigation, diverse and early diverging vertebrates including fish (Burt de Perera et al., 2016), reptiles (Wilkinson and Huber, 2012; Broglio et al., 2015), turtles (López et al., 2001), amphibians (Phillips et al., 1995), and tortoises (Wilkinson et al., 2007) all show the ability to build spatial maps of their environment and flexibly generate novel navigational routes to known places (Rodríguez et al., 2002). Model organisms for early bilaterians, such as flatworms seem to navigate only with taxis and perhaps response-based learning (Pearl, 1903; Luersen et al., 2014; Larsch et al., 2015; Gourgou et al., 2021) and show no ability to remember specific un-cued locations. Further, the neural substrates of such map-based navigation in vertebrates are uniquely vertebrate structures, such as pallial homologs of the hippocampus, suggestive of vertebrate origins. Importantly, the map-based navigational tests that early diverging vertebrates, such as fish, pass do not require planning, but they do require a spatial map that can generate novel homing vectors to well-learned locations in three-dimensional space. There is indeed evidence of sophisticated 3-dimensional navigation in arthropods, however, the degree to which they truly represent map-based representations and the degree to which these abilities are representative of early bilaterians is unclear. Some evidence suggests that arthropods fail at map-based navigation tasks (Benhamou et al., 1990; Wehner et al., 1996, 2006; Walker, 1997), while other evidence suggests that they indeed can build map-like memories (Boles and Lohmann, 2003; Menzel et al., 2005, 2011). Further, many impressive abilities of arthropods emerge from mushroom bodies (Perry et al., 2013; Cope et al., 2018), which is a cortex-like structure believed to have evolved independently (Farris, 2008).

In the context of interval timing, diverse and early diverging vertebrates such as fish (Sumbre et al., 2008), birds (Bateson and Kacelnik, 1997; Ohyama et al., 1999; Buhusi et al., 2002), non-human primates (Gribova et al., 2002), and mice (Roberts and Church, 1978; Gallistel et al., 2004; Buhusi et al., 2005) all show the ability to remember the precise timing between two cues. In contrast, invertebrates show a weak perception of time, if one at all (reviewed in Abramson and Wells, 2018). Further, the neural substrates of interval timing in vertebrates seem to be uniquely vertebrate structures (Buhusi and Meck, 2005; Yin and Meck, 2014).

In the context of omission learning, vertebrates such as dogs (Cole and Wahlsten, 1968), mice (Kamin, 1957; Avcu et al., 2014), and fish (Woodard and Bitterman, 1973; Abramson et al., 1988; Portavella, 2004; Vindas et al., 2014) have demonstrated the ability to learn from omission. By omission learning, I refer to the ability to learn an association based on a predicted event not occurring, as opposed to learning from a stimulus presentation or offset. In contrast to vertebrates, invertebrates seem to fail on such omission learning studies, and only learn from stimulus offsets (Abramson et al., 1988; Wenner and Wells, 1990; Sanderson et al., 2013; Abramson and Wells, 2018).

VTE is a behavior whereby an animal pauses at choice points and toggles its head back and forth, and “plays out” each option vicariously (reviewed in Redish, 2016). Recording studies have corroborated the hypothesis that animals are playing out these options vicariously, hippocampal place cells are shown to vicariously encode place sequences of each possible path (Johnson and Redish, 2007; Gupta et al., 2012). VTE has been shown across rodents, nonhuman primates, and humans (reviewed in Redish, 2016). Further, uniquely mammalian structures, such as the prefrontal cortex, are highly implicated in such VTE behavior. In contrast, I am not aware of any studies that have shown VTE behavior in non-mammals. Taken together, this is suggestive that VTE emerged in early mammals.

Counterfactual learning is when an animal learns from an alternative choice they could have made but did not actually make. Learning from counterfactuals has been shown in rodents (Lewis, 2014; Steiner and Redish, 2014), nonhuman primates (Abe and Lee, 2011), and humans (Zhang et al., 2015). Further, uniquely mammalian structures, such as the orbitofrontal cortex, are highly implicated in counterfactual learning (Gilovich and Medvec, 1995; Camille et al., 2004; Coricelli et al., 2005, 2007). In contrast, I am not aware of any studies that have demonstrated counterfactual learning in non-mammals. Taken together, this is suggestive that counterfactual learning emerged in early mammals.

A key test of the ability to engage in episodic memory is whether an animal can answer an unexpected Question about its own past. The ability to answer such unexpected Questions has been shown in mammals including dogs (Fugazza et al., 2020), rats (Crystal, 2013), and nonhuman primates (Menzel, 1999). The neural substrate of such episodic memory seems to be the hippocampus reactivating distributed representations across the neocortex (Eichenbaum et al., 2007). Such episodic memory has also been shown in pigeons (Zentall et al., 2008) and cephalopods (Billard, 2020). However, birds are poor model organisms for the amniote common ancestor, and cephalopods are a poor model organism for the bilaterian common ancestor; both seem to have independently evolved many unique brain structures (Roth, 2015). I am not aware of any tests of episodic memory, whereby animals answer unexpected Questions, in amniotes outside of birds and mammals. Taken together, this is suggestive, but far from conclusive, that this type of episodic memory emerged in early mammals.

