Alignment Science

1. The recurrence of alignment across substrates and scales

A quarter of a millennium ago, Adam Smith observed something that would change the world. He saw that to get people to contribute to outcomes distant or higher than themselves, you don’t need to make them directly care about those outcomes. Instead, the architecture of the marketplace and the broader economy can guide people, as if by an invisible hand, to participate in ends that are not their own (A. Smith, 1776). In fact, Smith argues that this works better: people who pursue their own self-interest promote the interests of society more frequently and more effectively than those who claim to serve the public good.

Economics has advanced its understanding of this phenomenon significantly in the centuries since without dislodging the central claim. The architecture of the marketplace—property rights, contracts, legal enforcement, monitoring, etc., and of course, prices—really does guide people to serve other people’s interests and participate in large-scale patterns they don’t know or care about. It doesn’t do so by getting people to care about other people’s interests; instead, it shapes what it means to pursue one’s own self-interest—to wit, it often becomes advantageous to make something that someone else would want to buy—and constrains and channels the effects of one’s behavior, like how a decision to buy more bread than usual is communicated across the system as an increase in the price of bread.

Increasingly, the same kinds of patterns have been observed in other sciences.

I first noticed this when learning about Lisa Feldman Barrett’s theory of constructed emotion (Barrett, 2017). Despite the name, this is not really a theory of emotion per se, although it grew out of it. It is a theory of how the brain regulates the body, where psychological phenomena emerge both as a consequence of bodily activity and as a dynamical regime governing that same activity (this kind of circular causality recurs across many of the phenomena we’ll discuss). In Barrett’s view, the brain does not micromanage detailed bodily processes, instead working with relatively high-dimensional information (Sennesh et al., 2022), nor does it come equipped with evolutionarily designed modules for producing psychological phenomena and the stereotypical behaviors and facial expressions associated with them. Instead, the brain’s job is to regulate the body by efficiently allocating resources—oxygen, salt, glucose, etc.—across competing demands from everything from muscles and organs to subsystems like the immune system and the endocrine system.

That sounds a lot like economics.

More examples started popping up. A revolution in motor behavior associated with Nikolai Bernstein said that behavior didn’t occur by the brain commanding the details of bodily movement (Bernstein, 1967). Bernstein observed that the brain would face a huge degrees-of-freedom problem: there are way too many muscles, fibers, joints, etc., that can exist in too many combinations, with respect to too many environmental conditions shifting in kind, degree, and relevance, to centrally command behavior. Instead, behavioral tasks and requirements have to be translated into constraints and affordances that are meaningful at the scales of the components that assemble a motor behavior; the components then can be relied upon to contribute their own competencies to the collective discovery of the solution.

Economists will note that this sounds a lot like the Hayekian knowledge problem and the associated critique of central planning relative to decentralized markets (Hayek, 1945). Later work would generalize Bernstein’s ideas to a more general science of behavior and organization (Turvey, 1990).

A few years later, I found a third example: Michael Levin’s bioelectric approach to morphogenesis (Levin, 2021a). In high school biology, we all learn that cells assemble a larger organism by following a genetically prescribed blueprint. But genes do not specify a trajectory for every perturbation the developing organism might encounter. One really incredible example of this is an experiment perturbing the locations of craniofacial organs in a developing frog (Vandenberg et al., 2012). The organs then navigated to their correct locations. This cannot plausibly depend on a separately specified correction trajectory for every possible displacement, especially when the right moves depend on the paths the other organs are taking. Instead, the organs have to somehow be guided to navigate to the right position by a live, real-time system—as if by an invisible hand.

I noticed that bioelectricity seemed to play a very similar role in the organization of multicellular organisms to the role prices play in the organization of the economy. Michael Levin and I wrote a paper about this (Levin & Lyons, 2026). Then, along with Léo Pio-Lopez, we wrote two more papers. The first extended the bioelectric-price analogy to an analogy between cancer and externality (Lyons et al., 2026b). The second examined the cross-substrate emergence of distributed control systems we called virtual governors (Lyons et al., 2026a), borrowing a term from Norbert Wiener (Wiener, 1965). This paper drew on examples from many phenomena, not just biology and economics. By observing a common abstract regulatory object across many fields, our virtual governors paper is a proper “alignment science” paper—a paper about how alignment works.

