Machine Metaphors in Biology III: A Dialectical Synthesis

Introduction

The first part of this trilogy began with a claim that has guided everything that followed: metaphors are not ornaments, but cognitive instruments. They organize attention, import expectations from one domain into another, and silently influence what scientists consider measurable, manipulable, and ultimately explainable. This is why machine metaphors became so important in biology: the clock transformed order into mechanism, the engine transformed metabolism into energetic throughput, the computer transformed heredity, control, and cognition into problems of information processing. Each metaphor made previously opaque biological phenomena scientifically tractable, and the history reconstructed in Part I was therefore not a history of scientific error, but of extraordinary epistemic success.

Part II started from the paradox generated by that success. Every time the metaphor encountered a limitation, the machine changed. The rigidity of the clock was answered by the thermodynamic engine; the engine’s inability to explain regulation was answered by computational architectures. The real problem therefore appeared only when this evolving analogy ceased to be recognized as an analogy at all, when the claim that organisms can be modeled computationally was silently transformed into the stronger claim that organisms are fundamentally computers, and when the same movement was eventually extended to minds, societies, ecosystems, and finally the universe itself. The target of Part II was therefore not computation per se but computational imperialism—the inflation of a productive explanatory strategy into an exhaustive ontology. Yet the antithesis created its own problem.

Rejecting pancomputationalism does not automatically give us a better biology. Saying that organisms are “wholes,” “processes,” “subjects,” “autonomous systems,” or “emergent organizations” may correctly diagnose what classical mechanicism leaves out, but it does not necessarily tell an experimental biologist what to measure tomorrow morning. There is a recurring temptation in anti-reductionist thought to replace a technically limited framework with a conceptually rich but operationally poor vocabulary. “Everything is connected” may be true, but by itself it explains almost nothing. “The whole determines the parts” may contain an important insight, but until we specify what relation of determination is involved, at which scale, over which time interval, and under which interventions, the claim remains closer to an intuition than to an experimentally discriminable hypothesis.

Part II therefore ended with a challenge to organicism itself: a theoretical biology that criticizes mechanical philosophy but cannot be operationalized risks becoming what I would call descriptive vitalism. This is where the synthesis must begin. So what is next? Is there an organicist alternative that goes beyond our current machine metaphors in biology while remaining operationalizable enough to explain real biological phenomena and useful enough for experimental settings? Can we preserve everything that mechanistic biology gave us (decomposition, intervention, causal reconstruction, quantitative modeling, prediction where prediction is possible) without assuming that the boundaries, components, constraints, and variables entering those explanations are fixed once and for all? I think we can, but only if we stop treating the organism–machine relation as a one-way street.

For most of the history reconstructed in this trilogy, machines functioned as the source domain and organisms as the target. But analogical modeling need not proceed in only this direction. Barwich and Rodriguez have recently emphasized precisely this reciprocity: biological discoveries can change our understanding of what machines are just as machine concepts change our understanding of organisms. The truth is that much criticism of the organism-as-machine metaphor attacks a machine ideal inherited from the eighteenth, nineteenth, or twentieth century even though artificial systems have themselves become stochastic, adaptive, distributed, reconfigurable, evolutionary, embodied, and increasingly hybridized. The proposal is then not to diminish organisms until they resemble old machines, but to expand our concept of machines in light of what modern systems can become.

This reversal changes the question. The interesting problem is no longer merely whether organisms are machines, but what kind of machine would an organism have to be? That question will guide the last installment of this trilogy. The answer I develop is not a new metaphor that should simply replace the computer. Nor is it an invitation to stretch the word machine until it means any organized physical system. Instead, I will argue that biological explanation requires a transformed conception of mechanism in which relevant constraints are not merely given from outside but can be produced, maintained, and eventually modified by the system itself. This requires distinguishing mechanicism from mechanistic explanation, recovering the organizational insights of organicism without its explanatory vagueness, and developing theoretical languages capable of representing relations of production, constraint, constitution, maintenance, interpretation, and anticipation.

Organizational approaches provide an unusually promising starting point. Their central principle, organizational closure, has reappeared in sufficiently different traditions that its recurrence is itself theoretically informative: living systems do not merely undergo transformations, they participate in producing some of the constraints under which those transformations remain possible. However, before expanding our current organizational narratives, there are many other matters we should clear up. What exactly is a metaphor? What is a machine? What do contemporary biologists mean by a mechanism? How does a time-invariant relational organization coexist with the historical becoming of an organism? And if a future machine could genuinely produce some of its own semantic primitives, would the hardware–software distinction on which modern computation rests remain intelligible? These questions determine whether the synthesis will become a genuine research program.

Metaphor and Machine

Throughout this trilogy I have used the word metaphor in a broad sense. Part I described metaphor as a controlled displacement of structure from one domain into another. When DNA is described as code, the cell as a factory, or the heart as a pump, the comparison does not merely replace one word with another. It transfers an architecture of relations. Codes have alphabets, syntactic rules for encoding and decoding, errors and interpreters. Factories have inputs, outputs, specialized machinery, assembly processes, control structures, and supply chains. Pumps possess operating regimes, failure modes, and well-defined mechanisms that allow them to function. In this way, biological concepts become tractable because the source domain brings with it a partially organized way of looking at the target.

This general understanding is consistent with a large literature on scientific metaphors. For example, Kampourakis emphasizes that metaphors have representational and heuristic roles precisely because they allow familiar conceptual structures to participate in the investigation of less familiar phenomena. Yet the transfer is necessarily selective: some similarities become foregrounded while others recede. Importantly, this is not a defect unique to scientific metaphors. Scientific abstraction itself works by selective omission too. A model or simulation that retained every feature of its target would no longer simplify anything. The difficulty arises because two very different types of failure can occur, and until now, I haven’t always clearly distinguished between them in the first two installments of the present trilogy.

The first problem concerns the incompleteness of a modeling relation. Robert Rosen gives metaphor a technical meaning in Life Itself. Roughly speaking, a metaphor can function as an incomplete modeling relation in which a formal structure is decoded toward a natural system without the empirically grounded encoding relation that would complete the modeling diagram. In that technical sense, a metaphor is not simply figurative language; it is a candidate formal representation that has not yet been justified through the full cycle relating natural entailment to formal entailment. Notably, this is not the same thing as the map–territory confusion I emphasized in Part II. A missing encoding is a feature of the modeling relation. Reification occurs when someone mistakes a representation for the thing represented. In other words, one can retain a metaphor responsibly while knowing it is incomplete; conversely, one can reify even an empirically successful model by treating its categories as the final furniture of reality.

There is, however, a second complication. Rosen’s definition of metaphor per se does not automatically condemn machine metaphors. Indeed, his own position is subtler. Metaphors are useful precisely because we may pursue them in the hope that they can eventually be transformed into genuine models. Moreover, by Rosen’s technical criterion, even a relational representation of an organism remains metaphorical insofar as it retains only selected relational properties of the natural system. Rosen’s deeper objection to machines appears elsewhere. It concerns what happens when one attempts “to complete” the modeling relation and discovers that the entailment structure characteristic of a living organization cannot simply be identified with the restricted entailment structure characteristic of what he narrowly defined as machines and mechanisms.

This correction changes the constructive question I am trying to develop here in Part III. Perhaps “what is the next metaphor?” is not quite the question we should ask. If metaphors are valuable partly because they can be developed into models, then a more productive question is: what kind of formal system can be completed into an empirically adequate model of a closed living organization? Basically, instead of searching endlessly for new source domains, we should ask which formal structures can represent functional entailment and eventually be tied back to experimental observations. This does not make metaphors obsolete, quite the opposite, it gives metaphors a more disciplined role in scientific endeavor.