One might think that vicarious trial and error is a requirement in order for “spatial maps” to be used by animals for map-based navigation, but this seems to not be the case. For example, fish do not show VTE behavior, but can remember locations in three-dimensional space (Karnik and Gerlai, 2012; Burt de Perera et al., 2016; Lucon-Xiccato and Bisazza, 2017; Wallach et al., 2018), and can generate novel paths to the goal, even if it requires losing sight of that goal and swimming further away from it at first (Gómez-Laplaza and Gerlai, 2010). This demonstrates some ability to generate spatial maps, some lightweight “working memory” (staying on task even though losing sight of a goal), and object permanence (Sovrano et al., 2018).

If much of the presumably intelligence spatial navigation tasks can be performed by animals without VTE, then what is the benefit of VTE? Some suggestive evidence can be seen in the superior performance of mammals on certain tasks requiring “hard choices.” For example, mammals tend to substantially outperform non-mammal vertebrates on detour tasks, where an animal has to make a roundabout path around a barrier in order to get to a goal (MacLean et al., 2014; Gatto et al., 2018; Macario et al., 2020). Mammals also seem to outperform non-mammals in delayed gratification tasks, whereby they have to resist choosing an immediate small reward in order to get a delayed larger reward (Stevens et al., 2010). Some evidence also suggests that early diverging vertebrates such as fish cannot learn to zero-shot update the reward value of places (Beyiuc, 1938).

The Bischof-Kohler hypothesis states that only humans can take actions to alleviate a need that they will have in the future, but do not currently feel (Bischof-Köhler, 1985). However, evidence now suggests that in addition to humans, many nonhuman primates are also capable of this ability (McKenzie et al., 2004; Mulcahy and Call, 2006; Naqshbandi and Roberts, 2006; Janmaat et al., 2014). In contrast, non-primate mammals such as rodents have been shown to be unable to anticipate future needs (Naqshbandi and Roberts, 2006). Consistent with this, the substrates of the ability to anticipate future need states and use them to change current decisions seem to be uniquely primate structures, such as the dorsolateral prefrontal cortex (McClure et al., 2004; Kim et al., 2008; Hare et al., 2009). This is suggestive that the ability to perform this task emerged in early primates.

Theory of Mind (ToM) refers to the ability to take the perspective of others and understand their intentions and knowledge. It is still controversial whether any animals outside of humans contain this ability. But the balance of evidence seems to favor the idea that many primates indeed have ToM, even if it is not as robust as that of humans. Diverse species of nonhuman primates pass “false belief tests” (Bräuer, 2014; Krupenye et al., 2016; Smith, 2016; Kano et al., 2019). Further, nonhuman primates can distinguish between accidental and intentional actions and can distinguish between someone “unwilling” to do an action and someone “unable” to do an action (Call and Tomasello, 1998; Tomasello et al., 2003; Call et al., 2004; Tomasello et al., 2005). Further, the two neural structures most implicated in the theory of mind, the superior temporal sulcus and the temporoparietal junction, seem to have uniquely emerged in early primates (Kaas, 2009). In contrast, the bulk of studies on non-primate mammals conclude that they do not have ToM (Byrne et al., 2001; Bräuer, 2014; Aldhous, 2015). There is some evidence of ToM in birds (Bugnyar et al., 2016), but as we have discussed, birds are a poor model organism for the amniote common ancestor. Taken together, this is suggestive that theory of mind, even in a primitive form, emerged in early primates.

Learning skills through observation has been demonstrated across diverse species of nonhuman primates (Tomasello et al., 1987; Meunier et al., 2007; Ferrucci et al., 2019). The neural substrate of learning through observation also seems to be uniquely primate structures, including the STS (Perrett et al., 1985; Puce and Perrett, 2003). A key feature of learning skills through observation in primates is the ability to infer the intention of movement, and not simply mirroring the movement. Consistent with this, the mirror neurons in the premotor cortex of nonhuman primates have been found to be selective for abstract goals (Rizzolatti et al., 2001; Fogassi et al., 2005). There is some evidence that fish and reptiles can learn paths through observation (Brown, 2015; Lindeyer and Reader, 2010; Wilkinson et al., 2010). However, this knowledge seems isolated to observing paths and is not readily passed down in generations (Lindeyer and Reader, 2010). Taken together, this is suggestive that learning motor skills through observation emerged in early primates.