Even more examples started coming after that. Power generators, as Wiener observed, collectively construct a virtual governor. Even simple mechanical systems, like tensegrity structures, see simple rigid struts become organized through a prestress pattern that enables distant adjustment to local problems, a hallmark of what Levin calls a “cognitive glue” (Levin, 2023a).

Common problems, and common solutions, recur across substrates, from single cells to global economies. The problems are similar in kind to the problem we face with artificial intelligence: how to get many agents, or competent components more generally, to contribute to some larger pattern while staying within desired bounds? But they predate AI, and often exhibit solution types that are not currently commonplace in discussions of AI alignment.

This suggests that alignment is not merely an AI problem, or a metaphor extended from AI to other spaces. Alignment is a recurring cross-substrate phenomenon worth treating as an object of science in its own right. “Alignment science” isn’t a cute term; it’s a plausible scientific discipline.

The standard division of the sciences into physics, chemistry, biology, etc., aren’t the only way to divide up scientific problems. Different people in different circumstances focusing on different problems, or different commonalities between them, might have produced a different division. Just like you can divide up a map of the world by countries, you can also divide up that same space by climate zones, weather systems, economic regions, mountain ranges, and more, you can divide up the same collection of scientific phenomena in different ways. The point isn’t to argue about which division is right or best; the question is whether the division produces useful scientific work by organizing the problems, guiding scientists to the right perceptions, constraining experimental protocols, and defining both the central questions of the field and what meaningful answers look like.

Several other cross-disciplinary generalizations of related patterns exist in the literature and have helped to inform the perspective I offer here. There is Wiener’s cybernetics, as alluded to previously. Also important is Haken’s synergetics (Haken, 1983) and Kelso’s related analysis of the dynamics of coordination (Kelso, 1995). More recent examples are the free energy principle (Friston, 2010), and Technical Approach to Mind Everywhere (Levin, 2022) and the associated collective intelligence framework (Watson & Levin, 2023; McMillen & Levin, 2024). Also of note is The Sensory Order (Hayek, 1952/1999), which could be considered an early domain-specific case study in alignment science. My hope is that alignment science adds to these perspectives with its focus on the conditions under which the activity of some component or process contributes robustly to a distant success condition, and on how that relation is created, maintained, altered, or shut down.

So this is what this essay is here to do: propose alignment science as a useful division of scientific study. To that end, I’ll attempt to sketch out the basic elements of that division: what alignment is, why alignment is hard, what makes alignment work, and what kinds of things start to become more salient when we consider the study of alignment as a science in its own right. The point of this sketch is, emphatically, not to declare definitionally what alignment is or how it works, but to generalize across existing empirical findings. Future findings, or current ones I am personally unaware of, will update this picture.

2. Definitions of alignment

I define alignment as the organization of the degrees of freedom of one process such that its activity becomes systematically responsive or contributory to the conditions of success at another process, system, or scale.

This is very broad and abstract, a consequence of trying to define a commonality across so many diverse physical instantiations. Rather than trying to fit every instance to this definition, empirical efforts focused on a particular alignment problem may find a definition or focus more suitable to that purpose.

Broader still, but perhaps more generally applicable, is this definition: alignment is the task of making success conditions at A become causally effective over the degrees of freedom at B. What “success conditions” are, and what “causally effective” means, may also depend on the empirical details of a particular case. Note that success conditions need not be represented or located at any single point but can emerge endogenously, as in a physiological setpoint; success conditions can also be historically selected, specified by another process in the system, or imposed or rewritten by an investigator for some purpose.

More general still, but no longer pretending to any degree of scientific precision, is this: alignment is the problem of making sure that the parts serve the whole. This definition is simple enough at a glance, but elides the problem that the parts construct, transform, and sometimes even disassemble the whole. Circular causality is frequently observed in alignment science; the whole is a product of alignment processes among the parts, which in turn are shaped, constrained, and channeled by the whole. Alignment is often not a problem of an already-existing, fully-formed system, asking “How do I get my parts to do what I want?”

