Machine Metaphors in Biology II: Limits and Challenges

Introduction

The first part of this trilogy ended with a paradox. Machine metaphors became dominant in biology not because they failed, but because they worked. They transformed organisms into scientifically tractable objects. They allowed us to decompose bodies into organs, organs into tissues, tissues into cells, cells into molecular pathways, and pathways into interactions that could be measured, manipulated, and eventually redesigned. The clock metaphor developed in the 17th century made order mechanical. The engine metaphor alluded to in the 19th century made metabolism energetic. The computer metaphor formulated in the 20th made control informational. Molecular biology made heredity readable; biotechnology made it writable. The result was one of the most successful scientific programs in human history. That success must remain the starting point of any serious criticism.

It is easy to attack an obsolete machine. One can point out that organisms are not clocks because they grow, or that cells are not steam engines because they reproduce, or that DNA is not literally a blueprint because development depends on far more than nucleotide sequence. But such observations, while true, do not explain why machine metaphors survived each criticism. They survived because the meaning of “machine” changed, evolving in tandem with our technology. When the clock became too rigid, the engine introduced flows. When the engine could not explain control, cybernetics introduced feedback. When feedback was insufficient to explain symbolic specificity, the computer supplied programs, codes, memory, and information. Machine metaphors repeatedly escaped its critics by becoming a different machine ideal.

This is why the real problem is deeper than asking whether organisms resemble artifacts. The deeper question is what happens when a successful metaphor ceases to be recognized as a metaphor at all. At that point, a comparison becomes an ontology. To say that a cell can be modeled as an information-processing system quietly becomes the claim that the cell is fundamentally an information-processing system. To say that evolution can be simulated computationally becomes the suggestion that evolution itself is a computation. And once computation is sufficiently generalized, the metaphor reaches its maximal extension: organisms compute, brains compute, societies compute, ecosystems compute, physical interactions compute, and finally the universe computes itself. This family of positions is usually grouped under the broad label pancomputationalism.

Yet, even this term conceals a taxonomy whose varieties exhibit significant differences. One version makes the modest epistemological claim that computational descriptions can be applied extremely broadly. Another claims that every physical system literally implements some computation. Still stronger versions imagine the universe itself as a computer executing a fundamental program. These positions should not be conflated. A world that can be described computationally is not automatically a world that is computation. Vincent Müller captures this distinction particularly well by separating the thesis that “the world is a computer” from the weaker claim that “the world can be described as a computer”; the latter may preserve computational language as a productive metaphor without establishing it as a literal theory of reality.

This distinction will structure the present essay. The purpose of Part II in the present trilogy is not to demonstrate that computation is useless, fictitious, or somehow “unnatural.” That would contradict everything established in Part I. Organisms unquestionably perform processes that can productively be described computationally. Nervous systems transform signals. Cells discriminate molecular states. Gene-regulatory networks implement conditional responses. Immune systems retain forms of memory. It is a fact that organisms integrate information across time in order to act. It is true that computational models have generated real discoveries throughout biology. The target is therefore not computation. The target is computational imperialism: the assumption that because computational description is powerful, computation must constitute an exhaustive ontology of life—or even of reality.

This distinction opens a more interesting historical path. Just as machine metaphors were evolving from clocks to engines to computers, alternative intellectual traditions were developing in parallel. Sometimes it appeared as vitalism, sometimes as organicism, sometimes as systems thinking, dialectical biology, autonomy traditions, relational biology, physical biosemiotics, process philosophy, or organizational approaches to life. These schools were never identical. They often disagreed profoundly with one another. Some produced rigorous mathematics; others remained philosophical. Some were compatible with mechanism; others defined themselves in opposition to it. Yet across their differences, they repeatedly returned to a common intuition: the organism cannot be understood merely as a collection of externally related parts undergoing externally specifiable transformations.

Parts in living systems depend on wholes that they simultaneously help constitute. Organisms transform the environments that select them. Biological functions depend on the organization that makes them functional. Constraints do not merely act on living systems from outside; living systems construct and maintain many of the constraints governing their own activity. Biological information is not merely transmitted; it must be interpreted within an organization that makes symbols consequential. And evolution does not simply explore a timeless catalogue of possibilities. It changes the very contexts in which possibilities acquire biological meaning.

Thus, the history of these alternative narratives forms a counter-history to the origin and evolution of machine metaphors. But a counter-history must avoid its own temptation. Rejecting reductionism or mechanical philosophy does not automatically give us a science of wholes. Saying “everything is connected” explains almost nothing. Invoking “emergence” without specifying a mechanism can become a sophisticated form of ignorance. Treating organisms as irreducible subjects can quietly reintroduce the very vitalism that mechanism once helped biology escape. The task of this essay, and of this entire trilogy, is, therefore, dialectical.

We must ask not simply where mechanism fails, but what exactly fails when mechanism is universalized; not simply whether organisms are different from machines, but which distinctions remain after we allow machines to become adaptive, distributed, stochastic, self-repairing, and intelligent; not simply whether computation has limits, but what kinds of limits are relevant to biological explanation. To find the answer, we must return to a fork in the road we encountered in Part I. We must return to Kant.

The road not taken: the organism as a problem for mechanism

In Part I, Immanuel Kant appeared at a crucial historical moment. Mechanistic philosophy had already transformed the scientific image of nature. Descartes had turned animal bodies into automata. Newtonian physics had demonstrated the extraordinary explanatory power of mathematical laws governing matter in motion. La Mettrie had pushed the analogy further, extending machinic materialism to human beings themselves. Kant accepted much of this scientific revolution. But organisms posed a peculiar problem.

Consider a watch. Its parts are organized for one another. A gear contributes to the movement of another gear; the spring powers the mechanism; the hands display the result. The watch is unquestionably organized. Yet the gears do not produce the spring. The spring does not repair the gears. The watch does not rebuild the watchmaker.Organisms are different. Their parts are not simply assembled into a functioning whole; the organized whole participates in producing, maintaining, repairing, and reproducing the very parts through which they exist. Cells generate tissues, but tissues also regulate cellular behavior. Organs depend on developmental processes that occur only within the organismal context they help bring into being. Components continuously degrade and are replaced while the organization persists.

Kant called such beings natural purposes. The importance of this concept is not that Kant offered a mysterious life force. Quite the opposite, he identified a structural difficulty in extending artifact-based explanation to organisms. In an artifact, organization is imposed by a cause external to the organized system. In an organism, organization appears recursively implicated in its own production. This distinction is the seed from which much later organicist thought would grow. However, the interesting difference is not that machines are “simple” while organisms are “complex.” Modern aircraft, semiconductor fabrication systems, neural networks, and global communication infrastructures are enormously complex. Nor is the distinction that organisms change while machines remain static. Modern machines learn, reorganize, update, compensate for damage, and alter their behavior.

The Kantian problem is more fundamental: Where does the organization that makes the system what it is come from? For a conventional artifact, the organization is predominantly heteronomous. Its constitutive constraints—the arrangements that make it a watch rather than a pile of brass—originate in a design and fabrication process external to the finished watch. For an organism, many constitutive conditions are endogenous. The system participates in generating the organization through which it continues to exist. This question did not disappear when Kant died. It merely changed vocabulary.

Vitalism: the wrong answer to a real problem

Even during the nineteenth and early twentieth centuries, dissatisfaction with purely mechanical explanations repeatedly generated appeals to vitalism. Vitalists were reacting to genuine empirical puzzles. Developmental biology was particularly troublesome. An embryo did not look like an artifact being assembled component by component according to a visible mechanical plan. It regenerated, compensated for perturbation, and often arrived at similar outcomes despite dramatically altered starting conditions.