It should be noted that there are many highly intelligent social behaviors that are observed across vertebrates well outside of the primate taxa. Empathy behaviors are seen in mice (Hofer, 1996; Bartal et al., 2011; Mogil, 2012; Rennie et al., 2013). Play has been observed in reptiles (Kramer and Burghardt, 2010), birds (Fagen, 1981), and mammals (Wojciech, 2009). Jealousy and fairness preferences have been observed in mice (Douglas, 2012). Reciprocity has been observed in fish (Brandl and Bellwood, 2015) and in mice (Viana et al., 2010). Complex understanding and implementation of social hierarchies have been observed in fish (Whoriskey, 1991; Grosenick et al., 2007; Reebs, 2010) and mice (Haller and Kruk, 2006). Kin recognition has been observed even in fish (Fricke, 1974; Spence et al., 2008; Reebs, 2010; Spence, 2011), reptiles, and non-primate mammals (Brennan and Kendrick, 2006). And movement imitation has been observed in reptiles (University of Lincoln, 2014) and rats (Seyfarth and Cheney, 2013). And even gaze following has been seen in reptiles (Simpson and O’Hara, 2018).

But despite the impressive complexity of these social behaviors, none of these in fact require the theory of mind or learning through observation, and all of them can be explained by the valence and mapping functionality of the vertebrate brain. Take empathy, for example, it merely requires cues of satisfaction or dissatisfaction of conspecifics to be “hardwired” to elicit similar valence responses in an individual the same way that predator smells are hardwired to elicit fear. Consistent with this, “pain” neurons in rats that get activated when in pain, also get activated when they see the pain in others (Netherlands Institute for Neuroscience—KNAW, 2019). Kin recognition, social hierarchies, jealousy, and play can all be explained in the same way—through cue recognition and hardwired valence. Gaze following and movement imitation can be explained as merely an impressively complex reflex.

The key features of human language that differentiates it from other forms of animal communication seems to be the ability to “name” objects and organize these names with “grammar” (Berwick and Chomsky, 2017; Terrace, 2019). Despite many attempts, non-human primates have been shown to be unable to learn such human-like language (Terrace et al., 1979; Thompson and Church, 1980). A key feature in the ontogenetic development of language learning in humans seems to be “joint attention,” whereby a child and a parent can jointly attend to the same object, enabling the process of “naming” (Carpenter and Call, 2013). Attempts to demonstrate joint attention in nonhuman primates have shown negative results (Warneken and Tomasello, 2006; Warneken et al., 2006).

Additionally, music seems to be a human universal, shown across all human cultures (Brown and Jordania, 2011). A key feature of music is the ability to engage in “beat-based timing,” whereby a human can tap to a beat. Despite this being trivial in humans, nonhuman primates cannot learn to synchronize taps with an auditory or visual metronome, even after a year of training (Zarco et al., 2009; Honing et al., 2012). Nonhuman primates also struggle with relative pitch perception more than humans do, struggling to generalize melodies to transpositions (Hulse et al., 1984; D’Amato, 1988; Ralston and Herman, 1995; Ghazanfar, 2002).

Taken together, this is suggestive that language with words and grammar, along with key features of music, emerged in early humans.

Many of the sensory neurons within early diverging bilaterians such as C. elegans can be interpreted as valence neurons. Many sensory neurons in C. elegans are directly modulated by internal states and thereby seem to directly encode valence as opposed to raw sensory information: hunger peptides modulate responses of neurons sensitive to food smells, pain, and various aversive smells (Davis et al., 2017; Lau et al., 2017; Rhoades et al., 2019). Further, temperature-sensitive neurons do not actually reliably encode temperature, but rather activate when a temperature rises above a homeostatic baseline—in other words, it is encoding the “negative valence” of temperature, not temperature per se (Luo et al., 2014). Food smells that trigger approach when hungry often have no effect when C. elegans is satiated (Davis et al., 2017). In C. elegans, some cues like carbon dioxide, which can signal both food as well as predators, shift from attractive when hungry to aversive when well fed depending on hunger state (Rengarajan et al., 2019).

It is likely that the cnidarian-bilaterian common ancestor had neurons that also contained neuropeptides that signaled internal states (Jekely, 2013), but the use of these peptides for such coordinated navigational decisions (locally search area, or relocated entirely, or rest) seems to either be unique to or have been highly elaborated in Bilateria. In Cnidaria, such neuropeptides seem to have primarily modulated independent reflexes, such as the likelihood of nematocyst release (Kass-Simon and Pierobon, 2007) or the swallowing reflex (Barron et al., 2010).

In early bilaterians, neuromodulators seem to have played a large role in navigational decisions, especially when it came to persisting navigational states. For example, in C. elegans, dopamine seems to trigger dwelling behavior whereby an animal engages in local area restricted search (Sawin et al., 2000), while octopamine seems to trigger roaming behavior whereby an animal relocates entirely (Churgin et al., 2017). Throughout Bilateria, neuromodulators have a remarkably consistent template for triggering affective states (Fellous, 1999; Lövheim, 2012).