Definitions are of limited use and, in my opinion, function only as pointers; they have fuzzy boundaries and do not pick out natural kinds. What alignment looks like is a relation in which variation in one component, process, or plan remains compatible with conditions relevant to the successful activity of another component, process, or larger organization. The challenge of alignment is in constructing, generating, maintaining, regenerating, transforming, and, when necessary, shutting down such relations.

This brings up a distinction between vertical alignment and horizontal alignment. Vertical alignment is the relationship between a component and a larger organization, such as that of a cell and a multicellular organism. Horizontal alignment refers to component-component relationships, such as how buyers and sellers in a marketplace are coordinated, or how drivers in traffic are coordinated. In both examples of horizontal alignment, the components’ plans are almost literally lined up with each other, i.e., aligned.

3. Five distinctions about alignment

Distinctions are more important than definitions. There are five basic points that help to separate the study of alignment specifically from interacting processes in general.

The first is that alignment is relational. It is meaningless to ask whether something is aligned in isolation. Alignment is to something (e.g., to a virtual governor). The same component can be considered aligned from one perspective and misaligned from another. For example, a worker can be aligned to the firm’s goal of maximizing profit, yet if the firm is producing a negative externality, this means that the worker’s activities are misaligned with the broader economy.

The second is that alignment concerns degrees of freedom. For a component’s behavior to be meaningfully considered aligned to something, there must be some possibility of it behaving otherwise. If there were only one physically possible trajectory, there would be nothing to align.

The third is that alignment requires consequential organization. Alignment rarely comes about because components are asked to please align to a purpose. It requires that a system be structured so that one process becomes responsive to another’s success conditions.

This consequential organization can consist of many things: motor synergies, price signals, laws, roles, traffic signals, the rules of the game, and more. None are obviously right or wrong in the abstract. Different alignment problems require different organizations; different organizations produce different alignment processes and different kinds of alignment. One may be good for forming a temporary coalition among a few components; another may be well-suited to scaling alignment over billions of agents across many years.

The important thing is to recognize that two systems can interact and influence each other without either being organized around the other’s success conditions or together coupled to a success condition at a higher scale. Such organization is what turns interaction into alignment.

The fourth is that alignment is multicausal, meaning that it is constructed from multiple elements at multiple scales along multiple timescales with no privileged point or scale at which the alignment or misalignment resides (Thelen, 1995). This means that alignment cannot simply be reduced to any of the following: shared goals, obedience, cooperation, value agreement, central control, optimization, similarity, or lack of conflict, nor is it necessarily the case that alignment problems are correctly or efficiently addressed with component-level fixes.

Fifth, alignment is often partial and multidimensional. A teacher can be aligned to a child’s long-term education and misaligned with a child’s immediate fun. When alignment occurs in a multiscale competency architecture (Levin, 2023b), it varies across scales pertaining to physical organization, abstract task organization, and time. More alignment along one dimension can mean less along another. Alignment is thus rarely “solved,” in a once-and-for-all sense, and is not necessarily reducible to a binary condition that exists on a single dimension. This is why it is often more useful to ask what kind of alignment a system exhibits than whether it is simply aligned.

4. Why alignment is hard

Speaking at a high level, there are four recurring challenges for alignment.

The first is underdetermination. A high-level target does not specify the local actions that realize it. “Maximize the firm’s profits” does not tell a worker what behaviors to engage in; “build this anatomical target” does not tell a cell what role to play in the larger structure. The components have to fill in the gaps.

The issue is more than a lack of high-level specification. There are often many behaviors or trajectories consistent with a high-level target. The canonical example is Bernstein’s degrees-of-freedom problem (Bernstein, 1967). The goal of moving from point A to point B does not specify whether the moving system should walk, run, crawl, jump, skip, hop, swim, drive, fly, etc. This continues: a decision to walk does not specify how fast, along what path, how uneven terrain or an obstacle should be dealt with, or how to recover from a perturbation. Any decision to engage in a particular movement fails to specify how that movement should be assembled across muscles, fibers, joints, nerves, etc.