Hans Driesch (1867–1941) performed many famous embryological experiments that became emblematic of this problem. Working with early sea-urchin embryos, Driesch separated blastomeres that some mechanistic “mosaic” models suggested should correspond to predetermined portions of the future organism. Instead, separated cells could regulate their development and form smaller but relatively complete larvae. The phenomenon posed a serious challenge to overly rigid conceptions of development as the mechanical execution of a prelocalized plan. Driesch eventually concluded that ordinary physico-chemical causation was insufficient. He revived an Aristotelian term, entelechy, to designate an organizing principle that could guide development without behaving like conventional material force.

Historically, this is easy to caricature. One can simply say that Driesch observed a difficult biological phenomenon, failed to explain it mechanistically, and inserted a mysterious force. In that sense, vitalism eventually became scientifically sterile. Once “life” itself becomes the explanation of what living systems do, the explanation becomes circular. But dismissing vitalism too quickly hides an important lesson. Vitalists were often wrong about the answer while being right about the existence of a problem. Biological development really does involve regulation. Organisms really do restore organization after perturbation. Biological wholes really do constrain what parts become. The failure of vitalism did not make those phenomena disappear.

The historical challenge was therefore to construct a third position. How could biology recognize the distinctive organization of living systems without postulating an immaterial essence? How could it resist reducing organisms to externally assembled machines without abandoning materialism? That project became organicism.

Organicism: a third way between machine and vital force

Organicism emerged not as one unified doctrine but as a family of attempts to take biological organisms seriously as an explanatory unit. Figures such as John Scott Haldane (1860–1936), William Emerson Ritter (1856–1944), E. S. Russell (1887–1954), Ludwig von Bertalanffy (1901–1972), Joseph Needham (1900–1995), J. H. Woodger (1894–1981), and later C. H. Waddington (1905–1975) differed greatly in their methods and metaphysics. What they shared was dissatisfaction with the forced choice between mechanism and vitalism.

As a pioneer in this narrative, John Scott Haldane used “organicism” to articulate a biology in which the activities of parts become intelligible only through their relations within the organism. Later, William Emerson Ritter’s The Unity of the Organism, published in 1919, similarly defended an “organismal conception” in which the whole is as necessary for explaining the elements as the elements are for explaining the whole. This was a subtle change. Reductionism asks: What are the parts, and how do they produce the whole? Organicism adds a second question: How does being part of this particular whole change what the parts are capable of doing? That difference matters.

A protein does not possess one universal function independent of context. Its activity depends on localization, chemical environment, binding partners, post-translational modifications, concentrations, cellular state, and developmental history. A cell removed from a tissue can change phenotype. A signal can have opposite effects in different contexts. Genes that are essential under one environmental regime can become irrelevant under another. The “part” is therefore not always a small machine with an intrinsic function waiting to be discovered. Its functional identity is relational. This does not mean that the whole “mystically” controls every microscopic event. It means that biological causation cannot always be reconstructed by assigning context-independent properties to pre-given components. In biology, there is no privilege level of causation.

The organismic tradition was, in this sense, already challenging what we might call substance metaphysics: the assumption that reality is fundamentally composed of entities whose identities are logically prior to their relations. Instead, organicists moved toward a relational, process-first picture. Relations can partly constitute the identity of what is related. This idea would eventually become central to fields such as relational biology, process ontology, dialectical biology, and contemporary theories of organizational closure. But before that maturation occurred, organicism found one of its most remarkable historical laboratories in interwar Cambridge.

In the 1930s, a remarkable group of scientists and philosophers gathered around what became known as the Theoretical Biology Club. Its members and associates included J. H. Woodger (1894–1981), Joseph Needham (1900–1995), Dorothy Needham (1896–1987), Dorothy Wrinch (1894–1976), J. D. Bernal (1901–1971), and C. H. Waddington (1905–1975), with a wider intellectual network intersecting with figures such as J. B. S. Haldane and Karl Popper. Erik Peterson’s historical reconstruction of this milieu is valuable precisely because it shows that organicism was not simply a vague reaction against molecular science. It was an attempt to build an intellectually serious alternative to both reductionistic mechanistic philosophy and vitalism.

All these thinkers were asking an important question that sounds surprisingly contemporary: What would a genuinely theoretical biology look like if biology did not merely borrow its explanatory categories from physics? The question was radical. As I analyzed in detail during the first part of this trilogy, after Newton, physics had become the model of scientific success. To understand something scientifically increasingly meant identifying elementary constituents, specifying laws governing their behavior, and deriving higher-level phenomena from those interactions. But perhaps biology required concepts that were not simply imported from the study of inert matter: Organization, Development, Differentiation, Regulation, Hierarchy, Historical contingency. These were not supernatural properties. They were biological properties demanding theoretical articulation.

Particularly, C. H. Waddington became one of the most influential inheritors of this project. His famous epigenetic landscape represented development not as the mechanical unfolding of a rigid genetic blueprint but as a structured process in which multiple interacting influences stabilize developmental trajectories. Genes mattered enormously, but they acted within regulatory and developmental systems. The “landscape” itself was a metaphor, of course. But notice what had changed. Machine metaphors imagine development as execution. The landscape metaphor imagines development as constraint-guided becoming. This conceptual shift anticipated several ideas that would later become central to dynamical systems biology: attractors, canalization, robustness, developmental plasticity, and the dependence of phenotype on networks rather than isolated instructions.

Organicism therefore did not merely say “the whole is more than the sum of its parts,” as modern complex systems science usually does. That slogan is too weak. The more radical claim was that the identity and behavior of parts can depend on the organizational context that the parts collectively generate. This is a circular relation, and circular causation would soon acquire a new scientific language.

Cybernetics: mechanism begins to turn back on itself

In the first part of this trilogy, we explore how cybernetics influenced the success of our current machine metaphor: the computer. However, the rise of cybernetics complicated any simple history of mechanism versus organicism. Norbert Wiener (1894–1964), W. Ross Ashby (1903–1972), Warren McCulloch (1898–1969), Walter Pitts (1923–1969), and other attendees of the Macy Conferences developed concepts such as feedback, control, communication, and regulation that transformed the scientific understanding of purposive behavior.

Yes, cybernetics was mechanistic, but it was not Cartesian mechanism in its simplest form. A thermostat provides the familiar example. Its behavior cannot be understood merely as a one-way chain from cause to effect. The system measures a variable, compares it (implicitly or explicitly) to a reference condition, acts on the environment, receives the consequences of that action, and adjusts again. In cybernetics, causality becomes circular, output returns as input, and behavior becomes regulation. This was a profound conceptual advance because it allowed scientists to model goal-like behavior without invoking an immaterial goal. A missile-guidance system can correct its trajectory. A thermostat can maintain temperature. An organism can regulate glucose. Purpose, cybernetics suggested, might be describable through feedback. The old dichotomy between mechanism and teleology began to weaken.

But cybernetics introduced a new ambiguity that still structures contemporary debates. Whose goal is being maintained? For a thermostat, the desired temperature is set by an external user. For a guided missile, the target is supplied by a designer or operator. For a living organism, the conditions relevant to continued existence arise from the organization of the system itself. This distinction between extrinsic goals and intrinsic norms is essential. Yes, cybernetics showed that machines could display goal-directed behavior. It did not automatically explain how a system could generate the conditions that make some states “good” or “bad” for itself. A thermostat does not care whether it survives. A bacterium does not require conscious concern, but there is an objective asymmetry between conditions compatible with its continued organization and conditions that destroy it.

This asymmetry—later discussed through concepts such as normativity, viability, adaptivity, and autonomy—is difficult to derive from the mere existence of feedback. Cybernetics therefore occupies an ambiguous place in our story. It expanded mechanism far enough to incorporate circular causation. And by doing so, it created the conceptual space from which some of the strongest alternatives to conventional machine metaphors would emerge. In this context, Humberto Maturana (1928–2021) and Francisco Varela (1946–2001) made one of the boldest moves in twentieth-century theoretical biology. Instead of asking primarily what organisms do, they asked what kind of organization makes a living system a self-maintaining unity in the first place. Their answer was autopoiesis, i.e., self-production.

