Bioelectric networks: the cognitive glue enabling evolutionary scaling from physiology to mind

Not only do normal tadpoles (starting from a standard, correct state) become frogs by rearranging the components of their face, but so do “Picasso tadpoles” in which all the organ positions have been scrambled. This is because the organs then move in novel paths and different distances, as needed, to achieve a normal frog craniofacial morphology, thus showing the capacity (C’) to move to the right region of morphospace from diverse starting positions (S1–S4) and despite various obstacles (local minima LM, in which less competent navigational agents would get stuck). Panels A,B,C’ by Jeremy Guay of Peregrine Creative; C courtesy of Douglas Blackiston and Erin Switzer, and taken with permission from (Vandenberg et al. 2011)Fig. 8Multiscale control of morphogenetic competency. A Early embryos of many species, including humans, when split, result not in half-bodies but in monozygotic twins, illustrating the capacity for cells to recognize departures from the normal path and adjust accordingly. B Cells of the kidney tubule in newts not only make up for artificially enlarged sizes (using fewer cells to make the same size of tubule and newt) but in fact can harness diverse low-level molecular mechanisms to achieve their target morphology: when cells get truly huge, instead of cell:cell cooperative mechanisms, the system uses cytoskeletal bending to allow cells to wrap around themselves to achieve the same anatomical outcome. This illustrates top-down control across scales of organization, as also exploited in animal behavior. C Another aspect of multi-scale architecture is the ability of cells to make collective decisions.

In the control of the melanocyte → melanoma transition in tadpoles (driven by a disruption of bioelectrical cues), stochastic behavior can be observed: some percentage of the animals in a given cohort stay normal (left side) and some become hyperpigmented (right side). However, this decision is always made at the whole animal level: each animal is either entirely normal or entirely hyperpigmented—regardless of the stochasticity, the bioelectric circuit enables the cells to in effect toss the same coin and make a system-level coordinated decision. The same is seen (C’) when left–right patterning cues are disrupted in early chick development. Here is shown the expression of the gene Sonic hedgehog, which should normally be expressed only on the left side of Hensen’s node (red arrow), not the right (white arrow). This can be randomized by various treatments to be right-sided (middle panel) or bilateral (right panel), but each side of the Hensen’s Node decides as a coherent system—speckling (decisions on individual cell level) is never seen. Each developmental domain makes a stochastic L vs. R decision as a unit, coordinating among all the cells as a tissue-level outcome. D In planaria, each fragment can regenerate an entire worm (a kind of holographic pattern memory). Critically, however, the anatomical decisions cannot be made locally according to any simple gradient scheme. When bisected, the anterior-facing and posterior-facing cells have radically different anatomical fates (building a head vs. a tail), but were direct neighbors before the cut (i.e., at the same position along the axis). What to do after injury cannot be decided by purely local cues, but must be derived by communicating with the rest of the fragment to determine where in morphospace they are located, and thus what must be done to reach the correct target morphology.

Panel A photo by Oudeschool via Wikimedia Commons; other images used with permission from (Blackiston et al. 2011; Levin et al. 2019) and Jeremy Guay of Peregrine Creative

The most basic component of proto-cognitive capacities is that of memory. Morphogenetic memory is ubiquitous in, for example, regeneration. A salamander whose limbs, eyes, tail, or other organs are amputated will regrow exactly the right structure and stop all of the complex cell proliferation and remodeling activity when, and only when, a correct structure is complete (Harris 2018; Pezzulo and Levin 2016). While no individual cell knows what a finger is or how many a salamander is supposed to have, the tissue collective clearly has a setpoint for anatomical homeostasis that functionally guides morphogenetic behavior (memory) as an error minimization (i.e., objective function) process. This pattern memory can, like any good memory, be modified by experience (it is stable, but also labile to the right kinds of stimuli). For example, repeated amputation of axolotl limbs leads to the tissue habituating to the loss of limb and eventually giving up trying to re-grow (Bryant et al. 2017). Trophic memory in deer (Bubenik and Pavlansky 1965; Lobo et al. 2014) enables an ectopic branch in their otherwise stereotypical antler structure to be formed year after year, at a specific point of earlier damage, long after the damaged antler rack has been shed (a remarkable example of memory of position in 3D space, used to guide the growth of bone and nerve toward a new pattern, by cells at the scalp that have to regenerate the large structure every year).