In neuroscience, this principle is called degeneracy (Edelman & Gally, 2001; see also Marder & Goaillard, 2006): there are many neural ensembles, pathways, etc., that might be sufficient to produce a memory, emotion, thought, or behavior. Beyond just the brain, degeneracy is a common principle in biological systems at many levels (Albantakis et al., 2024).

The second is distributed information. Relevant information for assembling a collective solution is not necessarily available to the whole collective but is dispersed across the system. This is the Hayekian knowledge problem (Hayek, 1945): no single component of the system, including at the highest scale, knows enough to solve the whole problem. This creates a problem of organizing interactions so that decentralized information enters system-level behavior.

A deeper problem is this: the relevant information may not exist prior to the interactions that create it. A market may not know the relative scarcities of apples and oranges until the interactions between buyers and sellers in the market produce the equilibrium relative prices for apples and oranges. Relative scarcity is a system-level property, yet individual buyer and seller knows only their individual plans for apples and oranges. Ultimately, there need not be a separately available system-wide measure of relative scarcity prior to the interactions through which prices emerge

Beyond economics, this problem arises in many domains. The brain does not know everything happening in the body, nor what local adjustments may be appropriate for the overall bodily situation. Genes cannot know ahead of time every perturbation and obstacle encountered on the developmental path to a mature organism.

The third is a mismatch between what is locally sensible and what is globally sensible. Alignment often works by arranging local conditions so that the components’ default behaviors collectively produce the desired global outcome. Many alignment problems can be framed in terms of a mismatch between what is locally incentivized and what is globally desirable.

In economics, externality is the general term for this mismatch; market failure (Bator, 1958) is one broad catalogue of mismatch sources. Cancer can be viewed as a local/global mismatch where a cell disconnects from the signaling network that integrates its dependencies with the larger multicellular organism (Levin, 2021b). Externality and cancer have comparable properties despite being studied in different sciences (Lyons et al., 2026a).

Local/global mismatch indicates the multicausal nature of alignment: alignment and misalignment cannot be attributed to the component in isolation. Instead, the fault may usefully be associated with the signaling system, the architecture of constraints and dependencies, or even the higher-scale goal. For example, a small cut to the wrist or neck can turn the behavior of an artery from aligned to severely misaligned in a matter of seconds. As Coase (1960) showed, the source of the local/global mismatch has no privileged position but is a policy or engineering question about where corrective work can most efficiently be done.

The fourth challenge is changing conditions. Alignment cannot be achieved once and then be considered solved. Alignment processes must respond correctly to perturbations—a market must adjust to changes in preferences; a fly needs to reach its destination even when a gust of wind may take it off-course. Turnover must be addressed: cells and people die and are born, firms go out of business and enter the market, some memories fade while new ones are created. Damage must be repaired, missing parts or organization regenerated.

A successfully aligned system may not remain aligned. Even success and progress can be threatening to alignment. An adolescent with growing limbs may find it difficult to coordinate. Infants may exhibit perseverative errors because they are gaining in competency (Thelen et al., 2001). Achieving a mature organism may cause senescence (Pio-Lopez et al., 2025).

Alignment is not just a problem of preservation and repair but also of change and disassembly. A sports game may require a rapid shift from offense to defense. Performing successfully on a reaching task may require destabilizing a previously useful memory (L. B. Smith et al., 1999). At any given time, an ostensibly aligned or aligning system may consist of many underlying processes of creation, repair, regeneration, transformation, generation, and destruction—and eliminating any of these might lead to pathologies elsewhere.

5. Alignment requires causal machinery

Simply assigning a higher-scale target to a collection of components is not itself an alignment mechanism. Giving the components seemingly desirable values or goals, or issuing them particular commands or specifications does not necessarily create alignment. Each component can want, in whatever sense, to achieve the higher-scale target and nevertheless collectively fail to assemble the solution.

In fact, real-world alignment processes often do not observe components having some kind of internal commitment to the higher-level goal. The economy relies on self-interested agents who do not know or care about the economy-level allocation of resources. There is no reason to believe that cells care about, or are even necessarily aware of, the human being they constitute. Even when components might affirm the right values, like how all members of a sports team might say that their goal is to help the team win, this does not mean they all have the same understanding of that objective, nor does that stated objective necessarily play a major role in constructing their perceptions and actions during the game. The most meaningful version of the goal might be a consequence of the in-the-moment playing of the game—something downstream rather than upstream.