In his How Life Works, Philip Ball calls for new, life-centered metaphors because post-genomic biology has repeatedly revealed the limitations of the genome-as-program and organism-as-machine pictures. Yet merely replacing machine metaphors with jazz ensembles, conversations, or whirlpools would not solve the deeper theoretical problem. Each new source domain would illuminate some relations and conceal others. The goal is not to discover the perfect metaphor since there is no perfect metaphor. The goal is to develop a science sufficiently reflexive that its metaphors can be transformed into models, its models can be confronted with experiments, and the discrepancies revealed by those experiments can force the metaphors themselves to change. That recursive relationship between metaphor, model, experiment, and conceptual revision may turn out to be more important than whichever machine metaphor happens to dominate next.

The problem becomes harder when realizing that machines are themselves a moving target. One of the central historical lessons of Part I was that the meaning of machine did not remain fixed. A seventeenth-century automaton, a nineteenth-century steam engine, a twentieth-century digital computer do not share one obvious internal architecture. What they share depends on the abstraction we choose, and this is why many classical objections to machine metaphors have an expiration date of their own. “Organisms learn but machines do not”; machine learning makes the claim historically obsolete. “Organisms exhibit decentralized behavior whereas machines are centrally controlled”; distributed computation complicates that distinction. “Organisms repair themselves whereas machines do not”; self-healing materials weaken the argument again. “Organisms are composed of living matter whereas machines are not”; programmable biobots immediately make the substrate boundary less comfortable.

Bongard and Levin therefore argue that much of the classic organism–machine debate compares living systems with machines as they were, rather than machines as they could be, recognizing that properties historically associated with machines are technologically contingent rather than conceptually essential. This is a powerful counterargument to simplistic organicism, but it does not follow that every future distinction will disappear. Nicholson’s critique remains important because he identifies a difference deeper than rigidity, determinism, centralized control, or material composition. He distinguishes the intrinsic purposiveness characteristic of organisms from the extrinsic purposiveness of conventional artifacts, distinguishing the heuristic usefulness of machine metaphors from the much stronger theoretical claim that organisms literally belong to the same organizational class as machines.

The tension between Bongard and Levin on one side and Nicholson on the other is very productive. If “machine” means a rigid, deterministic artifact assembled once from externally manufactured parts, then organisms are obviously not machines—but neither are many contemporary machines! If “machine” simply means any physical organization capable of producing some outcome, then organisms trivially are machines—but so are hurricanes, rivers, ecosystems, and perhaps almost anything else. The concept loses discriminatory power! The interesting territory lies between these extremes. A useful machine concept must remain restrictive enough that calling something a machine explains more than saying that it exists and changes, while remaining historically flexible enough to incorporate genuinely new modes of engineering.

Barwich and Rodriguez’s reciprocal account provides a clue. Machine metaphors do not require the source domain to remain unchanged. Biological discoveries can reorganize engineering concepts, and technological innovations can alter which biological similarities become salient. We therefore do not need to decide, once and forever, whether organisms “really are” machines. We can instead ask which organizational properties currently distinguish them and whether future engineering can reproduce those properties materially rather than merely imitate their outputs. That is a much harder question, and it is also truly experimentally useful. The difference between organism and machine would then not be a metaphysical wall but a research frontier.

“Mechanism” is neither mechanicism nor reductionism

Clarifying what is a machine forces us to clarify another word that has created considerable confusion throughout this debate: mechanism. As Nicholson well pointed out, at least three different ideas have historically traveled under that name. Mechanicism is an ontological thesis about what organisms fundamentally are. In its strongest classical form, it claims that biological systems are machines and that biological phenomena can ultimately be reduced to the interactions of machine-like components. As we explored in Part I, following the Newtonian revolution, the concept of a mechanism took on a much more concrete meaning. Nicholson termed it “machine mechanism,” and it refers to an actual device, such as gears, valves, pumps, levers, circuits, governors, or some analogous arrangement of parts that produces an operation in a machine-like structure.

Contemporary mechanistic explanation, however, is something much broader. In the new-mechanist philosophy of science, a mechanism is an explanatory organization of entities, activities, relations, and conditions responsible for a phenomenon. In Nicholson’s words, causal mechanism is “a step-by-step explanation of the mode of operation of a causal process that gives rise to a phenomenon of interest.” Nothing in this definition requires those entities to behave like gears. Cellular signaling, neural oscillation, gene regulation, immune activation, and developmental processes can all receive mechanistic explanations even when they are nonlinear, stochastic, distributed, context-dependent, and multiscale. This distinction matters because saying that “mechanism fails” can mean several entirely different things.

It is crucial that biologists keep looking for causal mechanisms. If disrupting a receptor abolishes a signaling response, we want to know why. If changing an enzyme alters a metabolic flux, we want to understand the causal organization responsible. If a tissue regenerates after injury, we want to identify the molecular, electrical, mechanical, and cellular dependencies that enable it. A theoretical framework that responds to such questions by saying “the organism is an irreducible whole” has not transcended mechanism. It has abandoned explanation. The synthesis I want to defend here is more conservative and, for that reason, stronger: mechanistic explanation remains indispensable, but it becomes incomplete when it treats as fixed the very conditions whose endogenous production is part of the phenomenon we are trying to explain.

The synthesis advocated here then goes beyond the usual criticisms of machine metaphors and mechanistic explanation. Under the modern conception mechanism, the relation between causation, mechanism, and explanation is no longer conceived as a one-way ladder. A causal vocabulary constrains what a mechanistic decomposition can represent. Mechanistic decomposition gives causal claims experimental content. But the explanandum itself determines which decomposition is relevant. A receptor can be the component of one mechanism, the product of another, the constraint on a third, and part of the material background of a fourth. There is no context-free “true decomposition” waiting underneath every biological system. Significantly, this perspectivism does not imply arbitrariness.

Some decompositions survive experimental intervention better than others. Some generate predictions. Some remain robust under alternative measurements and coarse-grainings. Some identify relations that disappear when the system is perturbed. William Wimsatt’s multi-perspectival realism is useful precisely because robustness across partially independent investigative routes offers an alternative to both a God’s-eye ontology and unconstrained conventionalism. The same organism can legitimately be decomposed differently for different explananda without every decomposition becoming equally good. Thus, what we need is a closure-sensitive mechanistic explanation, where local mechanisms remain and what changes is the explanatory boundary.

To see why this matters, it is useful to return to one of the strongest expressions of twentieth-century molecular mechanicism—Jacques Monod’s Chance and Necessity. Monod deserves more careful treatment than the simple caricature of a genetic reductionist. He identifies three characteristic properties: teleonomy, autonomous morphogenesis, and reproductive invariance. Teleonomy designates the apparent project-like organization of living systems. Autonomous morphogenesis refers to their ability to generate macroscopic structure through internal interactions rather than through external shaping. Reproductive invariance refers to the preservation and transmission of highly ordered organization across generations. Monod’s philosophy emphasizes that, among these properties, reproductive invariance allows autonomous morphogenesis to become mechanistically explicable, while teleonomy is ultimately subordinated to hereditary stability and natural selection.

There is an important philosophical reason for this ordering. Monod wants to protect what he calls the postulate of objectivity. Science should not explain biological purpose by assuming purpose as a fundamental principle of nature. If wings exist for flying, a scientific theory should not simply insert “flying” into the causal origin of the wing. Darwinian evolution offers a way out. Structures that look designed can emerge through variation and selection without foresight. Teleonomy is therefore real as an organizational property, but its historical origin must be explained through processes that do not themselves presuppose the future end. Any contemporary theory of agency that ignores this lesson risks simply renaming the explanandum. Saying that a bacterium “wants” nutrients does not explain chemotaxis. Calling a tissue “goal-directed” does not tell us how regeneration works. Invoking “intrinsic purpose” without specifying the organization through which some states matter to the system merely pushes the mystery into another word.