An autopoietic system is not merely a network of reactions, but an organization of processes whose activity continuously generates and regenerates the components and boundary relations required for that organization to persist. The canonical example is cells. Their membrane is not a container manufactured elsewhere and subsequently filled with metabolism. Metabolic processes contribute to producing and maintaining membrane components. The membrane, in turn, helps create the physical conditions under which those metabolic processes can continue. Neither side is intelligible in isolation, boundary and metabolism mutually enable one another, but this is more than feedback. Feedback modifies the behavior of a pre-existing machine. Autopoiesis concerns the recurrent production of the machine-like unity itself.

Maturana and Varela explicitly characterized living systems in terms of autonomy, self-reference, and self-construction, moving the analysis away from a list of material ingredients toward a circular organization of production. The difference can be stated simply: a conventional machine has an organization, while an organism must continually do its organization. Stop a clock and the clock remains. Stop a cell sufficiently and the cell disappears as a cell; its components rapidly cease to participate in the organization that made them living components. This brings us to one of the deepest limitations of classical machine metaphors. They tend to treat organization as a noun, but life requires us to treat organization as a verb.

Rashevsky and Rosen: from components to relations

A parallel route toward the same problem developed through mathematical biology. Nicolas Rashevsky (1899–1972), inspired by thinkers such as D’Arcy Thompson (1860–1948), began his career attempting something entirely compatible with the mechanistic scientific ideal: applying mathematical physics to biological phenomena. But over time he became dissatisfied. The problem was not that physics was wrong. The problem was that describing biological components in ever greater physico-chemical detail did not automatically yield a general theory of biological organization. Rashevsky, therefore, proposed a shift from “metric biology” toward “relational biology.”

Instead of beginning with the material details of a particular organism, relational biology would ask what organizational relations remain invariant across different material realizations of life. Robert Rosen (1934–1998), Rashevsky’s student, pushed this project much further. Rosen’s famous (M, R)-systems—metabolism-repair systems, later interpreted more generally in terms of metabolism and replacement—attempted to formalize a distinctive feature of organisms: the efficient causes (explained in the first part of the trilogy) required to maintain the organization cannot all remain external to the system. In ordinary machines, critical efficient causes are external: the factory makes the engine, the engineer repairs it; the programmer supplies the software, the user defines its task. But organisms are different.

In biological organization, somehow, the network of transformations must participate in producing or replacing the very catalysts, constraints, and functional capacities required for those transformations to continue. Rosen called this closure to efficient causation. Modern discussions often summarize the idea by saying that the catalysts or constraints (i.e., efficient causes) required by the system must themselves be generated within the organization. This “closure” can be easily misunderstood. Closure here does not mean thermodynamic closure. On the contrary, organisms are radically open to matter and energy. They eat, respire, excrete, absorb photons, exchange molecules, depend on environments, etc.

The closure to efficient causation referred by Rosen is organizational, not material. What closes is a network of causal roles, not the resources from the environment. This distinction is crucial because it shifts the organism–machine debate away from superficial properties. Yeah, a robot might move, learn, repair damage, update software, and even build another robot. However, none of those capabilities alone establishes the relevant form of closure life possesses. The deeper question is: Where do the constraints that make these processes functionally integrated come from, and how are they regenerated? That question will return in Part III of the present trilogy.

For now, what matters is that Rosen’s work turned an organicist intuition into a formal challenge. Perhaps the fundamental difference between organism and machine is not complexity, material composition, intelligence, behavior, or even reproduction. Perhaps it is a difference in causal organization.

Pattee and the epistemic cut: information requires more than symbols

A different route into the same territory was developed by Howard Pattee (1926– ). As well discussed across the first part of this trilogy, the computational metaphor had made “information” seem almost weightless. A bit can be represented by voltage, magnetic orientation, light, ink, or molecular configuration. This apparent independence from material substrate is one reason information became such a powerful cross-disciplinary concept. But biological information poses a special problem.

A nucleotide sequence does not interpret itself. DNA does not spontaneously mean “construct this protein” merely because one sequence can be correlated with one amino-acid sequence. The relation between rate-independent molecular vehicles and rate-dependent physical processes requires machinery: transfer RNAs, aminoacyl-tRNA synthetases, ribosomes, etc. The genetic “code” is therefore not merely a string of symbols. It is a physical relation between symbols and dynamics sustained by an interpreting organization.

Pattee described this using the concept of an “epistemic cut”, the distinction between relatively rate-independent symbolic descriptions and the rate-dependent physical dynamics they constrain and control. In biological systems, the genotype–phenotype relation is not an immaterial dualism, but an epistemic necessity to better understand biology. The cut is then physically instantiated,as both symbols and dynamics belong to a monism. This imposes one of the strongest challenges to naïve computational metaphors.

Computer science begins after the interpreter already exists. A program is meaningful because an architecture exists that implements an instruction set. An instruction set is meaningful because hardware has been organized so that certain symbolic differences trigger certain physical consequences. A file becomes executable because an entire stack of conventions, compilers, operating systems, processors, and engineered causal mappings already exists. In biology, however, the existence of the interpreter is part of the phenomenon requiring explanation. The cell must produce the machinery that interprets genetic descriptions, and much of that machinery is itself produced through processes that depend on previously existing interpretation. Biological symbol systems are embedded in a self-maintaining physical organization.

Here the computer metaphor becomes strangely circular. To explain the cell, we say that DNA is software. But software functions only because hardware knows how to interpret it. Where does the biological hardware come from? From processes partly guided by DNA. Then what interprets the DNA? The hardware. Thus, the explanation oscillates unless we explain the organizational closure connecting description and construction. This does not invalidate computational biology. It shows that computation presupposes an implementation relation. And in living systems, that relation itself has a history and a material organization.

The molecular revolution and the temporary eclipse of the organism

The rise of molecular biology after World War II made many organicist concerns appear old-fashioned. The reason is understandable. Molecular biology worked spectacularly. The discovery of the structure of DNA, our understanding of gene regulation and protein synthesis, and other similar achievements gave biology unprecedented causal precision. After the molecular revolution, researchers could manipulate specific molecular components and observe reproducible consequences. The gene became an extraordinarily powerful experimental handle. However, the success of this research program created what might be called an epistemological asymmetry.

Problems that yielded to biological decomposition advanced rapidly. Problems involving development, organism–environment reciprocity, integrated physiological organization, and biological agency proved harder. Science naturally accumulated more knowledge where its tools worked best. As well noticed by Levins and Lewontin in The Dialectical Biologist, the enormous success of Cartesian reduction as a research tactic can tempt us to infer that nature itself must be organized according to the ontology presupposed by the tactic. This is a subtle selection effect in the history of knowledge.

Imagine possessing a very powerful microscope. You begin to discover extraordinary things, and eventually, because everything you successfully study is visible through the microscope, you may start believing that what cannot be seen through it is not scientifically real. The limitation here lies not in the microscope, but in forgetting that it is a microscope, one of the many tools you could use to study the world. Reductionism works similarly. The mistake is not decomposition. The mistake is ontologizing decomposition.

This distinction allows us to criticize reductionism without becoming anti-reductionist. Sometimes the best scientific strategy really is to isolate a component. Sometimes causal intervention requires simplifying context. Sometimes the molecular scale is exactly the relevant scale. The problem begins when methodological success becomes a universal metaphysics in which every higher-order regularity must be treated as merely shorthand for lower-level events. At that point, the organism begins disappearing from biology.