Beyond a focus on the right values, alignment requires solving the problem of making success conditions at one locus causally effective over degrees of freedom elsewhere. This requires causal machinery: relationships and structures that make alignment happen.

An important alignment solution observed across substrates and scientific fields is mechanisms that transform the problem seen by the parts. A familiar example comes from economics. Rather than ask each member of the economy to be aware of and considerate of the relative scarcities of resources, the price system transforms the landscape of constraints and affordances perceived by each member. Then when each agent optimizes for their own individual goals, they collectively help to allocate resources efficiently.

An even simpler example can be seen in sports. A soccer player is rarely instructed to play in a way that maximizes the team’s odds of success. Instead, they are assigned a position and an associated role—“defend the left side.” History in the form of practice and previous games gives meaning to this role; the current evolution of the game guides perception and action in the context of this history. The soccer position transforms the problem seen by the soccer player.

This is the general pattern: rather than asking components, individually or collectively, to solve the global problem, the system instead alters what individual components sense, gain from, depend on, or can do so that the resulting local “default” behavior contributes to the larger outcome. In economics, this problem is explicitly studied as mechanism design, which asks how rules, incentives, and information structures can be arranged so that independently acting agents generate desired system-level outcomes (Hurwicz & Reiter, 2006). (A particularly interesting question is what happens when components are able to model and respond strategically to the alignment machinery.)

Other alignment mechanisms include reducing available degrees of freedom, as in locking joints to aid in motor control; removing persistently misaligned components, as when white blood cells kill cancer cells; and altering the components themselves, as when children are taught right from wrong or when workers gain competency from repeating a task. Internalized goals and values can thus be part of alignment’s causal machinery rather than alternatives to it.

The causal machinery of alignment is thus very broad and can consist of signals, feedback, constraints, incentives, couplings, dependencies, environmental structure, institutions, laws, norms, emergent and assigned roles, and physical architecture. The machinery can act on highly intelligent, capable components like humans, and it can also act on simple mechanistic components (Keim et al., 2019). Degrees of freedom, not a brain, are the basic prerequisite for alignability.

6. What does alignment science look like?

Simply noticing a common pattern across different scientific fields is not enough to make alignment science interesting. Alignment science should make new problems and possibilities visible. It should reorient scientific attention in useful ways. It should do the work that alignment mechanisms do: transform the impossible problem of understanding the universe’s phenomena into a locally manageable problem that enables the collective assembly of useful work.

At the level of generality employed in this piece, the basic function of alignment science is to enable cross-substrate and cross-mechanism comparisons, which can then enable existing problem spaces to borrow from others, or for candidate solutions to be identified for new problem spaces, or to prevent the emergence of predictable failure modes in solution design. The emergence of new general theoretical constructions, such as the virtual governor idea, or the generalization of the cognitive glue concept from biology to economics (Levin & Lyons, 2026), that are helpfully applied to various problems is a sign of the utility of alignment science.

An example of the work done by an alignment science perspective comes from the cancer-externality comparison (Lyons et al., 2026b). The theory yields a cross-substrate prediction: a misbehaving component might not be usefully modeled or corrected as a component-level flaw. Instead, restoring the component’s connections to the wider organization may correct its behavior.

Alignment science resists treating isolated component properties or a privileged microlevel decomposition as sufficient explanations. It emphasizes mechanisms insofar as they are relational, distributed, dynamical, and organized across scales. How the system decomposes its own self, and the world around it, can be more important than the decompositions the scientist finds convenient or intuitive (Bernstein, 1967; Hayek, 1952/1999).