The ability of morphogenetic machinery to modify its outcome due to experiences can be externally controlled (Fig. 9), analogous to the inception of false memories in neuroscience contexts (Liu et al. 2014; Ramirez et al. 2013; Vetere et al. 2019). For example, the number of heads in a regenerating planarian flatworm fragment is set by the state of a bioelectric circuit in the tissue (Beane et al. 2011). It has a built-in default of 1 (much like default, innate behaviors of brainy creatures), but is re-writable. Transient modification of the bioelectric state is remembered by the circuit, which can be experimentally re-set to read “2”: worms regenerating from such fragments indeed have two functional heads, one at each end (Durant et al. 2019; Oviedo et al. 2010). Remarkably, when cut again in multiple subsequent rounds of regeneration (with no further manipulation), the fragments continue to result in two-headed animals. This shift in the target morphology for regeneration is permanent (unless the animal’s bioelectric pattern memory is reset back to a one-headed state by an experimental manipulation). Thus, much like in nervous systems, the architecture underlying memory in this space consists of electrophysiological hardware with a highly reliable default behavior (guiding to the one-head region of morphospace) but also the capacity to be re-written toward other states, which ultimately feed into changes in gene expression and long-term changes in cell properties (Bischof et al. 2020; Pai et al. 2016; Pietak et al. 2019). The standing pattern of resting potential differences which instructively determines the number of heads that will be built is quite literally the memory of the collective intelligence of the body—a bioelectric state that guides its behavior in anatomical morphospace and can be modified by experience in physiological space.

This system matches all the key criteria for memory: it is long-term stable, yet labile (rewritable), enables discrete behaviors induced by stimuli, and can even be latent until needed (see below).Fig. 9Morphogenetic memory can be re-written non-genetically. A Normal planaria (left column) exhibit anterior gene expression in the head, and when cut into three fragments, produce normal one-headed worms. In contrast (A’), planaria in which the normal bioelectric pattern has been re-written by brief pharmacological targeting of ion channels (blue panels, green indicates depolarized regions on the voltage map) give rise to two-headed animals. Note that the bioelectrical map shown is a map of the animal pre-cutting. B Thus, a normal planarian body can store one of 2 (at least) representations of what a correct planarian should look like. Much as in the nervous system, somatic bioelectricity enables changes in how systems navigate morphospace based on experience, not only genetic rewiring. C Bioelectric circuit dynamics enable the planarian fragments to navigate a morphospace which contains attractors for 0, 1, or 2 heads as the target state which each fragment seeks to achieve. D Ability to reliably reach the right morphogenetic state is indeed a kind of memory which is stable but re-writable. Two-headed animals continue to give rise to two-headed animals upon further rounds of amputation (deviation into an incorrect region of morphospace), without changing the genetics of the cells. The two-headed state can be reversed by the same kind of technique, targeting ion flux to re-set the target morphology representation back to a one-headed configuration. E Kind of perceptual bistability seen in ambiguous images (such as this famous “2 faces vs.

vase” example) is also seen in morphogenetic systems: so-called Cryptic Worms have a destabilized target morphology memory, producing stochastically one-head or two-head forms upon each round of cutting. WT wild type, CRPT cryptic state, DH double-head. This ethnogram, applied to morphogenetic experiments, shows the transition probabilities when cut in water (H2O) or octanol (8OH, a gap junction blocker) or SCH28080 (SCH, a proton–potassium exchanger inhibitor). Images in panels B, C are by Jeremy Guay of Peregrine Creative. Others are used with permission from (Durant et al. 2017; Levin 2021a)