The neural mechanisms of associative learning are remarkably conserved across bilateria. Across bilateria, animals use cAMP as well as NMDA and AMPA receptors in learning processes (Kandel, 2001, 2006; Dubnau et al., 2002; Farooqui et al., 2003; Glanzman, 2010; Hawkins and Byrne, 2015). And further, “third factor” neuromodulators such as dopamine and serotonin are involved in gating both presynaptic and postsynaptic learning processes across both vertebrates and invertebrates including diverse invertebrates such as crickets (Hammer and Menzel, 1998; Farooqui et al., 2003; Vergoz et al., 2007), fruit flies (Burke et al., 2012; Liu et al., 2012), honeybees (Hammer and Menzel, 1998; Farooqui et al., 2003; Vergoz et al., 2007), and C. elegans (Kusayama and Watanabe, 2000; Qin and Wheeler, 2007).

Further consistent with shared mechanisms of learning and affect, even early diverging invertebrates, such as C. elegans and planarians, show addiction to similar chemicals that manipulate dopamine and other monoamines as vertebrates do (Kusayama and Watanabe, 2000; Barron et al., 2009, 2010; Devineni and Heberlein, 2009; Kaun et al., 2011; Søvik and Barron, 2013).

The proposed “breakthrough” of early vertebrates was the emergence of model-free RL of the kind which included spatial maps and sensitivity to interoceptive information. Studies of the function of the structures that emerged in early vertebrates corroborate this idea. Substantial evidence suggests that homologous regions of the pallium generate a spatial map across vertebrates. The ability to navigate spatial maps is abolished when hippocampus homologs (present in the pallium) are lesioned across many early diverging vertebrates (reviewed in Murray et al., 2018), including goldfish and turtles (López et al., 2003; Durán et al., 2010; Broglio et al., 2015). Further, border cells, head direction cells, place cells, and velocity cells have been observed in the hippocampus-like structures across vertebrates, including the lateral pallium of early diverging vertebrates such as teleost fish (Vinepinsky et al., 2018, 2020), which is what you would expect if the lateral pallium were the homolog of the hippocampus and the neural substrate of the “spatial map”. Further, lesions to hippocampal homologs that impair spatial navigation also impair time perception (Meck et al., 1984, 1987, 2012; Melgire et al., 2005; Balci et al., 2009; Yin and Meck, 2014; Lucon-Xiccato and Bisazza, 2017), consistent with the idea that the pallium incorporated representations of both time and space. Time cells have also been found in the hippocampal complex (Eichenbaum, 2014), and time has been shown to be represented in the hippocampus as a map, just as space is (Oprisan et al., 2018). The cerebellum, another uniquely vertebrate structure, is also shown to be critical to absolute timing tasks (reviewed in Breska and Ivry, 2016).

Temporal difference learning signals are also observed in common vertebrate structures such as midbrain dopamine neurons, lateral habenula, and the basal ganglia across vertebrates, including early diverging vertebrates such as teleost fish (Li, 2012; Cheng et al., 2014). Further, the specific circuitry of the basal ganglia, which is conserved even in the earliest diverging vertebrates such as the lamprey, has been shown to be entirely consistent with a class of model-free learning algorithms called “actor-critic” models (Grillner and Robertson, 2016).

The specific neural substrates of avoidance and omission learning also corroborate the idea that the pallial-BG-tectal implemented a model-free RL algorithm. In model-free RL, we would expect avoidance learning to work in the following way. When an animal experiences unexpected pain, there should be a decline in dopamine (negative reward prediction), and they are driven to escape via amygdala-brainstem circuitry. Whenever pain is offset, this should lead to rebound excitation of dopamine (positive reward prediction, or “relief”). If specific actions consistently precede the offset of pain, this dopamine burst will drive plasticity in striatal circuits, which learn the contingency between a CS predictive of pain and the dopamine reward of relief. If this happens enough times, the CS will cease to be scary (pavlovian responses fade) and will simply drive the avoidance behavior that has been reinforced. Such an interpretation of avoidance learning has been suggested by others (Oleson et al., 2012).

The evidence supports exactly the above mechanism for how avoidance learning works in vertebrates. First, as expected from the above model, it has been shown that the basal amygdala to basal ganglia circuitry is required for active avoidance but not escape (which requires a basal amygdala to central amygdala circuit; Bandura and Rosenthal, 1966; LeDoux et al., 2016). Second, it is shown that during initial learning dopamine declines in the striatum, but after avoidance is well learned, dopamine increases, and the increase in dopamine is predictive of the performance of avoidance (Oleson et al., 2012). Crucially, this dopamine does not increase in the striatum if the pain is always inescapable. It was even shown that stimulating dopamine neurons during aversive cue increases active avoidance performance (Wenzel et al., 2018) while inhibiting dopamine in the striatum prevents avoidance (Wenzel et al., 2018). Third, it has been shown that pain or other noxious stimuli offset drives dopamine bursts in the striatum (Navratilova et al., 2015). Fourth, it explains why pavlovian fear responses fade while avoidance can maintain itself (Annau and Kamin, 1961; Kamin et al., 1963; Blanchard and Blanchard, 1969; Starr and Mineka, 1977; Kapp et al., 1979; Mineka, 1979; Bolles and Fanselow, 1980). And fifth, it is shown that avoidance responses seem to show signs of being habitual (extensive avoidance training makes the amygdala not required for avoidance anymore (Lázaro-Muñoz et al., 2010), although still being necessary for the acquisition (Choi et al., 2010).