Alignment science can change what is treated as the object of explanation. Rather than beginning by isolating components in the lab and asking which of their properties produce an outcome, it is important to recognize that relationships do a great deal of work in a system. Context is not a source of noise but can in fact play a key role in constructing the phenomenon of study (Barrett, 2022). Competencies can be made to appear and disappear with no apparent change in the parts of the system ordinarily thought responsible for the competencies (Thelen & Ulrich, 1991; Gendron et al., 2020). Observing that making some variable A present or absent in a system yields the appearance and disappearance, respectively, of a phenomenon, is not necessarily evidence that A causes the phenomenon (L. B. Smith & Thelen, 2003). Signals might not be intrinsically meaningful but instead relationally real (Hayek, 1952/1999; Barrett & Theriault, 2025). A component’s perceptual, behavioral, and cognitive competencies should be tested, not assumed (Levin, 2022). A macrostate typically is not explained by a microstate that exists as a smaller version of the macrostate (e.g., a homunculus), nor is it necessarily useful to model a macrostate as a sum of several independent macrostates, the way a puzzle is arranged by sticking puzzle pieces together (Thelen & Smith, 1996; Barrett, 2017).

Certain questions acquire central importance in alignment science. What variables predict what a component is alignable to? What kinds of alignment do different dependencies and signaling systems lead to? When alignment processes construct a higher-level entity from the components, what predicts the interests of the constructed entity? What are the scaling rules for alignment? When failure types recur across substrates, can solutions be ported across substrates as well?

The most likely limitation for alignment science is that a cross-discipline generalization does not substitute for domain-specific knowledge. Alignment science can tell us that mutual dependency is a key part of alignment’s causal machinery; it does not tell us what mutual dependencies might make sense in a particular situation or how to construct, maintain, transform, and dissolve the dependencies when necessary. Alignment science will be most useful when it remains in constant contact with domain-specific research rather than becoming a purely abstract search for universal principles.

7. Conclusion: Beyond artificial intelligence

Currently, the topic of alignment is primarily associated with artificial intelligence. While it is noteworthy that research inspired by AI alignment has produced some work congruent with alignment science (Barasz et al., 2014; Tomašev et al., 2025; Pierucci et al., 2026; Paglieri et al., 2026; Jha et al., 2026), AI is not a central example in alignment science, either as a foundational example or as an application focus.

Alignment science takes a broad perspective on alignment, drawing from examples across many disciplines. It does not justify itself by potential applications to AI alignment problems. Instead, alignment is a scientific object simply worthy of study in its own right.

Separating alignment science from applications to AI specifically may ultimately be of benefit to AI applications. Consider: a NASA mission to the Moon will consist of some engineers dedicated to that mission, and some people who just like astrophysics and are willing to work at NASA to get paid to do it. The former benefit from the latter to some degree: they notice things others overlook, ask weird questions, think about certain problems for long periods of time. Alignment science can produce positive externalities for AI alignment without even necessarily studying it directly, let alone treating as the central theoretical problem and practical engineering focus.

Instead of restricting alignment to a narrow set, we should study alignment wherever we can find it. If alignment is a real cross-substrate phenomenon, then the relevant evidence is already scattered across many disciplines. There is only one way to find out what becomes visible when these cases are studied together.

References:

Albantakis, L., Bernard, C., Brenner, N., Marder, E., & Narayanan, R. (2024). The Brain’s Best Kept Secret Is Its Degenerate Structure. Journal of Neuroscience, 44(40). https://doi.org/10.1523/JNEUROSCI.1339-24.2024

Barasz, M., Christiano, P., Fallenstein, B., Herreshoff, M., LaVictoire, P., & Yudkowsky, E. (2014, January 22). Robust Cooperation in the Prisoner’s Dilemma: Program Equilibrium via Provability Logic. arXiv.Org. https://arxiv.org/abs/1401.5577v2

Barrett, L. F. (2017). The theory of constructed emotion: An active inference account of interoception and categorization. Social Cognitive and Affective Neuroscience, 12(1), 1–23. https://doi.org/10.1093/scan/nsw154

Barrett, L. F. (2022). Context reconsidered: Complex signal ensembles, relational meaning, and population thinking in psychological science. American Psychologist, 77(8), 894–920. https://doi.org/10.1037/amp0001054

Barrett, L. F., & Theriault, J. (2025). What’s Real? A Philosophy of Science for Social Psychology. In D. T. Gilbert, S. T. Fiske, E. J. Finkel, & W. B. Mendes (Eds.), The handbook of social psychology (6th ed.). Situational Press. https://doi.org/10.70400/BPQW3358