In this way, Monod was right to demand naturalization. Where I disagree is with the stronger architecture he builds around that demand. For Monod, invariance must precede teleonomy: selection can progressively refine teleonomic structures only after a system capable of preserving variation already exists. But this is not the only possibility. In fact, the opposition between “invariance-first” and “teleonomy-first” may itself be an artifact of the architecture through which the problem has been formulated. Consider what invariant reproduction actually requires in a contemporary cell. DNA does not preserve itself, it does not copy itself by virtue of sequence alone, and it does not interpret itself. In biological organization, the description helps produce the interpreter, and the interpreter makes the description functional.

Such circularity is not incidental and can be formalized in terms of relational diagrams. From this perspective, neither invariance nor teleonomy needs to appear fully formed before the other. Primitive hereditary differences could stabilize organizations exhibiting weak forms of viability-biased persistence. Those organizations could, in turn, improve the conditions under which hereditary differences were preserved. Improved inheritance could stabilize more elaborate interpretive constraints. Those constraints could generate more reliable self-maintenance. This loop could progressively tighten and the relevant origin-of-life story therefore becomes one of co-emergence. The mistake is then to imagine invariance and teleonomy as two independent modules waiting to be placed in temporal order. They may instead be two aspects of a progressively closing organization.

This changes how we interpret Monod’s “principle of objectivity” as well. Objectivity remains indispensable as an epistemic norm if it means that our explanations should remain answerable to evidence rather than preference. But objectivity should not be confused with perspective-free access to nature. Scientists choose variables, instruments, interventions, scales, boundaries, and explananda. We never encounter “the system as it is in itself” before deciding which relations are relevant. Wimsatt’s epistemology is useful here because it allows us to be realist without pretending to be omniscient: if a relational property survives multiple partially independent decompositions, interventions, and modes of observation, our confidence that we are tracking something objective increases. Objectivity is therefore not a view from nowhere. It is an achievement produced by triangulation.

This distinction matters because biological organization is deeply scale- and context-dependent. The same molecular object may appear as fuel in one explanation, constraint in another, catalyst in another, signal in another, and product in still another. Insisting that one description must be the ontologically privileged one does not make science more objective. It simply hides the perspective under which the object was first decomposed. Monod’s mistake was not his rejection of vitalism. It was treating one successful decomposition, the molecular logic of invariant heredity, as if its explanatory priority were automatically an ontological priority. And if Jacques Monod represents one pole of the twentieth-century debate, Richard Levins and Richard Lewontin provide an unusually clear expression of the other.

The Dialectical Biologist begins by acknowledging the extraordinary achievements of mechanistic molecular biology while arguing that Cartesian reductionism cannot function as a universal epistemology. Their objection is not simply that living systems are complicated, but that the traditional decompositions of organisms–environment, part–whole, cause–effect can itself distort the phenomenon. In the Cartesian picture Levins-Lewontin criticizes, parts are treated as ontologically prior to wholes, such that the properties of the whole are reconstructed from the intrinsic properties of components, and causation flows cleanly from independent variable to dependent variable, from environment to organism, from gene to phenotype, from lower level to higher level. Levins and Lewontin challenge each of these separations.

For instance, an organism does not simply confront an independently given environment. Organisms choose habitats, modify nutrient distributions, change chemical conditions, alter physical structure, construct nests and burrows, transform other species’ niches, and affect the selective pressures that later act upon themselves and their descendants. In this way, environments are not merely sources of external forces, but partly an outcome of organismal activity. The implications are profound. Physical factors acquire different biological significance depending on the organism’s organization. Organisms do not violate physical laws, but they selectively construct the conditions under which particular physical regularities become relevant to them.

Organisms modify the environment while the environment modifies organisms. Once this reciprocity is admitted, the clean distinction between subject and object of evolution begins to dissolve. The same logic applies internally: a biological part does not necessarily possess the same relevant properties in isolation that it possesses inside the organization; part and whole become reciprocally determining. This is the dialectical insight. Yet Levins and Lewontin themselves recognized a limitation. Their dialectical framework was powerful as critique, but the constructive application of dialectics to specific biological problems remained comparatively underdeveloped. Their own preface notes that earlier dialectical work had demonstrated considerable critical power while leaving a fully constructive methodology to future work. That future problem remains ours.

Organizational closure as the invariant beneath organicism

Suppose I say that organism and environment mutually constitute one another. How should I measure that complementarity? Suppose I say that the whole changes the causal properties of the parts. Which experimental intervention would distinguish that claim from ordinary context dependence? Suppose we embrace the premise that development is historically contingent. What mathematical object represents the relevant history? Suppose biological functions arise through reciprocal constraint. Which relations count as constraints, and when does their mutual dependence constitute closure? These questions expose the difference between a philosophical orientation and an operational theory. Crucially, I do not think the synthesis consists in choosing Levins-Lewontin over Monod. What we need is a formal and experimental language in which both demands become compatible. Dialectics in action.

Such a dialectical language should allow us to decompose without assuming ontological independence; identify local causes without pretending that all causal relevance is local; model relations without treating “relation” as magic; represent historical change without requiring one eternal state space; and characterize teleonomy without smuggling mysterious forces or an external designer into physics. This is the point at which relational biology becomes useful. The dialectical claim that parts and wholes mutually determine one another can be translated into a more specific question: does the system participate in producing the constraints through which its components acquire their relevant causal roles? That question can be formalized. More importantly, it can be experimentally tested.

Across the twentieth century, numerous researchers tried to characterize living organization using very different conceptual languages: Ashby developed the homeostat. Rosen developed (M, R)-systems. Von Neumann studied self-reproducing automata. Maturana and Varela developed autopoiesis. Gánti proposed the chemoton. Eigen and Schuster studied hypercycles. Kauffman investigated autocatalytic sets. Pattee developed semantic closure and the epistemic cut. Contemporary autonomy theorists speak about closure of constraints. Importantly, these frameworks are not equivalent and they should not be flattened into one doctrine. Their mathematical commitments differ, their biological targets differ, and in some cases their authors explicitly rejected one another’s vocabulary. Yet something keeps returning.

Living systems are repeatedly characterized as organizations in which the conditions enabling some processes are generated or maintained through other processes belonging to the same organization. That recurrence is important. In Closing the Loop, we noted that organizational closure appears across historically distinct frameworks despite the fact that many of them developed with surprisingly little mutual citation. In Wimsatt’s sense, this convergence can be treated as a robustness argument. If multiple partially independent conceptual routes repeatedly reconstruct something like organizational closure, the possibility increases that they are tracking a real organismal feature rather than merely reproducing one school’s terminology. But here “closure” must be interpreted carefully, as it does not mean energetic isolation.

Organisms are radically open to matter and energy. Nor does closure mean that every microscopic cause inside an organism belongs to one gigantic loop. Quite the opposite, the concept of organizational closure is relational and target-dependent. The stronger claim is that some constraints indispensable to the organization are produced, maintained, repaired, or replaced through processes belonging to that organization. In conventional machines, many causal conditions are externalized. An engine does not synthesize its own lubricant from environmental precursors to enable its continued operation. An organism, in contrast, internalizes far more of its causal conditions. This does not make it independent of the environment, but it changes the topology of dependence. That distinction is more useful than saying merely that organisms are “more complex” than machines.

Once closure becomes the explanandum, an interesting thing happens to causation. Part I began with Aristotle’s four causes because the history of mechanism can partly be read as a progressive narrowing of legitimate causal vocabulary. Material, formal, efficient, and final cause gradually gave way to a scientific ideal dominated by efficient causation: what prior event, force, interaction, or activity produced the later event? The temptation now would be to reverse the history and simply reinstall Aristotle. That would be a mistake; there is no reason to postulate four additional physical forces operating alongside electromagnetism, gravity, and the interactions described by contemporary physics. Nor do we need to adopt classical hylomorphism wholesale. The more modest and productive proposal is to interpret Aristotle’s causes as causal-explanatory roles.