Canguilhem: life does not merely obey norms—it establishes them

Georges Canguilhem (1904–1995) approached the problem through medicine, physiology, and the philosophy of biological normativity. A machine can malfunction relative to a norm imposed by its design. A broken pump is broken because we know what the pump was constructed to do. Its “normal” state is externally specified. But what does “normal” mean for an organism?

Canguilhem argued that living beings are not simply passive realizations of externally imposed norms. They establish norms through their ways of living. A physiological value can be normal in one context and pathological in another. Health is not always the maintenance of one fixed numerical optimum, it includes the capacity to tolerate variation, establish new equilibria, and respond to altered conditions. This insight complicates machine metaphors profoundly. If biological normativity were exactly like engineering normativity, there would have to be a designer specification against which the organism is evaluated. But organisms are products of evolution, development, and ongoing interaction with their environments. Their “goals” cannot simply be read from an external blueprint.

This does not make biological norms mystical. It makes them relational. A blood-glucose concentration matters because of how it affects the viability of a particular organized system. A temperature becomes damaging relative to the capacities of that organism. A perturbation becomes a signal only when some system can discriminate it and respond. Meaning, function, and normativity emerge from relations between organization and conditions of existence. Canguilhem’s deeper inversion is therefore historical as much as philosophical. Machines do not provide the original model from which life should be understood. Machines are themselves products of living organisms! Before there was a machine, there was an organism capable of constructing one.

In this way, the organism is not merely a defective machine. The machine is a peculiar extension of organismic activity. This inversion will become crucial in Part III of the present trilogy, because it suggests that the relationship between machines and organisms is not one-way. Biology has borrowed metaphors from machines, but organisms have also inspired us to build entirely new kinds of machines.

Dialectical biology: part makes whole, whole makes part

In The Dialectical Biologist, Richard Levins (1930–2016) and Richard Lewontin (1929–2021) gave reductionism one of its most explicit critiques ever formulated. Their argument is often compressed into the cliché that “the whole is more than the sum of its parts.” But their position is considerably more interesting and deep.

Levins and Lewontin challenge the assumption that parts possess fully determinate identities before entering wholes. In a Cartesian picture, the basic units come first. They possess intrinsic properties, they interact, and the whole emerges from those interactions. Thus, causality flows primarily upward. However, what happens when we reverse the asymmetry? In biology parts certainly contribute to wholes, but wholes also change the properties of parts; and because the whole is transformed by those changed parts, the relation itself evolves. In biological explanation, what counts as a biologically relevant “part” can depend on the question being asked and on the organization in which that part participates. This is not mystical holism. It is contextual causation.

Consider a cell. Even inside such a tiny organization, its differentiation state, gene expression, mechanical properties, metabolism, and fate may depend on signals generated by surrounding tissue. Remove it from that tissue and many of those properties change. So, what exactly is the “intrinsic” property of an isolated cell? The question may be badly posed. The cell’s biological “identity” was partly constituted relationally. In their marvelous book, Levins and Lewontin extend the same reasoning to the interactions between organisms and their environment. Traditional adaptationist narratives often imagine an environment that exists independently, presents a problem, and selects organisms according to how well they solve it.

But the truth is that organisms also modify environments: beavers build dams, plants transform atmospheres and oceans, microbes alter chemical gradients, earthworms restructure soil, animals construct shelters, etc. Organisms change the selective conditions encountered by themselves and others. Environment causes organism, but organism causes environment. The two are historically coupled. The organism becomes, therefore, both subject and object of evolution. This destroys one of the most persistent assumptions inherited from machine metaphors, that one dictating that there must exist a stable separation between system and external world.

A conventional computer receives inputs from an environment whose relevant variables have already been defined by the designer, but organisms participate in defining what counts as relevant. For a bacterium, a chemical gradient can become an affordance for movement. For another organism, the same molecule may be irrelevant. An environment is not simply “everything outside” organisms. It is a world structured by the organism’s capacities, vulnerabilities, and possible actions. This idea will reappear later in ecological psychology, enactivism, niche construction, and theories of biological agency. The environment is not merely input. The organism is not merely output. Their relation is constructive.

From mechanism to computation: the decisive conceptual inflation

Nowadays, the most powerful modern machine metaphor is not the clock, nor even the engine, but the computer. Part I of this trilogy reconstructed how this happened through the convergence of several intellectual revolutions: formal computation, information theory, cybernetics, molecular genetics, and eventually Artificial Life.

But an important distinction must now be restored. Alan Turing (1912–1954) and Alonzo Church (1903–1995) did not begin by announcing that the universe is a computer. They were addressing precise formal questions about effective procedure and computability. Turing’s machine was an abstraction designed to capture what can be accomplished by a rule-following symbolic procedure. Claude Shannon (1916–2001), when developing information theory, likewise did not establish that meaning is made of bits. Shannon intentionally abstracted away from semantic content in order to quantify communication under noise. All these were acts of conceptual discipline. The later problem arose through conceptual expansion.

Computation gradually came to mean many different things: symbol manipulation, information processing, causal transformation, dynamical evolution, state transition, cognition, control, and eventually, in some formulations of pancomputationalism, any physical change whatsoever. The broader the concept became, the easier it was to claim universality. But universality purchased through definition can become scientifically empty. This is the first major fault line. Suppose I say that “my laptop computes.” The claim seems informative, since it is true that laptops implement architectures deliberately constructed to manipulate representational states according to specified operations. Now, what else can we say it “computes”?

Suppose I say that a bacterium computes. The claim is less obvious but potentially productive. Perhaps I can specify molecular states functioning as inputs, regulatory transformations, memory, conditional responses, and outputs. It is true a computational description may reveal experimentally meaningful organization. Now suppose I say that a hurricane computes. Similar to the bacterium case, I can construct a mapping between atmospheric states and a formal dynamical process, but under that general construction I can also claim that a stone computes, a glass of water computes, a decaying isotope computes, and even a planet computes its orbit. But then, if everything computes, what information has the word “computation” added? This is the triviality problem.

Philosophers of physical computation have long debated what physical conditions are required for a system genuinely to implement a computation rather than merely to be describable through some convenient mapping. The strongest versions of pancomputationalism risk attributing vast numbers of computations to essentially every sufficiently rich physical system. More restrictive accounts therefore introduce additional requirements involving causal structure, counterfactual relations, functional organization, representation, or mechanistic mapping. The dilemma is highly relevant when discussing machine metaphors. If we define computation loosely enough, pancomputationalism becomes almost trivially true.

If “everything changes state, therefore everything computes,” then “computation” becomes a synonym for physical evolution. But if physics has not been explained computationally, then it has merely been renamed. This is a circular argument. If we define computation more strictly, however, pancomputationalism becomes substantive. But then we must show that every relevant physical system genuinely satisfies the stricter criteria. The claim is no longer automatic. This is why the deepest criticism is not “rocks obviously cannot compute.” The criticism is methodological: What explanatory work is gained by calling a physical process computation?

In science, a good scientific concept discriminates. “Living” distinguishes some physical organizations from others. “Catalyst” distinguishes a causal role. “Natural selection” specifies a process. If “computation” includes everything that undergoes lawful change, it ceases to discriminate. Pancomputationalism then risks becoming true at the cost of becoming empty.

The map is not the territory

A related problem concerns the difference between modeling something computationally and establishing that it is intrinsically computational. This distinction is familiar everywhere else in science. We model planetary motion using differential equations, but nobody therefore concludes that planets solve differential equations before moving. We represent fluids using finite differences, but the ocean does not contain a hidden finite-difference solver. We model populations using matrices, but populations are not matrices. Formal systems are maps. The phenomena are territory. The success of the map can reveal real structure in the territory, but correspondence does not establish identity.

Computation is especially vulnerable to this confusion because universal Turing machines can represent so many other systems. A computer can simulate an epidemic, a galaxy, an economy, a chemical reaction, a neural network, the evolution of a population, you name it. The extraordinary generality of computation thus creates the illusion that computational representation reveals the ontological essence shared by all these phenomena, but perhaps the opposite lesson follows. A representational framework capable of mapping radically different phenomena may tell us as much about the generality of the framework as about the ontology of what is represented.