This interpretation of the pallial-BG-tectal system is consistent with others who similarly suggest these structures together implement a model-free learning algorithm (Joel et al., 2002; Stephenson-Jones et al., 2013; Grillner and Robertson, 2016).

The specific proposed model of ACC-sensory function presented in this article is consistent with various observations. There is emerging consensus that the neocortex, especially the sensory cortex, implements some form of a generative model (reviewed in Kersten et al., 2004; Knill and Pouget, 2004; Parr and Friston, 2018). Consistent with such a generative model, imagining a stimulus activates the same exact representations in the sensory neocortex as the stimulus itself (O’Craven and Kanwisher, 2000; Doll et al., 2015; Pearson et al., 2015). Further, lesions to specific areas of the sensory cortex create impairments both in perception and imagination within the same modality (Bisiach and Luzzatti, 1978; Farah et al., 1992).

There is less consensus regarding the function of the prefrontal cortex, which is further confused by the fact that the nomenclature of prefrontal regions in rodents often confuses the homology of these regions between rodents and primates. Recent research suggests that PL, IL, cg1, and cg2 regions of the rat prefrontal cortex are all homologous with the anterior cingulate cortex and medial cingulate cortex of primates (Laubach et al., 2018; van Heukelum et al., 2020). Through this lens, emerging evidence is consistent with the proposal that the frontal cortex builds a model of intent, which is used to trigger internal simulations.

For example, lesion studies provide evidence that the prefrontal cortex triggers internally invoked simulations in rodents: frontal lesions create impairments in all three proposed forms of internally invoked simulations, including vicarious trial and error, episodic memory, and counterfactual learning. More specifically, rodents with lesioned or inactivated prefrontal areas reduce their vicarious trial and error behavior (Schmidt et al., 2019), no longer have goal representations in the hippocampus (Ito et al., 2015), and are impaired in episodic memory tasks (Frankland et al., 2004). Evidence also suggests rats with frontal lesions become impaired at causal reasoning tasks, consistent with an inability to engage in counterfactual learning (Jones et al., 2012). Frontal lesions also make rodents uniquely impaired at spatial navigation tasks that require preplanning (Granon and Poucet, 1995). Rodents with frontal lesions struggle to stay on task in an ongoing plan and often do actions out of sequence (Seamans et al., 1995). Lesions of some of these cingulate regions in rats impair their ability to incorporate “effort” into their decision-making, as if they are unable to actually imagine how hard taking an action would be (Walton et al., 2003; Schweimer et al., 2005; Hu et al., 2021). Disconnecting areas of the cingulate from the amygdala in rats has the same effect (Floresco and Ghods-Sharifi, 2006). Further, temporarily inhibiting the hippocampus during tests where mice are asked to recall episodic memories abolishes their ability to do so (Crystal, 2013), consistent with the model whereby the frontal cortex makes inquiries for an episodic memory through its projections to the hippocampus.

Recording studies within the prefrontal cortex of rats also show evidence of modeling “intent”. In complex tasks sequences, ensembles of neurons in the prefrontal areas of rats become highly selective for the specific places in the task sequence and reliably track progress towards imagined goals (Cowen and McNaughton, 2007; Fujisawa et al., 2008). Further, in working memory tasks when rats must do tasks from memory without the presence of any cues to follow, neurons in the prefrontal cortex show delay activity (Baeg et al., 2003). The observation that the prefrontal cortex of rodents, such as the ACC, becomes particularly activated by surprise (Bryden et al., 2011) and error (Totah et al., 2009) is consistent with the model whereby during hard choices, the ACC pauses behavior and triggers internally invoked simulations to play out possible futures and resolve any conflicts or errors.

Even primates, who have many more prefrontal regions, still have these more ancient agranular frontal regions such as the ACC, which is homologous to most of the prefrontal cortex of non-primate mammals (Laubach et al., 2018; van Heukelum et al., 2020) and share many of these functions. The ACC in humans also gets uniquely activated during “errors” (Dehaene et al., 1994) and conflict (Carter et al., 1998; Braver et al., 2001). The ACC in humans also seems to encode locations in task sequences (Koechlin et al., 2002). Single neuron recordings of ACC within nonhuman primates show single neuron level selectivity for different actions (Nakamura et al., 2005), as well as selectivity for the serial order of tasks irrelevant of the actual movements made (Procyk et al., 2000; Procyk and Joseph, 2001). In humans, hippocampal lesions also create severe impairments both in the ability to recall autobiographical events and imagining potential future events (Addis et al., 2007; Hassabis et al., 2007). ACC lesions in nonhuman primates also impair the ability to stay on task during delay periods (Rudebeck et al., 2014).