While the genome encodes the hardware (the cellular affordances, such as ion channels), the actual outcome is the result of context- and experience-sensitive electrophysiological software that this hardware supports (see (Bongard and Levin 2021; Bongard and Levin 2023; Nicholson 2014) for an in-depth discussion of the machine analogy in biology). The question, of how many heads a given cellular collective will produce, is, surprisingly, not directly encoded in its genome because exactly the same, wild-type set of cells can produce 1-, 2-, or 0-headed planaria depending on their history. The same is true of a single embryonic blastoderm, which normally produces 1 “embryo”, but can produce multiple conjoined individuals if the cell:cell communication is temporarily disrupted (Lutz 1949). Of course, despite its essentially epigenetic [in the broad, original sense of the word (Ginsburg and Jablonka 2009; Jablonka 2012; Jablonka and Lamb 1995; Jablonka and Raz 2009)] nature, the bioelectrical control system works together with genetic information. Evolutionary tweaks of ion channel properties that eventually lead to genetic assimilation of bioelectrically induced phenotypes is likely an important aspect of the evolution of body plans and other morphological features, much as the Baldwin effect is thought to be a key dynamic for the role of cognition and learning in evolutionary lineages (Baldwin 1896).

Three especially interesting analogs with behavioral cognition have been shown in this system. First, the advantage of such a system is the same as for neuro-cognitive architectures: it allows behavior modes that are not specifically genetically encoded, broadening the range of adaptive maneuvers that can occur in novel circumstances. The disadvantages are similar too: much as the mechanisms of brain function can be exploited by parasites using chemical (da Silva and Langoni 2009) or linguistic hijacking (Kurbel and Kurbel 2019; Panchal and Jack 2022), bioelectric control mechanisms also subject morphogenetic plasticity to potential hijacking (Williams et al. 2020). For example, much as host behavior can be specifically altered when parasites manipulate neurobehavioral systems to their own advantage (da Silva and Langoni 2009; Vyas and Sapolsky 2010; Webster and McConkey 2010), the planarian morphogenetic control system too can be hijacked not only by bioengineers (as described above) but also by commensal bacteria which can over-ride the default system and lead to two-headed regeneration (Williams et al. 2020).

Second, two-headed patterns can exist in a one-headed animal (prior to injury). This means that, analogously to the brain’s capacity for counterfactual thinking, a single physical structure [normal planarian body, with normal gene expression profiles (Durant et al. 2017)] can encode multiple memories of what the correct morphogenetic goal state is. In the case of a two-headed pattern in a one-headed host, this means that the bioelectrical pattern is not a read-out of what the anatomy is currently, but rather a representation of what it should be, weeks later, following regeneration if the animal is injured. It is in effect then, a counterfactual memory that is tied not to current state but to a possible future. This primitive system is one way to begin to think about the origins of mental time travel that brains enable, with all of its obvious adaptive advantages. It should be emphasized that the traditional lens of viewing development as an emergent property of a dynamical system that simply ends up in specific states does not facilitate the kind of research direction that enabled this new biology to be discovered (and in general, the open-loop emergence paradigm makes it very hard to see how novel outcomes can be rationally controlled). In contrast, it is specifically the search for an encoded representation as a pattern memory that guides morphospace navigation that led to the methods to directly observe (Figs. 5A,9A) and edit this information structure leading to permanent, top-down changes of the target morphology.

Finally, planarian regeneration offers an example of bistability (Pezzulo et al. 2021)—a common feature of perceptual systems with top-down control. Exposure to ion channel drugs (a physiological experience akin to electrical signals from a retina or other sense organ, not a genetic change) can place the bioelectric circuit into a state in which it cannot decide between two outcomes and flips back and forth stochastically at each regenerative event (recall of target morphology). Fragments from such “cryptic” worms (Durant et al. 2017) will regenerate as one-headed or two-headed upon each cut, randomly. These fascinating parallels between key features of cognitive and morphogenetic systems shed light on the origins of behavior and offer a simplified context in which to probe the complex aspects of conventional cognition.

Another interesting aspect of cognition is creative problem-solving: the ability of systems to achieve goals in novel circumstances or in ways that are different from their default [this is William James’ (James 1890) definition of intelligence—the ability to achieve the same goals by different means]. This definition is suitably cybernetic in its framing (emphasizing substrate-independent function and a degree of autonomous agency, not a specific neural architecture of phylogenetic position). At the same time, it reminds the experimenter that the formalization of goals and behavior are not absolute but observer-dependent: claims of goal-directedness (Clawson and Levin 2022; Levin 2023) must include a specification of a proposed problem space, a proposed ensemble of states that are hypothesized to be the goal states, and testable claims of how much and what kind of competencies to reach the goal can be expected when that system is stressed away from its normal course of events (Fields and Levin 2022; Levin 2022).