This was one of Robert Rosen’s central concerns with modeling relations. A formal system and a natural phenomenon are not identical merely because we establish a successful correspondence between them. The act of modeling itself creates a relation between physical processes and formal entailments. Confusing those sides collapses epistemology into ontology. This does not weaken the act of modeling, it simply clarifies what models are. A map can be extraordinarily accurate: it can predict, it can compress, and it can reveal hidden consequences. But no improvement in cartography, no matter how complex the map’s structure may be, transforms the map into the territory.

The map-territory distinction becomes particularly important in fields such as Artificial Life. A simulation of a hurricane does not make the room windy. A simulation of nuclear fusion does not produce nuclear radiation. A simulation of fire does not burn. These examples are deliberately obvious because they reveal a distinction that becomes less obvious when the target phenomenon is described functionally: some properties are substrate-sensitive. For instance, to realize combustion, specific energetic and chemical transformations matter. Computation, however, is multiply realizable; the same software, algorithm, or logical function can run on entirely different physical hardware.

What about life? Is it substrate-sensitive or multiply realizable? This is where the debate becomes difficult. Strong Artificial Life has sometimes defended the possibility that the right computational organization could literally instantiate life. The argument is not absurd. If being alive depends on organization rather than organic chemistry specifically, why assume that life cannot exist in another substrate? The problem is not substrate chauvinism, the problem is identifying which organization must be realized. If life requires a self-producing materially open organization, then simulating that organization is not necessarily the same as physically instantiating it. Simulations are not realizations.

A digital cell may contain variables representing nutrient flux, but the computer running the simulation is not materially maintained by those represented nutrients. A simulated membrane may be repaired by simulated molecules, but the silicon hardware does not depend on that simulated membrane for maintaining the causal organization enabling the simulation. The semantics of simulated variables are supplied from outside, by an external observer who interprets the simulation results and aligns them with physical reality. If a simulation runs without any impact on empirical reality or external observer understanding, it is a “pre-meaningful” process rather than an active physical cause.

Computability, predictability, and decidability are not the same thing

Many criticisms of computationalism become confused because they treat several distinct ideas as synonyms. A system can be deterministic without being predictively tractable, it can be computable without being efficiently computable, it can be simulable without allowing analytic shortcuts, and it can be unpredictable because of missing information, sensitivity to initial conditions, computational complexity, stochasticity, or formal undecidability. These are different limitations, and distinguishing them makes the critique of pancomputationalism stronger, not weaker.

Consider chaos. A chaotic system may be governed by completely deterministic equations. Its unpredictability arises because small uncertainties in initial conditions can amplify dramatically. That does not mean the dynamics are automatically uncomputable. It means long-term prediction can require impossibly precise knowledge. Undecidability is stronger. Cris Moore demonstrated that certain dynamical systems can encode universal computation so that questions about their long-term behavior inherit undecidability associated with the Halting Problem. Yet even here, later work shows that chaos, computational universality, noise, and predictability must not be collapsed into one phenomenon; in some formal settings, adding noise can destroy Turing completeness and restore computability of long-term statistical behavior.

This distinction matters enormously. It is tempting to say: “Chaos proves nature cannot be computed.” That is too strong. The defensible conclusion is subtler: “Even simple deterministic rules do not guarantee practical or universal predictability.” That already undermines one naïve inheritance from the Newtonian machine metaphor—the idea that complete laws plus initial conditions straightforwardly imply usable prediction. The machine may be deterministic, but the future can still remain inaccessible. Now we turn our attention to a related concept: undecidability. While both Gödel’s incompleteness theorems and Turing’s Halting Problem are often recruited as weapons against computational worldviews, they must be handled carefully.

Kurt Gödel (1906–1978) showed, roughly speaking, that sufficiently expressive consistent formal systems cannot prove every truth expressible within their arithmetic language. Alan Turing (1912–1954) showed that there is no general algorithm capable of deciding whether every arbitrary program will eventually halt. These results transformed our understanding of formal reasoning, but they do not straightforwardly prove that the physical universe is noncomputational. A computer can be computational even though there are questions about its behavior that cannot be decided by a universal algorithm. Indeed, undecidability arises precisely within computation. This produces an interesting reversal. The Halting Problem is not evidence that computation is weak in the colloquial sense, but it is a consequence of computation becoming sufficiently universal.

Therefore, the strongest anti-pancomputationalist argument cannot simply be: “Some questions are undecidable, therefore the universe is not a computer.” That inference fails. The more interesting conclusion is: “Even if the universe were computational, universal computation would not restore the Laplacian dream of complete prediction.” A computational universe, therefore, could still contain processes whose futures admit no general shortcut. This breaks an important historical link. Machine metaphors originally derived much of their prestige from predictability, but computation does not imply predictability. In fact, universal computation mathematically generates forms of irreducibility. The computer metaphor therefore partly dissolves the deterministic promise inherited from the Laplatian clockwork universe.

Rice’s theorem sharpens this point. Very roughly, any nontrivial semantic property of the function computed by an arbitrary program is undecidable in general. This means that there is no universal procedure that can inspect arbitrary source code and determine every meaningful behavioral property. Again, this does not prove that organisms transcend computation, but it challenges an extremely common intuition embedded in computational metaphors: “if we possess the program, we possess the explanation.” The above shows that this is not necessarily. Knowing the rules does not imply possessing a general shortcut to their consequences. This matters a lot in biology because “finding the code,” from molecular biology to neuroscience, has repeatedly been rhetorically associated with understanding the organism.

The Human Genome Project was sometimes culturally imagined as the acquisition of life’s source code. But even literal computer source code can resist global behavioral prediction. How much more cautious should we be when the supposed “program” is DNA embedded in a massively interactive, stochastic, developmental, ecological, and evolutionary system? The genome is invaluable, but sequence is not destiny, and a rule is not the same as its consequences.

Computational irreducibility and physical limits of computation

Stephen Wolfram (1959–) uses the phrase “computational irreducibility” to describe systems whose future behavior cannot generally be predicted by a shortcut substantially simpler than allowing the process itself to unfold. As we learned above, simple rules can generate behavior whose consequences are difficult or impossible to compress into closed-form prediction. But pancomputationalism sometimes responds to the limits of simulation by making a peculiar move. If no computer can efficiently simulate the universe, one can say: “The universe is computing itself.” This statement sounds profound, but we should pause.

If “the universe computes itself” means only that the universe undergoes whatever physical processes it undergoes, and then the claim risks becoming tautological. A falling stone becomes “the universe computing the stone’s fall.” A chemical reaction becomes “matter computing chemistry.” Evolution becomes “the biosphere computing its next state.” But what distinguishes this vocabulary from simply saying that physical processes occur? The computational metaphor may still be heuristically useful, but its explanatory content must be demonstrated, not assumed. When the only perfect simulation of a system is declared to be the system itself, the distinction between computation and physical happening begins to disappear. And once that distinction disappears, pancomputationalism risks winning by definition rather than discovery.

Physics also imposes limits on actual computation. Information storage requires physical resources. State transformations require finite energy and time. Quantum evolution has speed limits. Entropy constrains erasure. Bekenstein bound, Bremermann’s limit, and Landauer’s principle connect entropy, energy, and spatial confinement. The Margolus–Levitin theorem places a bound on the rate at which an isolated quantum system can evolve through distinguishable states as a function of available energy. These results matter. They remind us that real computers are physical objects. Turing machines are mathematical abstractions with idealized resources, but actual computation cannot be separated from thermodynamics, noise, finite memory, finite time, and embodiment. Still, we should avoid overstatement.