The supposed “default mode network,” consisting of the mPFC, ACC, hippocampus, and the posterior cingulate all become uniquely active both during the retrieval of autobiographical memories as well as during imagining potential futures (Hassabis and Maguire, 2009; Martin et al., 2011; Andrews-Hanna et al., 2014b). Although some areas of the default mode network in primates include uniquely primate areas (such as granular prefrontal cortex), the DMN in primates also includes more ancient areas such as the ACC, which is also part of the purported DMN in rodents (Lu et al., 2012; Stafford et al., 2014; Grandjean et al., 2020), suggestive that such a network was present even in the very early neocortex.

It is interesting to note how many behavioral abilities that are attributed to the neocortex are readily performed by many animals that do not have a neocortex. Take object recognition for example. Complex object recognition, even of human faces, has been shown across phyla, including fish (Newport et al., 2016; Schumacher et al., 2016). Object recognition that is insensitive to changes in rotation and transformation has also been shown across non-mammalian phyla, also including fish (Newport et al., 2018). Object identification despite occlusion, demonstrating inference, has also been shown across non-mammalian phyla, including fish (Sovrano and Bisazza, 2007). However, consistent with the proposed function of the neocortex, mammals seem to be unique in their use of “mental rotation” to solve object recognition tasks. Mental rotation has been suggested by studies in monkeys, humans, and sea lions (Mauck and Dehnhardt, 1997), and negative results have been found in pigeons (Hollard and Delius, 1982). This would again represent the usage of such “simulation” functionality of the neocortex.

If simulating is so adaptive, then why did it only evolve in mammals and not in non-mammal vertebrates? One possible explanation is the fact that early mammals were likely arboreal species (Fröbisch and Reisz, 2009). Navigating tree branches with far eyesight presents unique challenges and evolutionary pressures not previously experienced: namely, irreversible choices. As a small animal living in trees, they had to plan their route well in advance when navigating across branches. And it is likely they had to very regularly experience novel branches. This perhaps created pressure for “vicarious trial and error”. Consistent with this, computational models have found that the usefulness of “planning” is directly tied to visual range. Visual range in water is so poor that computational models suggest planning in water is barely useful at all (Mugan and MacIver, 2020). However, inconsistent with this proposal is the fact that there are plenty of invertebrate and reptilian arboreal animals without the neocortex. Another hypothesis, as suggested by Mugan and MacIver (2020), is that many intellectual abilities occurred only in mammals due to the unique evolution of endothermy, which enabled much faster neural processing than when ectothermic. The neocortical generative model perhaps came at a high computational cost, and as such, without endothermy, nonmammalian vertebrates were prevented from garnering this adaptation. Consistent with this, the non-mammal vertebrates that demonstrate the most impressive intellectual abilities, birds, also seem to have evolved both endothermy (Walter and Seebacher, 2009) and neocortical-like structures (Tosches, 2021) through their own convergent evolution.

When identifying the neural substrates of mentalizing, it is important to draw a distinction between emotional contagion (e.g., “emotional empathy”) and mentalizing (e.g., “theory of mind”; Shamay-Tsoory, 2011). Each is a different process, and evidence suggests they have separate neural substrates. Emotional contagion is the process of reflexively adopting the emotional state of a conspecific based on various cues that reveal their emotions. In contrast, mentalizing is the process of actively considering another’s perspective, knowledge, or emotions by imagining oneself in another’s shoes. Emotional contagion has been shown in non-primate mammals such as rodents (Langford, 2006) whereas, as discussed above, mentalizing seems to be unique to primates. Consistent with this, the key frontal substrate for emotional contagion seems to be the ACC, which is homologous to regions of the frontal cortex in early mammals. For example, certain areas of the ACC are activated both by one’s own experience of pain and by watching a conspecific experience pain (Derbyshire, 2000; Jackson et al., 2006). In contrast, the neocortical areas most commonly implicated in mentalizing and theory of mind are not the ACC but instead contained within the gPFC-PSC network. More specifically, metanalysis that has reviewed a multitude of imaging studies have primarily implicated the dmPFC (BA 8, 9), amPFC (BA 10), TPJ, and STS as areas that are uniquely activated by tasks that require the theory of mind (Carrington and Bailey, 2009; Van Overwalle and Baetens, 2009). The idea that emotional contagion evolved first, subserved by older frontal regions, and ToM then emerged later, subserved by newer primate frontal regions, has similarly been proposed by other (de Waal, 2008). Further consistent with this, damage to granular areas of mPFC leaves the ability to simulate past or future imagined scenes intact but impairs the ability to imagine yourself in that scene. In contrast, hippocampal lesions impair the ability to simulate complex scenes but leaves the ability to imagine yourself in those scenes intact (Kurczek et al., 2015).

Several studies have shown a linear relationship between the volume of orbital PFC, medial PFC, and social network size (Powell et al., 2010; Lewis et al., 2011; Powell et al., 2012). It has also been observed that gray matter specifically in mPFC increases when macaques move from smaller to larger social groups (Sallet et al., 2011). Further, one of the only frontal regions that is disproportionately larger in humans than other primates is BA 10 (Semendeferi et al., 2001). All of this is consistent with the concept that these areas of gPFC subserve mentalizing, and that mentalizing was a key feature in supporting larger social groups in early primates and even larger social groups in early humans.