One example of this plasticity of morphogenetic behavior (expanding on the default genetically encoded repertoire) is the fact that temporary exposure to a blocker of electrical synapses (i.e., an anesthetic) causes planarian fragments to create the heads of other species, with no genetic change required (Emmons-Bell et al. 2015; Sullivan et al. 2016). The bioelectric network incorrectly navigates to additional attractors in morphogenetic space that are normally used by species 100–150 million years of evolutionary distance—all without any genetic change needed (much like nervous systems allow an animal to dynamically remember or envision novel, distant scenarios without the need for genetic change to its brain hardware to enable each new thought). Much as with human patients exiting general anesthesia, who often hallucinate for a time while the brain network is finding its way back to the correct pre-anesthetic state, but then (usually) recover their correct personal identities, planaria that build the wrong species’ head shapes eventually remodel back to normal.

The robust regulative properties of bodies strongly emphasize the system’s ability to solve novel problems. For example, early mammalian embryos cut in half do not form two half-bodies (as any hardwired, purely emergent system would). Instead, each side recognizes the damage, makes up for it exactly, and creates one of a pair of monozygotic twins. Perhaps even more remarkable is the case of newt kidney tubules (Fankhauser 1945). By default, they consist of 8–10 cells in cross section. However, if the cells of the early embryo are artificially made to be larger, fewer cells will be used, resulting in the same (normal) tubule diameter and overall body size. Remarkably, this can be pushed to a fascinating extreme: if the cells are made to be enormous, a single cell will bend around itself, producing the normal size tubule diameter. This example illustrates not only the ability to reach the same anatomical state despite diverse and novel starting conditions with no need for periods of lengthy adaptation, but also the startling ability to call up diverse molecular mechanisms (cell:cell communication in normal conditions, but cytoskeletal bending in the case of huge cells) as needed in the service of a large-scale anatomical goal. This is an example of top-down control, in which lower level mechanisms are activated based on high-level needs—an essential feature of nervous system architectures which enables the same phenomenon in behavioral space (e.g., executive-level decisions and high-order goal states that filter down to control of actin dynamics in muscle cells as implementation machinery).

Another example of cellular plasticity beyond the default behaviors of the genetically specified hardware is that of Xenobots—the proto-organisms that spontaneously form from dissociated frog embryo skin cells (Blackiston et al. 2021; Kriegman et al. 2020, 2021a). These living forms are motile (via cilia normally used to distribute mucus along the frog’s skin) and self-directed, performing a variety of spontaneous behaviors. The most remarkable novel behavior is that of kinematic self-replication: Xenobots build copies of themselves by rearranging loose cells provided to them in the medium. This is Von Neumann-style replication that is not, to our knowledge, used by any other species on Earth. Having been deprived of the normal ability to reproduce, Xenobots arrive at a novel solution within 48 h of being created for the first time (they have no history of evolutionary selection to be a good Xenobot). These examples of real-time morphological and functional adaptation, including classic ones, such as Slijper’s Goat (Slijper 1942), which acquired, in its own brief lifetime and not millennia, the body structures needed for upright bipedal walking, reveal the prodigious capacity for morphogenesis to enact creative solutions to novel problems using the same hardware. This is another essential hallmark of cognitive systems.