Physical limits on computers do not automatically prove the universe is not computational. A pancomputationalist can simply say that the universe computes subject to those physical limits. The stronger conclusion is different: any meaningful claim that the universe computes must specify the physical theory of implementation rather than treating computation as an immaterial abstraction. This brings us back to Pattee’s epistemic cut. Computation is never “just information.” It requires a physics of symbols. Pattee’s distinction between syntax and pragmatics can be traced back to quantum mechanics, which often enters debates about computationalism through measurement. Some arguments claim that wave-function collapse is instantaneous, stochastic, irreversible, and therefore fundamentally noncomputational. This conclusion, however, is premature.

The interpretation of quantum measurement remains philosophically contested. Different interpretations assign different ontological status to collapse, branching, hidden variables, decoherence, observers, and probability. One should therefore resist using a controversial interpretation of quantum mechanics as if it were an experimentally settled refutation of computation. Yet measurement still raises an important conceptual question. A computation requires a mapping between physical differences and formal states. A measurement requires a mapping between physical dynamics and recorded distinctions. Who or what establishes that mapping? In engineered computers, designers build it. In scientific experiments, investigators define it. In living systems, organisms evolved structures that selectively transform environmental differences into internally consequential states.

That is the biologically interesting property. A receptor does not measure “the world in general,” but it is selectively sensitive; its discrimination matters because of the organization in which that measurement participates. The problem of measurement therefore becomes a problem of relevance realization (also known as the symbol grounding problem; the frame problem; and the problem of meaning), and relevance cannot be extracted from Shannon information alone. One of the most persistent conceptual confusions in biology concerns the word “information.” Shannon information quantifies uncertainty reduction in communication. It is an extraordinarily powerful mathematical concept. But Shannon deliberately bracketed meaning. A random sequence can contain high Shannon information while meaning absolutely nothing to an organism.

Biological meaning requires context. A nucleotide sequence becomes functionally significant through relations with molecular machinery. A hormone means something different to cells with different receptors. The same environmental signal can be irrelevant to one species, attractive to another, toxic to a third. Meaning is not a free-floating substance encoded in matter. It is relational. This is why the statement “DNA contains information” can be either useful or misleading depending on what follows. If we mean that nucleotide sequences constrain the production of other molecular sequences through a historically evolved coding relation, the informational language is precise. If we mean that DNA contains a complete symbolic description of the organism in the way software contains a complete program for a deterministic machine, the metaphor breaks.

Living systems are intercausal processes

As we learned in the first part of this trilogy, the blueprint metaphor was extraordinarily productive because it suggested a research strategy. Find the instructions, map them, and determine how they generate function. But biology repeatedly complicated this picture. Philip Ball’s recent synthesis provides multiple counterexamples to the computational metaphor, arguing that the genome cannot reasonably be treated as a self-sufficient instruction manual for the organism. The contemporary picture is distributed across genes, regulatory networks, protein dynamics, cellular organization, tissue mechanics, development, and active biological agency.

Again, this does not mean “genes do not matter.” That pendulum swing would be equally foolish. The point defended here is dialectical. Genes matter because of the systems that interpret and regulate them. Those systems matter because they are partly constructed through gene expression. Neither side is causally sovereign. The old debate “genes versus environment” is therefore badly posed. Genes are part of the environment of other genes. Cells construct environments for genes. Organisms modify external environments. External conditions transform gene regulation. Development itself changes what future signals can do. Life is not one-directional causation. It is intercausal organization.

A critical distinction between life and machines is not their complexity, but life’s processual identity. A conventional artifact persists materially when inactive. Turn off a laptop and it remains a laptop. Stop a watch and it remains a watch. Leave a car in a garage and it may slowly degrade, but its organization is not continuously dependent on active self-production in the same immediate sense as a living cell. Biological entities are different; their apparent stability is an achievement. Molecules turn over. Membranes are repaired. Proteins degrade. Ions leak and are pumped. Cells die and are replaced. Tissues remodel. Metabolic reactions continuously sustain concentrations far from equilibrium.

Organisms, in fact, appear thing-like only because processes occur at different timescales. Consider a whirlpool. At first glance, it looks like a tangible entity, but its persistence consists in an organized flow of water. Stop the flow of water and the whirlpool disappears. An organism is vastly more complex, but the ontological lesson is similar. Life is not merely a structure through which processes happen. Life is a structure constituted by processes. This is why process philosophy has returned so strongly to contemporary biology. The shift from substance to process does not deny relatively stable entities, it also explains stability as a temporally maintained organization.

Bernd Roßlenbroich’s organismic program makes a related point from another direction: reductionistic findings need not be rejected, but they gain biological meaning within the integrated properties of organismic systems. This provides a more precise criticism than saying “machines are static.” Modern machines are not necessarily static; examples include smart thermostats, autonomous cars, and industrial robots. The deeper contrast is, however, whether the system’s continued identity depends on the ongoing endogenous regeneration of the organization that constitutes it. That is a much harder property to engineer, and no machine built by humans to date has that characteristic.

The language of constraints offers a promising way to make these organicist intuitions scientifically precise. A constraint does not necessarily push matter the way a force does; it restricts possibilities. A membrane constrains diffusion. An enzyme constrains which reaction pathways become kinetically accessible. A ribosome constrains molecular dynamics so that nucleotide sequences become translated into amino-acid sequences. An anatomical structure constrains possible movement. The key insight developed in contemporary organizational biology is that living systems do not merely exist under constraints; they construct constraints, and those constraints can participate in constructing one another.

Montévil and Mossio formalize biological organization in terms of closure of constraints, distinguishing the relatively fast processes occurring in open thermodynamic conditions from constraints that remain sufficiently stable over the relevant timescale to channel those processes. Biological organization arises when such constraints stand in mutually dependent relations of production and maintenance, constraining energy flows to maintain the constraints enabling those flows. Constraint enables work. Work maintains constraints. The circle closes. This is a naturalistic route toward what older philosophers called formal and final causation. No ghost enters the machine. The causal topology changes.

The distinction between matter and organization described above is helpful for clarifying a common misunderstanding. It is a fact that organisms are open thermodynamic systems. They depend continuously on environmental matter and energy. To be alive, living entities constantly take in nutrients and expel waste. How, then, can anyone speak of “biological closure”? Because biological entities can be simultaneously open and closed in different causal modes. A cell is materially open. Molecules cross boundaries. Energy flows. Waste leaves. But organizationally, the network determining how those flows are harnessed exhibits closure.

Rosen expressed organizational closure in terms of closure to efficient causation. Similarly, autopoiesis describes a network that recursively produces the components participating in its own realization. However, as mentioned above, modern organizational approaches reframe these and other theoretical related ideas through closure of constraints. Importantly, these frameworks are not mathematically identical. They should not be collapsed casually. But they converge on a profound conceptual point: autonomy is compatible with dependence. Indeed, biological autonomy requires dependence; an organism isolated from all environmental exchange dies.

Autonomy therefore cannot mean independence from the world or the external environment. Autonomy means a specific mode of dependence in which environmental affordances and thermodynamic flows are incorporated through internally organized relations that help maintain the system’s own conditions of existence. This gives biological autonomy an apparently paradoxical structure: open to matter, closed in organization. Dependent, yet self-determining. This is precisely the kind of relation that current machine metaphors struggle to represent, and in the third part of this trilogy, we will explore its operationalization and realization through relational models.

Evolution does not merely search a pre-given possibility space

Computational metaphors frequently imagine evolution as search. There is a fitness landscape, possible genotypes occupy points, variation explores the landscape, and selection moves populations toward peaks. This metaphor can be extremely useful, but it contains an assumption that becomes problematic when universalized: the space of possibilities already exists.