The above evidence implicates only a relatively small portion of the granular prefrontal cortex (BA 8, 9, 10) is related to the process of mentalizing. This suggests that the general function of the granular prefrontal cortex is not mentalizing, but rather mentalizing is merely a function subserved by specific areas of the granular prefrontal cortex. However, a broader examination of the different types of tasks involved in mentalizing implicates a much broader suite of areas within gPFC. For example, within “ToM” it has been argued that there is a dissociation between “affective ToM” and “cognitive ToM” (Shamay-Tsoory and Aharon-Peretz, 2007). Affective ToM refers to the ability to take the emotional perspective of another, whereas cognitive ToM refers to the ability to take the physical sensory perspective of another. When examining the neural substrates of these different types of ToM, broader regions of gPFC get implicated in addition to just medial BA 8, 9, and 10. Specifically, evidence suggests that the vmPFC (BA 11, 12, 14) is specifically necessary for the affective aspects of ToM whereas dlPFC (BA 44, 45, 46) is specifically necessary for the cognitive aspects of ToM. vmPFC lesions selectively impair affective ToM and not cognitive ToM (Shamay-Tsoory and Aharon-Peretz, 2007)—although admittedly given that this was examined in humans, some of these lesions may have included areas of standard mentalizing mPFC (BA 8, 9, 10) as well as areas of the ACC (BA 24, 32). In contrast, dlPFC disruption seems to selectively impair cognitive ToM (visual perspective taking; Conson et al., 2015; Qureshi et al., 2020) but not affective ToM (Kalbe et al., 2010). Disrupting the right TPJ function also disrupts visual perspective taking, while disrupting dmPFC does not (Martin et al., 2020). Some of these dissociations may explain why there is evidence that some individuals with extensive lesions of ACC and mPFC can retain features of self-awareness (Philippi et al., 2012) and theory of mind (Bird et al., 2004).

Another dissociation may be in what specific type of inquiry an individual is making into another’s mind. fMRI evidence suggests that dmPFC and vlPFC get more activated when considering how to make an observed character feel better than when merely trying to identify the observed character’s emotion (Reniers et al., 2013). On the other hand, the same study found that the dlPFC gets more activated when a participant is asked to “imagine how they themselves would feel” in an observed character’s situation than when merely trying to identify the observed character’s emotion.

There may be even more nuanced dissociations between the roles of various areas of the granular prefrontal cortex and various features of mentalizing. When comparing activation during first-person perspective taking to third-person perspective taking, areas of vlPFC (BA 44) and premotor cortex (BA 6) seem to be uniquely implicated in third-person perspective taking (David et al., 2006). Further, fMRI evidence suggests that amPFC (BA 10) builds models of person-specific theory of mind, while dmPFC (BA 8 and 9) is used for general processes of the theory of mind (Welborn and Lieberman, 2015). Other fMRI evidence suggests that amPFC is activated in any self-referential thinking, and only selectively during the considering of one’s own mental state (or presumably someone elses’), does dmPFC also become activated (Andrews-Hanna et al., 2010, 2014a). Left dlPFC may be specific for considering the mental state of others, while right dlPFC may be specific for considering the mental state of yourself (Otsuka et al., 2011). Yet another dissociation that has been proposed is that the TPJ is responsible for making mental inferences about others (such as about goals or beliefs), while the mPFC is responsible for attributing overall personality traits about yourself and others (Van Overwalle and Baetens, 2009).

Consistent with the idea that mentalizing about yourself is repurposed for the task of mentalizing about others, these same areas of the granular prefrontal cortex also get activated when considering your own mind state, including dmPFC (Gusnard et al., 2001; Gallagher and Frith, 2003; Amodio and Frith, 2006; Gilbert et al., 2006; Ochsner, 2008; Van Overwalle, 2009; Bzdok et al., 2012; Moran et al., 2012), amPFC/vmPFC (Lane et al., 1997; Phan et al., 2004; Lotze et al., 2007), dlPFC (Otsuka et al., 2011), and TPJ (Kelly et al., 2014). The general idea that there are common substrates, especially in mPFC, for imagining yourself in a future or past situation as well as for imagining being in the mind of another has been reviewed elsewhere (Buckner and Carroll, 2007; Jenkins and Mitchell, 2011). When evaluating your own personality traits or receiving evaluations of yourself by others, the same mentalizing network in mPFC activates (Ochsner et al., 2005; Gilbert et al., 2006). Further consistent with this idea that theory of mind of others is “bootstrapped” on a generative model of yourself, is the fact that the concept of “self” emerges first in child development before the theory of mind emerges (Rochat, 1998; Ritblatt, 2000; Keenan et al., 2005).