Thus, coherent, effective organisms form despite not being able to count on having the right number of cells, cells of the expected size, or even the same number of chromosomes (in the case of planaria and polyploid newts). This ability to handle novelty, not only in its external environment but also in that of its component parts, is the envy of the robotics and AI communities (Aubin et al. 2022; Bongard and Levin 2021; Kriegman et al. 2021b). The on-the-fly competencies of the morphogenetic control system offers evolution the same thing that nervous systems eventually offered: the ability to not over-train on evolutionary priors and instead generate problem-solving machines. Much like with behavior, it is impossible for evolution to foresee all of the novel circumstances that organisms will be required to deal with [well beyond the handful of possible environments usually studied in phenotypic plasticity and epigenetic controls (Fraebel et al. 2020; Santos et al. 2015; West-Eberhard 2005a, b)]. The amazing inter-operability of life [e.g., chimeras, bio-tech hybrids, etc. (Clawson and Levin 2022)] is a testament to the fact that life uses a “play the hand you’re dealt” system. It produces cellular collectives which carry out, not a rote set of steps, but rather a suite of second-order functions [such as active inference (Pezzulo et al. 2015; Pezzulo et al. 2018a, b) and other computational tasks] to achieve coherent function in a wide range of changing and unpredictable scenarios, both with respect to environment and also its own parts.

One formalism for understanding these capacities, and for learning to rationally manipulate them, is that of navigation (Fields and Levin 2022). Both behavioral and morphogenetic competencies can be modeled as various policies for navigating problem spaces. For example, during metamorphosis, the tadpole of Xenopus laevis can reach the correct frog face region of morphospace not only from its default starting position (a normal tadpole configuration), but also from a wide range of scrambled configurations (Vandenberg et al. 2012)—using new paths through that morphospace that nevertheless end up in the same goal region (Friston et al. 2015; Pezzulo and Levin 2016). Neural wiring does this as well, finding new paths to achieve functional network architecture in mutant mice (Little et al. 2009). As with certain animals that have automated, built-in behavioral repertoires, there are embryos (e.g., the nematode C. elegans) whose development seems largely hardwired. However, the vast majority of model species appear to use a combination of default modes and ability to improvise.

The full range of cognitive capacities for navigation of morphospace has only begun to be investigated—it is unclear how many of the advanced concepts from cognitive science (place cells, path planning, etc.) will become relevant. Moreover, these concepts in neuroscience are themselves in flux (Keijzer and Arnellos 2017; Keijzer 2017; Levin et al. 2021; Lyon et al. 2021; Pinotsis and Miller 2022). Given that cognitive capacities are present very widely across the biosphere, it is likely that advances in developmental bioelectricity may help to scaffold our understanding of neural systems and behavior.

A number of additional competencies (beyond the ability to achieve the target morphology despite modification to starting position or internal structures) have been found. For example, when an ectopic lens (in tail tissue) is induced by bioelectric modulation, it is sufficient to misexpress the channel only in a small subset of the needed cells: once their organ-level goal is specified, these cells will secondarily recruit other (un-modified) host cells as needed to achieve the critical mass needed to produce a normal-sized lens. While the engineer re-specifies the morphospace target region for a set of cells via a simple trigger, the necessary downstream modules (size control, substructure such as lens/retina/optic nerve placement, etc.) are automatically activated and do not need to be micromanaged (Gallistel 1980; Powers 1973). This modular control architecture, implemented by bioelectric circuits that set properties of a developmental compartment, is fundamental to the action of the brain (Bizzi et al. 1995; Callebaut et al. 2005; Levin and Yuste 2022), and in both contexts provides an interesting counterbalance with mechanisms of global integration. More specifically, this kind of ability of the collective to ascertain needs and modify the behavior of the right number of components is seen across swarm intelligences beyond morphogenesis, from that of ants recruiting conspecifics to a task (Burchill et al. 2022; Collignon et al. 2014; Wilson 1980), to neurons being recruited according to cognitive load (Bryer et al. 2013; Rossi et al. 2012).

The mechanisms that enable collectives to make high-level decisions and deploy their components as needed (deform the action space of subunits toward the goals of the collective in a new problem space) are only beginning to be understood. However, those dynamics are likely related to two key properties of the gap junctions common to morphogenetic and neurocognitive systems (Peracchia 2004; Trosko 2007). First, as with ion channels, they are themselves voltage-sensitive valves, enabling feedback loops and historicity. Second, by providing direct connections between the intracellular milieus of cells, they provide an “owner wiping property” for stress signals and other molecular traces of experience: cells cannot tell whether a given memory molecule (e.g., calcium flux) belongs to itself or its neighbors. This leads to partial erasure of individual identity for cells in a network, and enables a “collective” to scale up the measurements, stored goal states, and actions toward the emergence of larger scale agents operating in new spaces. Thus, while not necessarily an essential component of cognition in every possible living system (e.g., exo-biological contexts), bioelectric networks have been exploited extensively by life on Earth to provide the integrated computations needed to scale the homeostatic competencies of the most humble self-reproducing units into agents with flexible, highly adaptive behaviors in complex problem spaces.