Optimization requires a domain. Search requires a search space. A computer program requires definable states. A dynamical system requires variables and a phase space. What if biological evolution changes the relevant variables themselves? Kauffman, Longo, and Montévil have argued that this is precisely what happens in open-ended evolution. New traits create new ecological possibilities. New organisms create new niches. New functions emerge for existing structures. Once flight evolves, “air as a navigable ecological domain” acquires new biological relevance. Once flowers and pollinators coevolve, new relational ecologies emerge. Once oxygenic photosynthesis transforms atmospheric chemistry, the space of possible metabolisms changes. The provocative claim is that we cannot prestage all future biological affordances in advance. The phase space itself evolves.

Longo, Montévil, and Kauffman therefore argue for “enablement” rather than “universal entailment” in biospheric evolution: organism–niche interactions continually alter the contexts within which new biological possibilities become meaningful. This proposal is controversial, as it contradicts the premise that prediction is the quintessential form of explanation. However, it should not be presented as a mathematical theorem proving biology “uncomputable,” but it identifies one of the deepest conceptual differences between conventional computation and open-ended evolution. A Turing machine operates over a pre-given formal architecture. Its alphabet is specified. Its transition rules are specified. The space of possible machine configurations is mathematically definable. Evolution appears to do something stranger. It generates novelties that alter what later novelties can mean.

The issue is not merely that we do not know which future state will occur. We may not know how to specify in advance the biologically relevant dimensions in which future states should be described. That is a fundamentally different kind of unpredictability and Kauffman often illustrates this problem using the possible uses of an object. Consider a screwdriver. It can turn screws, but also it can pry open a lid, it can prop a window, it can scratch a surface, it can serve as a paperweight, it can become part of a sculpture, it can block a door, etc. No list seems naturally complete. More importantly, the possible uses depend on contexts that may not yet exist. A screwdriver, and any other object, can acquire a future function through relations with technologies, organisms, practices, and environments not yet invented. This is the logic of an affordance.

Affordance, in the sense the term holds in ecological psychology, are not simply intrinsic properties of an object. They are relational possibilities for action. A branch affords perching for one organism, food for another, camouflage for another, building material for another, and irrelevance for countless others. Biological possibility is therefore relational, relative to a goal. And relations can be historically created, not structurally predefined. This challenges the idea that evolution can always be represented as exhaustive search over a fixed possibility space. The conclusion is not that computers cannot approximate evolutionary processes. The conclusion is that a model with a fixed ontology may fail precisely where evolution changes the ontology relevant to subsequent dynamics. The most interesting novelty is not a surprising answer, but a new question becoming possible.

Stephen Jay Gould’s famous thought experiment in Wonderful Life asks us to imagine rewinding the tape of evolutionary history and playing it again. Would the same forms evolve? Would humans return? Gould’s answer emphasized contingency: tiny historical differences could redirect later evolution. The thought experiment has sometimes been used to argue that life cannot be algorithmic, but that conclusion goes too far. A stochastic algorithm can generate highly contingent histories, a computational process can be path-dependent, and randomness can be modeled computationally. Therefore contingency alone does not refute computation. But Gould’s thought experiment remains important for another reason. It attacks the idea that biological explanation consists solely in deriving inevitable outcomes from initial conditions and timeless laws.

In biology, historical explanation matters because what exists now changes what can happen next. Evolution inherits, exapts, repurposes, and builds upon accidents. A feather that eventually participates in flight did not need to originate “for flying.” An old structure can acquire a new function. A new function can transform selection pressures. In organisms, history changes possibility. This is one place where machine and evolutionary metaphors diverge. Machines are usually understood from their intended design. but organisms must also be understood through what happened to their lineage. The present contains residues of paths not taken. The same logic undermines a simplistic picture of adaptation. In the conventional engineering analogy, the environment defines a problem and the organism evolves a solution. That framing is often useful, but it renders the organism too passive.

Organisms do not merely solve environmental problems, they change which problems exist. A beaver dam changes water flow, sedimentation, vegetation, predator access, temperature, and the ecological conditions encountered by future generations. Microbes change pH. Plants alter light environments. Animals redistribute nutrients. Earth’s atmosphere itself has been transformed by life. The organism–environment relation is reciprocal. This matters for computational metaphors because “input” and “output” cease to be stable categories. The organism changes its input distribution through its outputs. More deeply, life does not merely react within a fixed problem space, but in ways that reshape the conditions of future action. This is already a bridge toward the third part of the trilogy, but we are not ready to cross it yet. First, we must confront pancomputationalism at its strongest.

Pancomputationalism: theory, hypothesis, or metaphor?

The phrase “the universe is a computer” can mean several radically different things. One interpretation is methodological: any physical process can, in principle, be modeled computationally. Another is implementational: every physical system realizes some form of computation. Another is ontological: physical reality is fundamentally computational. Another is digital-physical: the universe literally evolves through discrete informational updates comparable to a cellular automaton or quantum computer. These positions are often blended rhetorically, but that creates an unstable argument.

When challenged ontologically, the pancomputationalist can retreat to methodology: “I only mean that everything can be described computationally.” When challenged methodologically, the stronger language returns: “But computation is what nature is doing.” The ambiguity makes the thesis difficult to falsify. Müller’s critique is useful precisely because he separates the two core possibilities: the world is a computer, or the world can be described as one. The latter can remain a powerful metaphorical and scientific framework without entailing the former. This distinction dissolves much unnecessary conflict. I have no objection to describing an organism computationally when computation reveals structure. My objection begins when a successful description is treated as an exhaustive identity.

The same principle applies to every metaphor discussed in this trilogy. The heart really does behave like a pump in important respects. It is not therefore nothing but a pump. The cell really does exhibit factory-like production. It is not therefore nothing but a factory. DNA really does participate in coding relations. It is not therefore a complete software program. Living systems really do perform computations. It does not follow that life is exhausted by computation. Metaphors become dangerous not when they are wrong. They become dangerous when their domain of validity becomes invisible.

Another problem concerns scale. Suppose some biological subsystems compute. Does the whole organism therefore constitute one computer? Not necessarily. Suppose every local physical interaction can be represented informationally. Does the universe therefore form one unified information-processing architecture? Again, not necessarily. Both examples instantiate a fallacy of composition. Properties do not automatically transfer from parts to wholes. Every human has a birth date, but humanity does not have one birth date in the same sense. Each neuron produces electrical activity, but the brain is not therefore merely one giant neuron. Scale introduces organization. This makes claims such as “the universe is a quantum computer” more complicated than they initially sound.

Even so, theoretical computer science contains models that exceed conventional Turing computation under unusual assumptions. These models are philosophically valuable because they test the boundaries of what “computation” means. One example is Malament–Hogarth spacetimes in general relativity. Malament–Hogarth spacetimes are theoretical spacetime geometries in which one worldline can contain an infinite amount of proper time while another observer reaches a relevant event in finite proper time, opening formal discussions of “supertask” computation.

Does this prove the universe is a hypercomputer? No. Does it refute pancomputationalism? Also no. These scenarios depend on particular physical and idealizing assumptions, and their realizability remains speculative. Their real philosophical value is different. They demonstrate that the Church–Turing notion of computability and the physically possible are conceptually distinct questions. Turing computability characterizes a formal class of procedures. Whether every physically realizable process falls within that class is an empirical and theoretical question about physics. One cannot answer it simply by redefining “physical process” as “computation.” That would assume what must be demonstrated. This fact was also recognized by Robert Rosen.

At this point, it is worth pausing. A critique becomes weaker when it collects every conceivable objection and treats them all as decisive. The arguments examined so far do not have equal force. The Halting Problem limits universal algorithmic prediction, but computational systems themselves exhibit undecidability. Physical bounds constrain actual information storage and processing but can be incorporated into physical theories of computation. Hypercomputational spacetime models are theoretically provocative but empirically speculative. These are important constraints on naïve computational triumphalism, but they are not knockdown refutations of every computational ontology. Together, the arguments reviewed here concern four deeper issues.