Material causation asks what is transformed or what resources make a transformation possible. Efficient causation asks what locally produces, channels, or brings about the transformation. Formal causation asks what makes one transformation rather than another specific, admissible, or selected. Final causation asks in virtue of what contribution a transformation counts as functional for the organization within which it occurs. Notice the distinction is perspectival, but not arbitrary. A single molecular structure can occupy several causal roles. An enzyme may be an efficient cause insofar as its interaction accelerates a reaction. The same enzyme’s three-dimensional specificity can play a formal role insofar as it constrains which substrate is transformed. In another biological mechanism, the enzyme itself becomes material to be produced, degraded, transported, or repaired. Its contribution to maintaining a viable metabolic organization can ground a functional or final role.

Above all, nothing supernatural has been added. We merely have changed the questions. This distinction becomes especially important under organizational closure. Suppose a constraint $C$ enables process $P$. If $C$ is externally supplied, then a local efficient-causal story may be sufficient for explaining $P$. But if $C$ is itself produced or repaired through processes belonging to the organization, several additional questions immediately appear. What material resources instantiate $C$? What makes the new structure produced at time $t+1$ relevantly the “same” constraint as the one it replaces at time $t$? Which organizational relations preserve its specificity? Why does regenerating $C$, rather than producing some other physically possible structure, count as maintenance? The efficient-causal chain alone does not answer these questions.

This is the central intuition behind what we can begin calling closure-induced causal pluralism. The claim is not that Aristotle was totally right and Rosen proved mathematically; it is an explanatory dependence claim. Once endogenous regeneration of efficient causes becomes part of what we want to explain, additional distinctions concerning material realization, specificity, organization, and viability-relevant contribution become necessary to say what the closure physically amounts to. This is how the statement “living systems are closed to efficient causation” can coexist with the statement “efficient causation is not sufficient to explain living systems.” The first concerns topology within a causal register. The second concerns what causal roles must be distinguished to explain how that topology is materially realized. These are different claims.

Final causation deserves special caution because no concept in this discussion carries more historical baggage. To naturalize finality, we do not need future states literally reaching backward through time to cause present events, we do not need conscious intention, we do not need an external designer, and we do not even need to identify every attractor with a goal. The relevant question is much simpler: in virtue of what does an activity count as contributing to the organization whose continued existence makes that activity possible? Suppose process $A$ contributes to maintaining process $B$, while $B$ participates in regenerating the conditions under which $A$ remains possible. The effect of $A$ is then not merely correlated with system persistence, but it participates in maintaining the organization through which $A$’s own functional role is realized.

This recursive dependence offers a route toward intrinsic biological normativity. But I would now be more conservative than the wording that appeared in earlier chapters of this trilogy. Final causation should not simply be defined as the inversion of efficient causation. Two processes can causally support one another without thereby generating a biologically meaningful function; reciprocal causation is not automatically teleology, additional conditions matter. The effect must contribute to the maintenance, viability, reconstruction, or reproduction of an organization for which that contribution is relevant. The organization must, in turn, participate in sustaining the conditions under which the contributing process remains realizable.

This subtle distinction matters a lot because otherwise every stable feedback loop becomes teleological. Yes: a hurricane maintains structures through reciprocal physical processes, a flame can contribute to sustaining the conditions under which combustion continues, and a crystal can propagate ordered structure. However, we should not infer from every form of circularity that the system possesses the same kind of intrinsic functionality as a cell. Organizational closure may be necessary for the account I am developing, but closure alone is not the whole account. This is exactly why a synthesis is preferable to a slogan.

Constraints: where mechanism and organization meet

The concept doing much of the work here is constraint. Unfortunately, constraint is another word that can become so general that it explains everything and therefore nothing. In classical mechanics, constraints restrict configurations or trajectories relative to a mathematical description. In biology, the term has been broadened to describe relatively persistent structures or conditions that selectively channel faster processes. As an example, we can conceive enzymes as constraints over reaction pathways. But a molecule is not intrinsically “a constraint” merely because it exists; constraint is a relational role. In A Matter of Time, we define the relevant biological notion operationally: a candidate constraint must be identified relative to a process that it canalizes, a characteristic interval over which it remains sufficiently stable, and an organization for which its contribution matters.

This temporal relativity is crucial. Following our example above, an enzyme can be effectively invariant relative to a reaction occurring in milliseconds while itself being synthesized and degraded over minutes or hours. Organisms therefore do not preserve matter, they preserve relations through material replacement. That is one of the deepest differences between living organization and the ordinary artifact ideal. A clock can preserve its organization because its gears persist, but a cell preserves its organization precisely by replacing the matter that realizes it. This is why a process ontology becomes scientifically useful rather than merely philosophical. Identity is not located in unchanging material components. It is located in the historically sustained organization through which changing components continue to occupy relevant roles.

Furthermore, constraints provide a bridge to mechanistic biology because they can be experimentally perturbed. We could disable a constraint, then ask what fails and whether the system can restore, reroute, or replace the constraint. At that point, “the whole determines the parts” stops being a metaphysical proclamation and it becomes a realizable intervention. This leads to what I think is one of the most practically important consequences of the entire trilogy. Organizational approaches and their closure-sentitive explanations should not replace mechanistic biology, but they should change what counts as a complete and informative non-reductionist mechanistic explanation for certain explananda.

Consider protein synthesis. A conventional molecular mechanism might include mRNA, ribosomes, tRNAs, aminoacyl-tRNA synthetases, initiation and elongation factors, energy carriers, and their interactions. This explanation is legitimate, enormously useful, and experimentally grounded. A closure-sensitive explanation, in contrast, asks a second layer of questions. How are the ribosomes produced? How are the synthetases maintained? How are membrane gradients preserved? Which metabolic processes provide the necessary precursors? How is the translation apparatus repaired after perturbation? Which dependencies are supplied by the environment, and which are regenerated by the focal organization? The point is not to continue recursively until the entire universe appears in the diagram. The explanandum determines where we stop.

A Matter of Time therefore proposes a target-relative experimental workflow rather than a metaphysical verdict. First, one specifies the phenomenon, spatial scale, temporal window, observables, candidate organizational boundary, and viability variables. Second, candidate constraints are distinguished from faster processes and an enablement graph is constructed in which edges represent production, maintenance, repair, or modulation. Third, candidate constraints are selectively perturbed such that their regeneration and external dependence are tested. Finally, the inferred closure is challenged across alternative decompositions, boundaries, perturbations, and timescales.

In this way, organizational closure becomes falsifiable. A claim about autonomy is weakened if a supposedly internal constraint is actually supplied by the experimenter, or if the organization cannot regenerate or substitute it under the conditions in which regeneration was predicted, or if the inferred closure disappears under modest changes of scale or decomposition. There is therefore no need for an experimentum crucis proving that “an organism is closed”. We should instead construct an evidence profile. This is exactly the kind of methodological move organicism has historically lacked, and it is why I think the dichotomy between reductionism and holism is badly formulated. Decomposition is not the enemy. We need decomposition in order to identify the relations. The mistake is believing that because we can decompose experimentally, the decomposition is unique and the decomposed parts must be ontologically independent.

Another benefit of this framework is that it prevents the phrase “top-down causation” from becoming a placeholder for mystery. When somebody claims that the whole causes the behavior of its parts, several very different relations may be hiding underneath which are not necessarily one homogeneous causal relation. A better formalism should therefore type its arrows. This simple move allows us to retain the relational insight of dialectical biology without inventing a ghostly whole that pushes downward on molecules. On the one hand, the organism as a whole matters because organization changes which causal capacities components instantiate. On the other hand, the components matter because the organization has no material existence independent of their activities. There is reciprocal dependence without requiring reciprocal efficient forces at every level.