The argument in this article is that the overall function of mentalizing was not only applied to understanding the cognitive and emotional states of yourself and others (“theory of mind”), but also for anticipating future needs. The idea being that mentalizing about yourself in the future is the fundamental mechanism by which you can anticipate what your intentions and desires will be in that future situation, and thereby generate behaviors necessary for fulfilling those needs. A multitude of findings are consistent with the idea that the areas of the gPFC-PSC network specific for the cognitive theory of mind, such as the dlPFC, seem to also be key substrates for anticipating these future need states. Damage to the dorsolateral prefrontal cortex (dlPFC) dramatically impairs performance on tasks requiring abstract goal representations such A-not-B tasks (Diamond and Goldman-Rakic, 1989), consistent with the idea that dlPFC is the neural substrate of such goal representations and that it cascades these goals down the motor hierarchy where representations are progressively more “habitual”. dlPFC uniquely activates when considering future rewards but not present rewards (McClure, 2004; Tanaka et al., 2004; Berns et al., 2007; McClure et al., 2007; Kim et al., 2008). dlPFC is selectively activated when choosing a delayed reward over an immediate reward (McClure et al., 2004, 2007; Weber and Huettel, 2008). It is activated when avoiding temptation such as when dieters exert self-control (Hare et al., 2009). Damage to dlPFC leads to impairment in the ability to give up immediate rewards for future ones (Figner et al., 2010). And temporary inactivation of the dlPFC impairs people’s ability to forgo an exciting high-risk high reward option for a lower-reward lower-risk option (Knoch et al., 2006). Evidence also suggests that some types of “rule-based” categorization are unique to primates—consistent with the idea that gPFC enables the presentation of abstract goals and intentions (Soto and Wasserman, 2014).

This article also argues that the function of mentalizing in the gPFC-PSC network was also applied to enable learning through observation. There are several pieces of evidence for this. The same neurons in the premotor cortex activate for the “intentions” within yourself as when observing those same intentions in others (Amodio and Frith, 2006)—these have been called “mirror neurons.” Further, such mirror neurons have been identified in nonhuman primates (Gallese et al., 1996; Rizzolatti et al., 1996; Fogassi et al., 2005) while I am not aware of reports of mirror neurons in non-primate mammals. Although there is still controversy surrounding whether these mirror neurons are in fact modeling others’ movements (Hickok, 2009; Churchland, 2011).

Taken together, the above evidence suggests that a key, if not the primary, function of the gPFC-PSC network was to enable mentalizing—defined as building a generative model of one’s own mind state, inclusive of intentions, emotions, and knowledge. And this function was thereby applied in numerous adaptive ways, such as modeling the mind of others (theory of mind), anticipating future needs, and learning through observation.

There are compelling parallels between the connectivity of the AF-BG network and modern machine learning language models. The most popular language models, which have been shown to be capable of remarkably flexible sentence production and completion, are long-short-term memory (LSTM) models. LSTM models have “working memory” gates whereby the model can learn to maintain “context” from past words and sentences to the current word production (Mikolov et al., 2015; Dennis Singh and Lee, 2017; Ororbia et al., 2017). This is similar to the function proposed by frontal-cortex-basal-ganglia loops (O’Reilly and Frank, 2006). Further, these networks are recursive, whereby the next word is a function of the previous word, similar to the extensive recursive connectivity of both posterior and frontal cortex. The unique lateralization of language and music relative to other neocortical functions may be a requirement for the rapid sequence prediction in language and melody production.

Consistent with this interpretation of the AF-BG network, the arcuate fasciculus, and the basal ganglia are both highly implicated in language and music. Damage to AF impairs fluency, comprehension, and verbal working memory (Binder and Desai, 2011; Schomers et al., 2017). The development of the AF is correlated to the development of language abilities in children (Friederici, 2011; Yeatman et al., 2011; Skeide et al., 2016; Goucha et al., 2017; Schomers et al., 2017). The strength of connections in AF is correlated with word learning performance (Lopez-Barroso et al., 2013). It has long been known that language related tasks such as sentence generation activate and require a functioning Broca’s area, Wernick’s area, and the basal ganglia (Brown et al., 2006). However, less appreciated is the fact that music tasks such as melody generation activate the exact same network (Koelsch et al., 2002; Brown et al., 2006), although with a bias towards the right hemisphere, while language is biased towards the left hemisphere (Sammler et al., 2015). Consistent with a common neural implementation, children with language impairments also show musical impairments (Jentschke et al., 2008).

The direct projection from the motor cortex to laryngeal motoneurons, which is not found in nonhuman primates, is clearly the projection by which humans can control speech sounds (Simonyan and Horwitz, 2011). However, this was likely an addition after language had already emerged (Hewes et al., 1973; Corballis, 2002)—and not the fundamental modification that enabled language. Evidence for this can be seen in the fact that complex gestural languages (with words and grammar) are seen in deaf communities all over the world, none of which requires these descending laryngeal projections. Further, the neural substrates of such gestural language are highly overlapping with those of vocal-verbal language, implying a shared circuitry for “language,” independent of the modality (Neville et al., 1997). As such, I hypothesize that the big breakthrough that unlocked music and language was in fact the AF-BG network, and direct control of laryngeal motor neurons came later, enabling verbal language to replace the shorter-range gestural language.