Bioelectricity is fundamentally a mechanism to scale computation. While bioelectric states do control cell-level properties, such as plasticity, proliferation, differentiation, etc. (Levin 2021a), the real power in this system is in determining large-scale behaviors at the tissue and organ level (Harris 2021; Levin and Martyniuk 2018). It has been shown to control size (Daane et al. 2018; Perathoner et al. 2014; Yi et al. 2021), organ identity, and whole body axes (Levin 2021a). A critical (and brain-like) aspect of bioelectrical networks is the hierarchical organization of functionality and the association of complex morphogenetic activity with simple stimuli (triggers). Much like the central nervous system (CNS) allows complicated multi-step behaviors to be triggered by a low-information content stimulus, a brief and transient bioelectrical signal can induce whole eyes (Pai et al. 2012) and appendages (Adams et al. 2007; Tseng et al. 2010) in which all the internal details are handled autonomously.

This key feature is currently beginning to be explored for regenerative medicine applications. For example, amphibian tails (including spinal cord) and limbs can be induced to regenerate by specific bioelectric states triggered by ion channel drugs or optogenetic stimulation (Adams et al. 2013; Tseng et al. 2010). Severe brain defects in the frog model, induced by chemical teratogens or mutation of a critical neurogenesis gene (Notch), have been rescued by reinforcing appropriate bioelectrical signaling (Pai et al. 2020, 2015; Pai and Levin 2022). Especially in the case of Notch mutation-induced malformations, the fact that such fundamental hardware defects can be in effect resolved “in software” (by drugs modulating Vmem to force the correct pattern) suggests the potential power of these interventions for addressing defects and traumatic injury for which we have few viable treatments today. This potential is akin to how behavioral accommodations can often make up for even severe structural defects (Slijper 1942). Similarly, tumorigenesis has been shown to be controllable by modulation of bioelectric state—normalizing cancer by reconnecting cells to the electrical network that harnesses them toward adaptive tissue homeostasis (Chernet et al. 2016, 2015; Chernet and Levin 2013b), a promising alternative to current toxic chemotherapy approaches.

Efforts to control the information processed by bioelectric networks toward particular outcomes are now being guided by computational tools at multiple levels of organization, applying Marr’s (Marr 1982; Peebles and Cooper 2015) taxonomy to control of body form just as neuroscience addresses the problems of behavior at levels from that of synaptic proteins to psychiatric therapies (Adams et al. 2016; Friston et al. 2014; Wang and Krystal 2014). However, it is likely that this work has only begun to scratch the surface of what is possible. Biomedicine today remains largely focused on the hardware—micromanaging genes and pathways toward desired functionality— and has not yet internalized the fundamental wisdom of the multi-scale approach in the neuro-behavioral sciences. The research summarized above suggests that behavior-shaping and training paradigms for cells and tissues will likely enable much greater control of morphology, gene expression, and pathway function than can be realistically achieved bottom-up (just as training animals for specific behaviors is more efficient than attempting to run their muscles directly like puppeteers). Transformative improvements in capabilities in bioengineering and regenerative medicine are within reach, if we learn to appropriate as much as possible of the multi-scale wisdom of neuroscience and generalize it beyond neurons.

The impacts of behavioral neuroscience, and of its advances in explaining cognitive and proto-cognitive capacities, lie well beyond classical neural cells. Understood broadly, developmental bioelectricity provides an entry-point into unifying adaptive “behavior” and problem-solving intelligence in diverse spaces in a way that makes it natural to think of plant, microbial, and even synthetic life (Baluška and Mancuso 2012; Baluška et al. 2022; Baluška and Reber 2021a, b; Bassel 2018; Calvo et al. 2020, 2017; Debono and Souza 2019; Martinez-Corral et al. 2019; Prindle et al. 2015; Reber and Baluška 2021; Schofield et al. 2020; Solé et al. 2016; Souza et al. 2017; Urrios et al. 2016; Yang et al. 2020) using the same conceptual tools from behavioral and physiological sciences. The on-going debate around representation and morphological computation is likewise being enriched by data in this field (Keijzer 1998, 2001).