This is also where contemporary philosophy of mechanisms and organizational biology can genuinely learn from one another. New mechanists have developed sophisticated tools for analyzing constitutive relevance, intervention, decomposition, and multi-level explanation. Autonomy theorists emphasize that the conditions enabling those mechanisms are themselves biologically maintained. Neither perspective needs to eliminate the other. The synthesis is not “mechanism plus a mysterious whole,” but a mechanism whose explanatory boundary can expand to include the production of its own enabling conditions. Having clarified how constraints serve as a bridge between mechanism and organization, it is worth exploring the formalization of this connection.

If constraints provide the bridge to experiment, relational biology provides one candidate mathematical language for the organization itself. Rosen’s radical move was not simply to draw triangular diagrams, but to treat the functional relations among processes as primary objects of analysis. This matters because state-space approaches normally begin with a list of variables and possible states, together with rules describing transitions among them. That strategy has been extraordinarily successful throughout physics, chemistry, engineering, and biology. Nothing in the present argument makes state spaces illegitimate. The problem appears when the variables, boundaries, or transition relations themselves become part of the phenomenon.

A conventional state-space description asks: given this space of possibilities and these transition rules, what trajectory follows? A relational description can ask a logically prior question: what organization makes these transformations possible at all? Rosen’s $(M,R)$-systems and Hofmeyr’s later $(F,A)$-systems attempt to capture functional entailment rather than merely movement through predefined states. In Closing the Loop, we used $(F,A)$-systems because they provide a biologically richer representation of fabrication, assembly, symbolic specification, and semantic interpretation. Their organization is closed to efficient causation while remaining open to free-standing formal causes, allowing adaptive syntax to coexist with self-maintaining semantic relations—a clear limitation of state-space-based approaches.

The biosemiotic interpretation of Rosen’s relational biology is particularly suggestive. Syntax concerns relations among symbols, semantics concerns the way symbols become physically interpreted, pragmatics concerns what the resulting interpretations do for the organization, and the symbol vehicle itself has a material realization. Once these causal roles are coupled through a self-constructing organization, the biological code is no longer a disembodied string or code. It becomes a material relation among description, interpreter, activity, and consequence. This is what computational metaphors often conceal. Computer science normally starts after the symbol system exists, but biology has to explain how the symbol system became possible.

Experimental organicism

Relational biology, however, has a serious limitation. Its diagrams are time-invariant, but organisms are not! Development canalizes trajectories, environmental interactions leave traces, and evolution changes the conditions under which future evolution occurs. So, how can a static relational diagram represent such a system? We can start by associating characteristic temporal scales with the processes represented in an $(F,A)$-system. Each relation then acquires a timescale over which it operates and another scale over which the relevant constraint is produced, maintained, or replaced. As shown in Closing the Loop, this temporalization can be redescribed probabilistically as an asynchronous dynamic Bayesian network, allowing different processes to update on different schedules rather than forcing the organization into a synchronized clock.

However, as I have clarified in A Matter of Time, a relational diagram can remain time-invariant at the level of its entailment structure even though every physical realization is diachronic because its material carriers are precarious. This distinction prevents temporalized relational biology from collapsing back into exactly the state-space ontology it was intended to supplement. The asynchronous Bayesian network is not the organism, but one admissible stochastic realization of that organization. The relational layer constrains what must be preserved functionally, the temporal layer describes how particular material carriers instantiate those relations through history, and the state layer therefore sits over the relational layer rather than replacing it. The same organizational pattern may be instantiated through different substrates, timescales, developmental histories, or environmental couplings; that is exactly what we expect from life!

Importantly, if this picture is to become more than philosophy, relational biology itself must be reformulated. Historically, Rosen began developing $(M,R)$-systems in the late 1950s, when category theory was still young. His work was conceptually radical but mathematically idiosyncratic. Some notation is historically dated and some arguments remain disputed. In fact, some biological assumptions, such as those inherited from early molecular biology, require revision. Rosen should therefore not be treated as scripture or pinnacle. His ideas should be treated as a research program to be matured. Modern category theory offers tools that simply did not exist in mature form when much of relational biology was constructed.

Crucially, category theory does not become a theory of life merely because it contains sophisticated mathematics. Every categorical construction must eventually return to a modeling relation. Which biological process instantiates this morphism? Which intervention distinguishes two proposed diagrams? Which equivalence is biologically meaningful rather than merely formal? Which observation would falsify the selected closure structure? A mathematically elegant organicism that cannot answer such questions would reproduce the same descriptional vitalism I criticized earlier, only in categorical notation. The real promise of modern category theory is not abstraction for its own sake, but the possibility of designing new experimental questions.

The strongest test of an organizational theory is not whether it retrospectively describes a living cell, but whether its principles allow us to construct, perturb, or distinguish systems in ways conventional frameworks do not. Hofmeyr’s biochemically realizable $(F,A)$-systems remain important because they translate abstract functional closure into recognizable cellular processes. Closing the Loop uses that framework precisely because fabrication, assembly, symbolic specification, catalytic closure, structural closure, and information handling can be related to concrete biochemical organization. Nevertheless, relational models have rarely been subjected to the kind of systematic experimental testing expected of mainstream mechanistic biology. If we want to take organicism to the next level, this should change.

One promising direction is to treat the reconstruction of organizational closure as an experimental inference problem. We might perturb candidate constraints independently where possible, measure how their disruption propagates, determine whether restoration comes from internal processes or external support, compare different decompositions, track turnover rates, map dependencies across timescales, ask whether the same functional organization survives material substitution, and search for transitions where a relation changes from environmentally supplied to internally regenerated. Such experiments could be especially revealing in protocells, minimal cells, synthetic consortia, developmental systems, and biohybrid constructs.

In this way, the goal is not to “prove life,” but to detect organizational transitions. A Matter of Time proposes precisely this kind of profile-based reasoning, emphasizing that no single signature is sufficient. In the non-exhaustive, preliminary metrics we proposed, strong evidence should emerge from convergence among semantic closure, measurement–control complementarity, anticipatory modulation, affordance reconstruction, syntactic open-endedness, and viability-corrected skill acquisition. This is a very different philosophy of experiment from searching for one magic molecule or one universal life score. It treats living organization as a pattern of relations whose empirical reality must survive multiple interventions. That is how organicism becomes laboratory science.

There is another problem relational biology can no longer postpone: multicellularity. Most classical organizational models are implicitly cellular. This is understandable. Cells provide an unusually clear scale at which membranes, metabolism, heredity, catalytic organization, self-production, and environmental exchange meet. But multicellular organisms complicate autonomy. A differentiated cell inside an animal is often not autonomous in the same sense as a free-living bacterium; it depends on circulatory systems, extracellular matrices, endocrine signals, immune regulation, neighboring cells, developmental history, and organism-level physiological organization. Yet multicellular systems display capacities that isolated cells do not.

Where, then, is organizational closure in multicellular entities? The wrong answer is to assume that one scale must monopolize it. A more promising approach is to represent multiple organizational closures and ask how they couple. Cells may retain partial autonomy while becoming components of a higher-level organization. Some constraints may be produced locally, while others may be maintained at tissue level. Others may depend on organism-wide circulation, but still others may be stabilized through ecological or social interaction. This immediately creates a richer picture of agency. Rather than asking whether the organism is an agent, we can ask which measurement–control organizations exist at which scales, how they constrain one another, and which forms of anticipatory organization emerge only through their coupling.