Bator, F. M. (1958). The Anatomy of Market Failure. The Quarterly Journal of Economics, 72(3), 351–379. https://doi.org/10.2307/1882231

Bernstein, N. (1967). The Coordination And Regulation Of Movements. Pergamon. http://archive.org/details/bernsteinthecoordinationandregulationofmovements

Coase, R. H. (1960). The Problem of Social Cost. The Journal of Law & Economics, 3, 1–44. https://doi.org/10.1086/466560

Edelman, G. M., & Gally, J. A. (2001). Degeneracy and complexity in biological systems. Proceedings of the National Academy of Sciences of the United States of America, 98(24), 13763–13768. https://doi.org/10.1073/pnas.231499798

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Gendron, M., Hoemann, K., Crittenden, A. N., Mangola, S. M., Ruark, G. A., & Barrett, L. F. (2020). Emotion Perception in Hadza Hunter-Gatherers. Scientific Reports, 10(1), 3867. https://doi.org/10.1038/s41598-020-60257-2

Haken, H. (1983). Synergetics, an introduction: Non-equilibrium phase transitions and self- organization in physics, chemistry, and biology (3rd ed.). Springer.

Hayek, F. A. (1945). The Use of Knowledge in Society. The American Economic Review, 35(4), 519–530.

Hayek, F. A. (1999). The Sensory Order: An Inquiry into the Foundations of Theoretical Psychology. University of Chicago Press. https://press.uchicago.edu/ucp/books/book/chicago/S/bo8930377.html (Original work published 1952)

Hurwicz, L., & Reiter, S. (2006). Designing Economic Mechanisms. Cambridge University Press.

Jha, K., Cicala, F., Arcas, B. A. y, Richards, B. A., Jaques, N., Kleiman-Weiner, M., & Niklasson, E. (2026, September 9). Tapes Together Strong: The Co-evolution of Computation and Cooperation. arXiv.Org. https://arxiv.org/abs/2609.10817v1

Keim, N. C., Paulsen, J. D., Zeravcic, Z., Sastry, S., & Nagel, S. R. (2019). Memory formation in matter. Reviews of Modern Physics, 91(3), 035002. https://doi.org/10.1103/RevModPhys.91.035002

Kelso, J. A. S. (1995). Dynamic patterns: The self-organization of brain and behavior (pp. xvii, 334). The MIT Press.

Levin, M. (2021a). Bioelectric signaling: Reprogrammable circuits underlying embryogenesis, regeneration, and cancer. Cell, 184(8), 1971–1989. https://doi.org/10.1016/j.cell.2021.02.034

Levin, M. (2021b). Bioelectrical approaches to cancer as a problem of the scaling of the cellular self. Progress in Biophysics and Molecular Biology, 165, 102–113. https://doi.org/10.1016/j.pbiomolbio.2021.04.007

Levin, M. (2022). Technological Approach to Mind Everywhere: An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds. Frontiers in Systems Neuroscience, 16, 768201. https://doi.org/10.3389/fnsys.2022.768201

Levin, M. (2023a). Bioelectric networks: The cognitive glue enabling evolutionary scaling from physiology to mind. Animal Cognition, 26(6), 1865–1891. https://doi.org/10.1007/s10071-023-01780-3

Levin, M. (2023b). Darwin’s agential materials: Evolutionary implications of multiscale competency in developmental biology. Cellular and Molecular Life Sciences: CMLS, 80(6), 142. https://doi.org/10.1007/s00018-023-04790-z

Levin, M., & Lyons, B. (2026). Cognitive glues are shared models of relative scarcities: The economics of collective intelligence. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 384(2320), 20240528. https://doi.org/10.1098/rsta.2024.0528

Lyons, B., Pio-Lopez, L., & Levin, M. (2026a). Alignment Is to a Virtual Governor: A Theory of Coordination in Diverse Intelligence (No. 2026070220). Preprints. https://doi.org/10.20944/preprints202607.0220.v1

Lyons, B., Pio-Lopez, L., & Levin, M. (2026b). From Cancer to AI Alignment: Tackling Externalities Through Homeostatic Principles (No. 2026040056). Preprints. https://doi.org/10.20944/preprints202604.0056.v1