First, implementation: without principled criteria specifying what makes a physical system implement one computation rather than arbitrarily many mappings, pancomputationalism risks triviality. Second, model–ontology confusion: the ability to describe a system computationally does not establish that computation exhausts what the system is. Third, constitutive organization: living systems participate in producing the constraints, boundaries, and interpretive machinery through which their own processes acquire functional organization. Fourth, historical openness: biological evolution may transform the relevant state variables, affordances, and possibility spaces rather than merely traverse a fixed predefined domain. These challenges show that computation, by itself, may not specify the organization that makes a computation biological. That is the question worth keeping.

Organicism’s own failure: when “the whole” explains nothing

The history reconstructed here can easily become triumphalist in the opposite direction: mechanism was reductionist, organicism understood the truth, and modern biology should become holistic. Organicism, however, has often suffered from a severe methodological problem. It identifies genuine limitations of reductionism but does not always tell us what experiment to perform next. “The organism is an integrated whole.” Fine. How integrated? Through which processes? At what timescale? Which components are constitutively necessary? What observations would falsify the claim? What intervention distinguishes genuine organizational closure from mere causal interdependence? How do we measure autonomy? How do we compare two candidate organizations? When holism cannot answer such questions, it risks becoming descriptive vitalism.

The vital force disappears linguistically, but its explanatory role survives. Instead of saying “entelechy organizes the embryo,” we say “the whole organizes the parts.” But unless “whole” is operationalized, have we explained anything? This is precisely why I do not think the solution is to abandon mechanism; the solution is to transform it. Reductionism gave science something organicism frequently lacked: experimental traction. A future biology must preserve that virtue. Otherwise, its philosophical sophistication will be purchased at the cost of scientific usefulness. This is the central challenge for Part III in our trilogy. Still, the contemporary return of organismic thinking should not be mistaken for nostalgia. The more detailed our molecular knowledge becomes, the clearer it becomes that context is not noise around the mechanism.

We can now dissolve one of biology’s oldest false dichotomies. Neither “reduce the organism to molecular machinery,” or “treat the organism as an irreducible whole” is satisfactory. Pure reductionism loses constitutive relations. Pure holism loses experimental traction. A better framework must allow decomposition without assuming ontological separability. We can study a part while remembering that its causal role may depend on the organization. We can manipulate a gene without pretending the phenotype is encoded linearly. We can model a regulatory network computationally without declaring the cell a computer. We can identify molecular mechanisms while asking how those mechanisms are produced, maintained, and coordinated. We can use machine metaphors while refusing to let them become metaphysics by inertia.

This is a dialectical move. The opposite of reductionism is not vagueness. The opposite of mechanicism is not magic. The alternative is a relational mechanism. Mechanism in which relations can be constitutive. Mechanism in which constraints can be self-produced. Mechanism in which wholes can change the causal roles of parts without violating physical law. Mechanism in which history matters. Mechanism in which the variables relevant to future dynamics need not remain permanently fixed. Mechanism capable of explaining not merely how a system operates, but how it participates in producing the conditions under which its operation remains possible. The strongest challenge is not that organisms possess a magical capacity absent from all possible machines. It is that our dominant concept of machine has historically externalized crucial causal conditions that organisms internalize.

This reframes the question completely. Rather than asking: “Are organisms machines?” We should ask: “What kind of machine would an organism have to be? Can we build one?” That is a scientifically productive question. The failure of perfect modularity is exactly what one should expect from constitutively integrated processual systems. Biological parts are not always Lego bricks, their identities and behaviors depend on relations. This does not imply abandoning engineering. It implies that biology may force engineering to evolve. Just as the steam engine transformed thermodynamics and the computer transformed theories of information, attempts to construct living systems may transform our concept of machine. Synthetic biology could become the experimental arena in which organicism finally proves whether it has anything useful to offer. A philosophy of life that cannot help us build, manipulate, or experimentally distinguish forms of organization risks remaining philosophy alone. This is why the future belongs neither to naïve mechanical philosophy nor to romantic holism. It belongs to a synthesis.

What would such a synthesis require? First, it must preserve the achievements of mechanistic explanation; we still need decomposition, intervention, quantitative models, prediction where possible, and falsifiability. Second, it must stop assuming that all relevant causal organization comes from local interactions among pre-given parts; it must represent constraint production and history dependence. Third, it must distinguish computation from the organization that makes computation meaningful; measurements require selective couplings and biological information must be physically embedded. Fourth, it must treat organism–environment relations as reciprocal; organisms do not merely receive inputs, they alter the worlds from which future inputs arise. Fifth, it must accommodate changes not only in state but in the organization of state spaces themselves; open-ended evolution does not merely explore, it constructs. Part III will ask whether these requirements can be formalized. That is the crucial transition. Organicism has spent more than a century telling us what conventional mechanism misses. Now it must show us what to build instead.

Conclusion

The history of machine metaphors is often told as a conflict between two camps. On the hand, defenders of mechanism. On the other, defenders of organicism. One side decomposes, the other integrates. One side explains, the other contextualizes. One side builds machines, the other insists that life is different. I think this framing has outlived its usefulness. Part I of the present trilogy showed why machine metaphors conquered biology. They earned their authority: they made life experimentally tractable; they gave us physiology, molecular biology, genetics, computational biology, biotechnology, and synthetic biology. Their success is not an embarrassment to explain away. It is the thesis of this trilogy.

The present essay has reconstructed the antithesis. From Kant’s natural purposes to the Theoretical Biology Club, from Rashevsky-Rosen’s relational biology to Pattee’s epistemic cuts, from Canguilhem’s normativity to Levins-Lewontin’s dialectics, from process ontology to contemporary organizational closure, many alternative traditions have repeatedly insisted that something disappears when living systems are treated merely as externally organized collections of parts. But the lesson of this history is not that mechanism has failed, it is that mechanism is incomplete. Historically, machine metaphors progressively absorbed what earlier versions excluded.

Clocks captured order but not metabolism. Engines captured energy transformation but not control. Cybernetic machines captured regulation but not self-production. Computers can revise internal rules, but the question remains how the organization that gives those revisions meaning is constituted and maintained. Every new machine humans created expanded the metaphor. Every expansion revealed another boundary. Pancomputationalism represents the maximal expansion of the current metaphor. If every physical process is called computation, the machine finally encompasses everything. And precisely at that moment, the metaphor risks explaining nothing.

Living organization highlights multiple limitations of our current mechanistic ideal. The history reconstructed here suggests that the most productive response is to build better machines. Machines whose relevant constraints are not all externally fixed. Machines that participate in producing the organization that makes them operational. Machines whose measurement–control relations can be reconstructed rather than merely parameterized. Machines whose functions are grounded in their own conditions of continued existence. Machines in which hardware and software cease to be cleanly separable because description, interpretation, and construction participate in one recursive organization.

Perhaps calling such systems “machines” will eventually stretch the metaphor beyond recognition. Or perhaps they will constitute a new class of neomachines, devices that currently elude our conception of what a machine could be. Either possibility forces us to confront a problem that neither classical mechanism nor twentieth-century organicism solved completely. The mechanists gave us powerful tools but an impoverished ontology. The organicists recovered organization, context, process, and reciprocity but often lacked the formal and experimental machinery required to turn those insights into a constructive operational research program.

Neither side is sufficient. That is why the next step cannot simply be another critique. We now need a synthesis. A theory capable of representing functional closure. A physical account of constraints that explains how causes can be produced by the systems in which they operate. A framework that distinguishes externally programmed computation from systems that construct and revise their own measurement–control organization. After tracing the success and limits of machine metaphors, a new question naturally arises: what must machines become before the distinction stops being useful? That is where the final part of this trilogy must begin.