In this way, multicellular agency need not be a simple scalar amplification of cellular agency. The transition may reorganize what counts as an individual, where control is located, and how autonomy is distributed across levels. This is one reason I think category theory could become genuinely useful. The challenge is not merely to draw a larger network, but to formalize relations among organizational diagrams. Once multiple organizations share an environment, the boundary problem becomes even more interesting. A cell alters the chemical environment in which other cells operate. A microbiome alters host metabolism. A virus changes constructed niches. A plant modifies soil. A beaver modifies hydrology. A technological device modifies the affordances available to other organisms.

Organizational closure therefore cannot imply ecological isolation. What appears organizationally “external” at one scale may become constitutive at another. In Part II, I emphasized that organism and environment are not two objects completed independently before interaction; their relevant properties emerge relationally. Relational biology, however, gives us the possibility of sharpening that statement. Instead of merely saying that two organisms co-construct an environment, we can ask how their respective closure structures become coupled, which constraints remain internally regenerated, which become jointly maintained, and whether new higher-level closures emerge. Closing the Loop already identifies the coupling of multiple semantically closed systems as an open direction toward multicellular organization.

One can imagine an entire ecology of organizational diagrams: some cooperate, some exploit, some become obligately dependent, some fuse, and some generate new symbols whose interpretation exists only at the collective level. From this perspective, the major transitions of evolution could eventually be studied as transformations in the topology of closure rather than merely increases in component number. That is speculation, but it is a mathematically and experimentally structured speculation, and that difference matters. However, open-endedness creates perhaps the hardest mathematical problem in the whole program. Yes, a state-space model presupposes a fixed dynamical description, but a dynamic Bayesian network also presupposes variables. So what happens when evolution changes the relevant variables?

This is where the distinction between exploration and construction becomes unavoidable. If evolution simply moves through a gigantic but fixed possibility space, then in principle every novelty already exists as an unvisited point. But biological history often appears to create new relations of relevance. The evolution of a new organ creates new modes of interaction. A new metabolic pathway makes previously unusable resources significant. A new sensory modality converts formerly irrelevant environmental differences into signals. Symbiosis changes what one organism can do by coupling it to another. Technology transforms the human affordance landscape even faster. As we can see, a fixed state space can often be reconstructed retrospectively.

The harder question, however, is whether it could have been specified prospectively without already knowing which distinctions future evolution would make relevant. I do not think this issue has been solved. In my own framework, a full mathematics of topology-changing asynchronous dynamic Bayesian networks remains undeveloped. The next formal step may require representing transitions among state spaces rather than trajectories inside one. Perhaps categories of possible organizations will become more useful than one universal phase space. Perhaps the mathematical object should encode transformations of measurement relations themselves. Perhaps neither approach will work. But the question has changed. Instead of asking only “what state comes next?”, we can ask “what makes a new kind of state describable?” That is the kind of question biological open-endedness deserves.

On neomachines

In 1976, Czech director Václav Mergl adapted Anatoly Dneprov’ sci-fi story Crabs on the Island into the short animated film Krabi (Crabs). The film presents self-reproducing robotic systems designed in a military context. Given material resources, the artificial creatures reproduce, undergo selection-like competition, and ultimately exceed the controlled situation in which they were intended to operate. Although von Neumann also conceived of a device of this kind during the same decade, its realization remains beyond the reach of our current technology. Suppose that, at some point in the future, we succeed in creating non-organic devices capable of generating and maintaining their own constraints without external intervention in the same way that life does. Let us call this next design objective a “neomachine.” Such a term needs immediate qualification.

I am not using neomachine as a synonym for the “new mechanism” literature in philosophy of science, nor am I proposing a mysterious new ontological entity. In A Matter of Time, I explicitly avoid calling closure-sensitive mechanistic explanations a new kind of “neomechanism.” A neomachine as hypothesised here is instead a design constraint; it would be an artificial physical organization capable, to some nontrivial degree, of producing and maintaining the measurement–control relations required for its own continued operation, while remaining capable of revising some of those relations through its history of interaction. This is more demanding than adaptive control, more demanding than self-repair, more demanding than self-reproduction, and more demanding than generative novelty.

Imagine a system that encounters a physical regularity for which it possesses no sensor. A conventional learning system cannot learn from a variable it cannot measure. A neomachine, in the strongest sense, would need some capacity to construct a new measurement relation: to modify its own material organization such that a previously irrelevant environmental difference becomes detectable and actionable. Now imagine that this new sensor changes the actions available to the system, those actions alter the environment, the altered environment changes the system’s viability conditions, and the system subsequently reorganizes its measurement–control architecture again. At that point, adaptation is no longer movement through a predefined state space; the system is participating in the reconstruction of the space.

This is related to Peter Cariani’s notion of emergent semantic functions and to the broader problem of how systems can generate new sensors and effectors rather than merely optimize existing ones. It is also where the computer metaphor starts to break from inside. Modern computation depends on a powerful abstraction: software can vary while hardware remains sufficiently stable. That separation is never physically absolute, but it is operationally extraordinarily useful. A program can be copied from one machine to another because the machines implement sufficiently equivalent interpretation relations. The same source code can operate across multiple physical devices because layers of architecture preserve the meaning of the relevant symbols. Biology, however, does not fit this picture comfortably.

Cellular “hardware” is constantly replaced: proteins degrade, membranes reorganize, molecular complexes assemble and disassemble. The structures responsible for interpretation are themselves products of interpreted descriptions. DNA contributes to constructing ribosomes and enzymes, but ribosomes and enzymes are required to make DNA biologically consequential. The hardware produces the interpreter of the software, and at the same time, the interpreted software participates in reconstructing the hardware. The distinction becomes circular. Importantly, this does not mean biological information is unreal, but that biological information is materially grounded. The syntax is effective only because a historically produced organization maps it into transformations. This is what semantic closure captures and it can be formalized using relational biology.

In $(F,A)$-systems, the organization remains closed to efficient causation while permitting adaptive formal causes, producing a relation in which symbolic specification and material construction are recursively linked. A sufficiently advanced neomachine would therefore force us to ask whether programmability can survive when the architecture interpreting the program is itself continuously rebuilt by the system. Perhaps it can, but then “program” will no longer mean quite what it means in a laptop. This is another reason why I believe that reconstructing Rosen’s relational biology in modern mathematical terms will prove highly fruitful, as it could allow us to conceive models of computation alternative to those based on state transitions, something that is not strictly equivalent to Turing machines without falling into the hypercomputation realm.

One possibility is that the distinction will become scale-relative. Some relational invariants could function as software relative to faster material processes while themselves being reconstructed at slower timescales. Another possibility is that what persists will be neither hardware nor software but a hierarchy of mutually regenerating constraints. This is not a problem current computer science has solved because conventional computing was designed precisely to avoid it. Engineering normally seeks stable interfaces, but life survives by rebuilding them. Research programs such as unconventional computing already enable data processing with significantly lower energy consumption and high parallel processing capacity. The approach suggested here could help us overcome the “black box” nature of reservoir systems by creating from scratch a new logic for programming distributed computational architectures.

The same framework suggests an alternative direction for AI research. Instead of asking only how to make systems better at human-defined benchmarks, we could ask how new measurement–control relations emerge. Instead of increasing model size indefinitely, we could investigate architectures in which sensing and action become materially coupled. Instead of treating embodiment as an optional peripheral connected to a central computational core, we could study systems in which bodily organization determines what can be represented. Instead of assuming a permanent distinction between learning software and fixed hardware, we could investigate architectures that construct and modify some of their own physical interfaces. Instead of treating intelligence as one scalar capacity, we could map profiles of autonomy, anticipation, semantic closure, and open-endedness.

This would be a genuinely bio-inspired AI program in a stronger sense than copying neural-network motifs, leading us to genuine open-ended systems and realistic digital twins. The most radical lesson biology can offer computation may therefore have very little to do with making the next predictor larger, as the current transformer program suggests. It may involve changing what we mean by computation. This is the direction I have called computational enactivism, a name that is deliberately dialectical. It rejects neither computation nor enactivism: computation captures syntactic transformation, enactivism emphasizes that meaning emerges through situated interaction. The point is to embed computation inside an organization rather than elevate computation into the ontology of the organization.