Marder, E., & Goaillard, J.-M. (2006). Variability, compensation and homeostasis in neuron and network function. Nature Reviews Neuroscience, 7(7), 563–574. https://doi.org/10.1038/nrn1949

McMillen, P., & Levin, M. (2024). Collective intelligence: A unifying concept for integrating biology across scales and substrates. Communications Biology, 7(1), 1–17. https://doi.org/10.1038/s42003-024-06037-4

Paglieri, D., Cross, L., Genewein, T., Leibo, J. Z., Tomasev, N., & Vezhnevets, A. S. (2026). A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms (arXiv:2609.04170). arXiv. https://doi.org/10.48550/arXiv.2609.04170

Pierucci, F., Galisai, M., Bracale, M. S., Prandi, M., Bisconti, P., Giarrusso, F., Sorokoletova, O., Suriani, V., & Nardi, D. (2026). Institutional AI: A Governance Framework for Distributional AGI Safety (arXiv:2601.10599). arXiv. https://doi.org/10.48550/arXiv.2601.10599

Pio-Lopez, L., Hartl, B., & Levin, M. (2025). Aging as a Loss of Goal-Directedness: An Evolutionary Simulation and Analysis Unifying Regeneration with Anatomical Rejuvenation. Advanced Science (Weinheim, Baden-Wurttemberg, Germany), 12(46), e09872. https://doi.org/10.1002/advs.202509872

Sennesh, E., Theriault, J., Brooks, D., van de Meent, J.-W., Barrett, L. F., & Quigley, K. S. (2022). Interoception as modeling, allostasis as control. Biological Psychology, 167, 108242. https://doi.org/10.1016/j.biopsycho.2021.108242

Smith, A. (1776). An Inquiry into the Nature and Causes of the Wealth of Nations: II (1st ed.). W. Strahan & T. Cadell.

Smith, L. B., & Thelen, E. (2003). Development as a dynamic system. Trends in Cognitive Sciences, 7(8), 343–348. https://doi.org/10.1016/S1364-6613(03)00156-6

Smith, L. B., Thelen, E., Titzer, R., & McLin, D. (1999). Knowing in the context of acting: The task dynamics of the A-not-B error. Psychological Review, 106(2), 235–260. https://doi.org/10.1037/0033-295X.106.2.235

Thelen, E. (1995). Motor development: A new synthesis. American Psychologist, 50(2), 79–95. https://doi.org/10.1037/0003-066X.50.2.79

Thelen, E., Schöner, G., Scheier, C., & Smith, L. B. (2001). The dynamics of embodiment: A field theory of infant perseverative reaching. The Behavioral and Brain Sciences, 24(1), 1–34; discussion 34-86. https://doi.org/10.1017/s0140525x01003910

Thelen, E., & Smith, L. B. (1996). A Dynamic Systems Approach to the Development of Cognition and Action (Reprint edition). Bradford Books.

Thelen, E., & Ulrich, B. D. (1991). Hidden skills: A dynamic systems analysis of treadmill stepping during the first year. Monographs of the Society for Research in Child Development, 56(1), 1–98. https://doi.org/10.2307/1166099

Tomašev, N., Franklin, M., Jacobs, J., Krier, S., & Osindero, S. (2025). Distributional AGI Safety (arXiv:2512.16856). arXiv. https://doi.org/10.48550/arXiv.2512.16856

Turvey, M. T. (1990). Coordination. American Psychologist, 45(8), 938–953. https://doi.org/10.1037/0003-066X.45.8.938

Vandenberg, L. N., Adams, D. S., & Levin, M. (2012). Normalized shape and location of perturbed craniofacial structures in the Xenopus tadpole reveal an innate ability to achieve correct morphology. Developmental Dynamics, 241(5), 863–878. https://doi.org/10.1002/dvdy.23770

Watson, R., & Levin, M. (2023). The collective intelligence of evolution and development. Collective Intelligence, 2(2), 26339137231168355. https://doi.org/10.1177/26339137231168355

Wiener, N. (1965). Cybernetics, Second Edition: Or the Control and Communication in the Animal and the Machine (2nd edition). Mit Pr.