The remarkable fundamental mechanism that enables a true emergent collective intelligence—a mind (at whatever scale of sophistication)—is ancient, and is also responsible for the plasticity and robustness of morphogenesis. Evolution re-used some of the same computational strategies, for binding competent signaling subunits into networks with memory and problem-solving capacity, to navigate a diverse set of spaces (physiological, anatomical, behavioral, and linguistic). Bioelectricity offers a tractable and powerful entry-point into understanding this process, because it serves as the cognitive medium of collective intelligence—whether of neurons in the brain, or of cells in a body trying to achieve anatomical outcomes. Thus, firm conceptual (and disciplinary) distinctions between the science of the brain and those of the body are largely artificial hold-overs from past limitations of technology and evolutionary understanding, and are increasingly being erased (Anderson et al. 2012; Beer 1995; Pezzulo and Levin 2015; Pfeifer et al. 2007; Willems and Francken 2012). Future work will address the mechanisms by which developmental bioelectricity sets up the structure of the CNS (Pai et al. 2015) and the ways in which neural signals control morphogenesis (Belgacem and Borodinsky 2015; Borodinsky et al. 2012; Herrera-Rincon and Levin 2018; Herrera-Rincon et al. 2017); beyond the molecular mechanisms linking these two systems, there is also much opportunity for unification into a single underlying conceptual architecture.

Evolution exploits the generic computational properties of such networks (learning, generalization, counterfactual memories, representation, distributed control, etc.) at many scales, building flexible problem-solving engines instead of fixed solutions to specific environments (Moczek et al. 2011; Oyama 2000; Sultan et al. 2022). While bioelectricity is an especially convenient modality, proto-cognitive capacities (such as learning) are found also in biochemical networks and biomechanical networks (Biswas et al. 2021; Stern et al. 2020a, b; Watson et al. 2010). Emerging frameworks that focus on a ubiquitous multi-scale competency architecture, where each layer has some degree of behavioral and proto-cognitive capabilities that await discovery, are already generating new capabilities and driving novel research programs (Aubin et al. 2022; Bongard et al. 2006; Davies and Levin 2023; Kriegman et al. 2021b; Levin et al. 2017; Merrild and Rasmussen 2018; Pezzulo 2020; Pezzulo et al. 2021; Pezzulo and Levin 2015, 2016, 2018; Pfeifer et al. 2005; Taylor et al. 2016). The implications of this approach (Fig. 10) range far beyond behavioral and developmental sciences, to encompass synthetic bioengineering of novel cognitive life forms (Clawson and Levin 2022; Ebrahimkhani and Levin 2021; Kamm and Bashir 2014) and related fields, such as robotics and AI. All of these disciplines are beginning to intersect in the exciting emerging field of diverse intelligence research, which will not only provide numerous useful applications in engineering and biomedicine, and help to understand our evolutionary history, but most importantly, will shed light on the deepest philosophical problems of the origin and nature of possible cognitive Selves.Fig. 10Future of the diverse intelligence field.

A Because of the deep ability of living cells to form collectives with adaptive functions under novel circumstances, many diverse combinations of evolved material (cells, tissues), designed material (engineered smart materials or implants), and software (AI systems) are viable. The future is likely to include a massive number of highly diverse agents with various degrees of behavioral competency (A’: hybrots, cyborgs, biorobots, etc.) which will be the subject of an expanded field of animal behavior research that is not limited to the set of natural animals here on Earth. B Mindmap of the emerging field at the intersection of the sciences of animal behavior and morphogenetic control, showing the many disciplines whose ideas feed into this new consilience, and the deep concepts and practical impacts that progress in this field will have. Images in panels A and A’ are by Jeremy Guay of Peregrine Creative