Closing the Loop was an initial attempt to integrate these traditions. The paper interprets semantic closure as a self-referential condition in which symbols construct and are interpreted by the organization they help sustain; temporal extension then makes anticipation possible, while environmental coupling connects the organization to affordance realization. The term “computational” therefore does not mean “life is a Turing machine.” It means that syntactic transformation remains one legitimate layer of biological organization. The term “enactivist” prevents that layer from becoming self-sufficient. Computation then becomes one process inside life rather than life becoming one process inside computation. Truly bio-inspired computing should implement measurement and control mechanisms, just as life does.

This brings us back to one of Part II’s deepest problems: machines normally inhabit possibility spaces, but evolution constructs them. That statement must be handled carefully. The stronger and more defensible claim is epistemic and organizational. The variables through which future biological possibilities become relevant are historically generated. Before eyes evolved, “visual camouflage” did not exist as the same ecological affordance. Before oxygenic photosynthesis transformed Earth’s atmosphere, entire metabolic opportunities and hazards were absent. Before multicellular bodies existed, many developmental niches for cell specialization did not exist. Before human technology, radio frequencies did not function as communication affordances for our species.

In all these cases, the physical possibilities were constrained by the universe all along. However, the biological relevance was not. This is why retrospection can create the illusion of pre-specification. Once a novelty exists, we can add a variable representing it, we can reconstruct a state space containing it, and we can simulate its dynamics. But the fact that a retrospective model can contain a novelty does not show that the novelty’s relevance was available as a biologically meaningful coordinate before the history that produced it. This is the difference between physical possibility and constructed affordance. Open-ended evolution lives in that gap. If that claim is correct, it suggests a particularly ambitious class of experiments. Rather than measuring only adaptation within a fixed sensor–effector architecture, we could search for systems that acquire new measurement relations.

Imagine a synthetic protocell exposed to an environment containing a physical regularity that initially has no functional consequence. Over repeated cycles, the system evolves or constructs a molecular interface that couples that regularity to metabolism, such that the environmental difference becomes a signal. Now suppose the new signal supports an action that changes local conditions. The action creates a new ecological niche and the niche stabilizes the interface. A semantic primitive has emerged not because an experimenter assigned a label but because the relation has become integrated into the system’s viability organization. This is the kind of transition a theory of semantic closure should eventually be able to detect. It would be far more informative than training another controller on another benchmark. It would tell us something about how relevance originates. This is also where synthetic biology and artificial life could converge, experimentally constructing conditions under which meaning becomes materially grounded.

The importance of bioethics

It is impossible to finish a contemporary trilogy on machine metaphors without discussing artificial intelligence. The reason is not that AI has solved the problems described above, quite the opposite, historically AI shows how quickly behavioral success can harden into ontological interpretation. Large language models generate fluent explanations, code, essays, mathematical proofs, plans, translations, and conversations. Their usefulness is not in question, but because their outputs inhabit the same symbolic medium humans use to display intelligence, LLMs are uniquely vulnerable to what I would call a simulacrum problem. Since their performance looks intelligent, we therefore begin treating the underlying organization as if it must resemble whatever we mean by intelligence in ourselves.

The scientific community itself remains divided over how far this inference should go. Melanie Mitchell and David Krakauer describe an active disagreement over whether contemporary large language models possess anything appropriately described as humanlike understanding, and they emphasize that part of the dispute arises because “understanding” itself is theoretically unstable. That uncertainty should make us more precise, not more dismissive. It would be wrong to say that language models merely copy sentences from their training data. They generalize, they compose, they can produce novel outputs, they acquire nontrivial internal structure, and they can sometimes solve tasks that were not explicitly represented as memorized examples. However, none of those observations settles whether they instantiate the organizational properties of biological agency described earlier. These are different questions.

One way I currently find useful for thinking about large language models is as technologies for accessing and recombining an enormous residue of human cultural activity. Books, articles, code, websites, conversations, documentation, arguments, stories, and explanations are not raw data in any philosophically innocent sense. They are products of human communities, embodying centuries of conceptual selection. Language itself contains compressed distinctions generated through collective history. Training a model on cultural artifacts therefore exposes it to the products of intelligence before the model performs a single inference. Modern language assistants are also shaped through explicit feedback procedures. Human judgments have been extensively used to rank outputs, guide model behavior, and construct preference models, although contemporary systems increasingly mix human and AI-generated feedback.

Calling this “merely imitation” misses something important. Culture is a real cognitive technology. Humans also inherit languages, norms, tools, concepts, and institutions they did not invent individually. The interesting scientific question is therefore not whether LLMs are contaminated by human influence—every cultural intelligence is—but how the organization producing their capacities differs from biological, developmental, and social forms of intelligence. This is precisely where I think current AI research could become much more interesting. Language models provide extraordinary instruments for studying how capabilities are encoded in cultural products, how symbolic regularities can support generalization, and how new behavior can arise from recombining structures distributed across enormous corpora.

This may teach us a great deal about culture, but it does not automatically tell us how intelligence originates in a living system. The product and the source are different explananda, and the organizational framework developed here helps separate them. A language model may display impressive competence while having no metabolism. That fact alone does not disqualify it from every possible form of intelligence, but it does matter if the question is biological autonomy. LLMs do not ordinarily manufacture the data center that runs them, they do not replace failed processors, they do not maintain the electrical grid, they do not determine the physical sensing channels through which its inputs enter, and they do not construct the basic tokenization and implementation architecture through which its symbols are interpreted. Their viability conditions are overwhelmingly supplied by an external technological infrastructure, by an engineer.

Again, none of this makes LLMs useless. A microscope is useful without being autonomous, a scientific simulation can discover something without being alive, and a search engine can transform cognition without becoming a biological agent. The problem begins when capability, autonomy, agency, understanding, and life are collapsed into one scalar dimension called “intelligence”. The threshold framework in A Matter of Time gives us a different vocabulary. A system might exhibit sophisticated goal pursuit while lacking closure, it might display history-sensitive control while relying on externally maintained measurement relations, and it might generate enormous novelty while remaining inside a fixed ontology of inputs and outputs. Different profiles are possible. This is scientifically more interesting than asking whether a model is simply “intelligent” or “not intelligent”.

The ethical stakes become clear when we realize that machine metaphors are not merely descriptive, but constructive. When researchers speak about biological “chassis,” “circuits,” “platforms,” “modules,” and “devices,” the language encourages particular experimental practices. This is not necessarily bad, as engineering metaphors make living systems tractable. They encourage standardization, they permit modular design, and they make construction imaginable. But as Joachim Boldt has shown, this vocabulary also foregrounds control while backgrounding evolutionary change and ecological interdependence, precisely the properties relevant when engineered organisms leave tightly controlled settings. Nora Vaage similarly argues that the living-machine metaphor carries assumptions about human control over biological material that are simultaneously epistemological, ontological, and ethical.

This does not mean researchers should stop saying “genetic circuit.” It means the metaphor should come with a warning label. A circuit in electronics normally does not mutate, it does not reproduce, it does not exchange genes with neighboring circuits, and it does not alter the ecology in which future circuits operate. Nevertheless, a “genetic circuit” can. The analogy remains useful precisely where its disanalogies remain visible. This becomes even more important under the neomachine program I discussed above. The closer an engineered system approaches endogenous constraint production, semantic reconstruction, and open-ended agency, the less adequate ordinary engineering assumptions become. This advocates for the use and development of new metaphors in biology and to build ethics into research programs before such a technology exists.

Robert Rosen understood this point unusually well. In a 1997 interview later transcribed online, he was asked about the possibility that his theory of living systems might be used to synthesize organisms. Rosen responded that such a capacity could, in the long run, be more dangerous than the development of atomic fission in the last century because it would permit fundamental alteration of living organization while our understanding and social wisdom remained limited. One does not need to accept the comparison literally to take the warning seriously. Rosen’s concern was not merely that synthetic biology could become powerful and dangerous. Biology as a whole was already powerful. The deeper worry was that a theory capable of reconstructing the organizational principles of life would change the type of intervention available.

Nowadays genetic engineering modifies components. A more matured engineering science of organizational closure might eventually modify the conditions under which components acquire functions. That is a different order of intervention. Such a development would also make the usual separation between basic and applied theory difficult to maintain. If a mathematical framework tells us which constraints must be closed for a system to become autonomous, then the same theory that explains autonomy may become a blueprint for constructing it. At that point, theoretical biology will no longer be ethically downstream from engineering. The theory itself becomes dual-use knowledge. Rosen’s warning leads to another question that technological narratives often postpone: who benefits?

Suppose the organizational approach eventually contributes to self-maintaining therapeutic systems. Imagine synthetic cells capable of repairing tissues, biohybrid devices that adapt locally inside a patient, artificial metabolic systems that manufacture compounds with radically lower energy requirements, distributed living sensors that respond to environmental contamination, and new architectures inspired by biological measurement and control that dramatically reduce computational energy consumption. Each possibility is exciting, but none guarantees equitable benefit. A technology can be scientifically revolutionary and socially exclusionary. It can solve one biological problem while concentrating ownership. It can lower energetic costs while increasing infrastructure dependence. It can improve treatment while remaining inaccessible to the populations carrying the greatest disease burden. It can generate ecological benefits in one region while externalizing risks elsewhere.

These distributional questions are not optional moral decorations added after the “real science” is finished; they influence which problems are investigated, which systems are optimized, which risks are tolerated, and whose failures count as unacceptable. The same lesson applies to AI. If technological success is evaluated only by what a system can do, then the social organization through which that capability is deployed disappears from the model. Yet an algorithm deployed by a hospital, military institution, school, corporation, or scientific laboratory does not have the same consequences simply because the code is identical. Context is causal; institutions are part of the effective machine. This is another place where the relational perspective generalizes beyond biology.

There is one final ethical consequence of machine metaphors that I think deserves more attention, and this is the fact that they eventually return to describe us. If cognition is computation, and computers increasingly perform cognitive tasks, then a seductive conclusion follows: humans are replaceable computational devices. But this inference depends on defining a human activity by its output under an utilitarian lens. A teacher produces explanations, a programmer produces code, a scientist produces hypotheses and papers, a physician produces diagnoses, and a writer produces text. Then, if another system generates a sufficiently similar output, the function appears replaceable. This logic is perfectly coherent for many machines because artifacts are commonly individuated by externally assigned functions.

If two pumps move the same quantity of water with comparable reliability, substituting one for the other may be straightforward. Human practices are not always organized that way. A teacher is embedded in trust, responsibility, institutional norms, developmental relationships, and a local history with students. A scientist does not merely generate plausible sentences but participates in experimental communities that assign responsibility for claims. A physician’s role includes accountability, judgment under uncertainty, and relationships with patients. A writer participates in a history, audience, and social context that cannot be reconstructed from the text alone. Automation can certainly transform or eliminate tasks, that is an empirical and economic reality, but task substitution is not equivalent to organismal or social equivalence.

Machine metaphors become politically consequential when a decision about how institutions should distribute the benefits of automation is redescribed as an inevitable consequence of what humans “really are”. Claiming that “humans are information processors; another information processor can produce the same output; therefore the human is obsolete” is not a theorem, but a metaphysical argument wearing the clothing of engineering. Part I showed how machines became metaphors for organisms. The social danger appears when machine metaphors return and become a theory of human worth. This insight could become one of the most practical contributions of theoretical biology to emerging biotechnology and ethics. A theory developed to understand life may eventually help us decide how much life-like autonomy not to build.

Conclusion

The history of machine metaphors in biology is about productive analogies that repeatedly transformed the sciences that used them. The clock gave life mechanical order, the engine gave it energetic throughput, the computer gave it control. Each metaphor succeeded because some part of the organism genuinely could be understood through machines. And each metaphor eventually failed because the organism did something machines left outside. The first installment of this trilogy reconstructed that success. The second examined what happens when success becomes metaphysics. This third part has tried to do something more difficult: preserve the achievement while changing the conceptual architecture.

The dialectical synthesis presented in these three chapters is therefore not anti-mechanistic. Mechanistic explanations remain indispensable. What must disappear is the assumption that mechanisms always operate inside boundaries whose relevant components, variables, constraints, and purposes can be treated as externally fixed. Living organization makes some of those conditions endogenous and this observation leads us from classical and conventional mechanism toward closure-sensitive mechanism. Such a transition constitutes the first step that, in the long run, leads from what we currently term a “machine” toward the hypothetical “neomachine” I discussed above.

But the neomachine should not become another metaphysical slogan. We do not yet know how to build one. We do not yet possess a complete mathematics of topology-changing relational organizations. We do not yet know how semantic primitives can be experimentally generated rather than externally assigned. We do not yet know how cellular models of closure should be generalized to tissues, holobiots, economies, and ecologies. We do not yet know whether hardware–software separation remains meaningful in systems that continuously reconstruct both. These of course are not reasons to abandon the program, but reasons the program exists.

The central methodological lesson is therefore modest and it points toward an experimental question: can we construct a physical organization in which symbols participate in producing the machinery that interprets them, in which measurement and control are maintained through the organization they regulate, in which the relevant constraints are regenerated across multiple timescales, and in which some of those relations can themselves be revised without destroying organizational continuity? If the answer is no, we will have learned something profound about the distinction between organisms and artifacts. If the answer is yes, it would mean that biology compelled us to invent a different kind of machine (one we can barely conceive of today) rather than implying that organisms were, in fact, computers from the very beginning.

This possibility is exciting but it should also make us cautious. Rosen’s ethical warning becomes more relevant the closer theoretical biology approaches synthetic biology. A framework capable of explaining how living organization produces its own conditions of existence may eventually enable technologies that do the same. Those technologies could transform medicine, sustainable manufacturing, environmental engineering, and information processing. Unfortunately, they could also produce new ecological risks, new forms of concentration and inequality, and new dual-use problems. Václav Mergl’s Krabi remains a useful cultural reminder that self-production is not merely another feature added to an artifact. It changes the relation between control and construction.

We therefore need better machines, but before that, we need better distinctions. The machine metaphor became successful because it allowed scientists to make life intelligible by comparing it with the most sophisticated artifacts they knew how to build. Perhaps the next stage will reverse the direction of explanation. Rather than continually forcing organisms into the image of our machines, we can allow organisms to reveal how impoverished our machines still are. So far we have built machines to understand life, and the time has come when life could teach us what machines might become.

And if one day we construct an artificial system capable of producing and repairing the constraints that constitute it; of maintaining the physical interfaces through which its world becomes meaningful; of anticipating possible futures; of creating new measurement and control relations; and of changing some of the possibilities available to it without losing the continuity of its organization, the important conclusion will not be that Descartes was right. but that the word machine no longer means what Descartes meant. Every machine metaphor has an expiration date. Because, when science works properly, the phenomena eventually force metaphors to evolve.

The future of machine metaphors in biology therefore lies neither in abolishing machines nor in declaring that everything is one. It lies in making the relation reciprocal, in transforming metaphors into models, models into experiments, experiments into realizations, and realizations back into a new understanding of what life (and perhaps one day a machine) can be, always maintaining a bioethical perspective and reflecting on the social repercussions that our scientific developments could have.