The word emergence has become suspiciously convenient. Consciousness emerges from neural activity, life from chemistry, markets from individual exchange, temperature from molecular motion. These claims may all be true. None of them explains the phenomenon it names.
In Emergence explains nothing and is bad science (October 2025), John Heil argued that the word masks ignorance and mistakes gaps in explanation for gaps in reality. His first complaint is easy to recognize. When someone asks how consciousness arises from matter and the answer is that consciousness is an emergent property of sufficiently complex matter, the explanandum has simply been repeated in different words. We wanted to know how the transition occurs; instead we were given a label for the fact that it occurs.
His second complaint is the harder one, and it is usually the one that goes unanswered. Philosophers disagree about causation too, Heil notes, but there the disagreement is anchored: everyone can point to uncontroversial cases, and competing theories are theories of the same thing. Emergence has no such anchor. Competing accounts of it look like accounts of entirely different phenomena. A word whose instances have nothing in common is not a concept; it is a homonym with good press.
The right response is to concede that charge and split the word. There are four situations routinely called emergence, and they differ in which of three questions gets a negative answer. Is the macrostate physically determined by the microstate? Is it computationally accessible from the microdescription? And does the microdescription provide the best explanation of the macrobehavior?
When we simply do not yet know how the higher-level behavior follows, and no principled barrier stands in the way, the emergence is epistemic. It is a fact about us, and Heil’s diagnosis of it is correct. When the microdynamics fix the outcome completely but afford no feasible shortcut to it, determination holds and accessibility fails; call this computational emergence. When both hold and the higher-level description still explains better, because coarse-graining has revealed regularities insensitive to large classes of microscopic variation, the emergence is structural. And when determination itself is denied, so that something remains unaccounted for after the physical state and laws are fully specified, the emergence is strong.
The first three remain compatible with ordinary physical explanation. The fourth adds something to the ontology. Heil is right about the fourth, and right about the first: epistemic emergence needs no separate treatment, since once the gap is merely ignorance, ordinary inquiry either closes it or does not. The disunity he complains of is real, and these are different phenomena wearing one word. What does not follow is that the word should be abandoned, or that higher-level explanation is a symptom of ignorance.
Determination Is Not Computational Accessibility
Suppose the complete physical state of a gas fixes everything that happens to it. Every collision, trajectory, and fluctuation follows from the underlying dynamics. Pressure, temperature, viscosity, and shock waves require no additional substance or force. In that sense the system is ontologically reducible: nothing at the macroscopic level is left physically undetermined by the microscopic state and laws.
But complete determination does not imply that every macroscopic fact is cheaply available from the microscopic description. A simple dynamical rule can generate behavior whose later states cannot be reached by any route substantially shorter than running the dynamics. Cellular automata make the point cleanly. Their update rules fit on a few lines, yet predicting the state after a billion steps can require computing something close to those billion intervening steps.
The obstacle takes several forms. Some systems are difficult to predict because the required computation grows too quickly. Chaotic systems introduce a different problem: small uncertainty in initial conditions can be amplified until long-range prediction becomes impossible in practice. Computationally universal systems impose a stronger limit, since some questions about eventual behavior are formally undecidable.
Neural networks show the same structure outside physics. Their elementary operations are fully specified, yet knowing them does not let us predict the representations training will produce without running something close to the training. Evolutionary systems too: variation, inheritance, and selection are describable by local rules, and those rules do not make future adaptations available in closed form.
All of these undermine the same naive inference. Knowing that lower-level rules determine an outcome does not supply a feasible shortcut to it. The system remains lawful and its future remains generated by lower-level dynamics; what fails is computational accessibility.
Being fixed by a process is one thing; being able to bypass the process is another. A complete microscopic description may contain all the relevant information while providing no tractable route to the particular higher-level fact we want. Nothing here requires new ontology. What has to go is the assumption that a lower-level specification gives us computational possession of the higher-level outcome.
Computational Accessibility Is Not Explanation
Even when a macrostate can be calculated from microphysics, the calculation may still be a poor explanation. Consider pressure. A gas pushing against the wall of a container can be described as an astronomical collection of molecular collisions, and nothing prevents that description from being physically complete. But if we want to understand how the gas responds when volume changes, pressure and temperature are the useful variables.
Those variables do not represent additional forces layered onto the molecules. They are abstractions over vast classes of microscopic states. Many different molecular configurations correspond to effectively the same pressure, so the macrovariable discards almost all microscopic information while preserving what matters for the behavior under investigation.
Thermodynamics survived statistical mechanics for this reason. Discovering that temperature is constituted by microscopic motion did not make temperature scientifically obsolete. It explained why a compact higher-level description could succeed across wide microscopic variation. The same pattern recurs throughout science. Biology does not become meaningless once organisms are known to consist of molecules, and computer science survives the reduction of software to transistor states. A lower-level description can be more fundamental while a higher-level description is more explanatory, because composition and explanation answer different questions.
Almost any system can be summarized by throwing information away, so coarse-graining alone is not enough. A scientifically useful emergent variable must preserve stable structure across many microscopic realizations and support approximately autonomous dynamics that can be modeled without continually reconstructing the discarded microdetail.
A better definition is therefore:
An emergent variable is a coarse-grained variable whose behavior exhibits stable, approximately autonomous regularities across many distinct microscopic realizations.
Approximately autonomous does not mean causally independent of the underlying physics, and it should not be glossed as merely convenient. The higher-level variables support a dynamical description whose predictions remain accurate under the interventions the domain admits, without reconstruction of the microstate at each step. Fix the class of interventions and the question becomes checkable: either the coarse-graining survives them or it does not. Convenience is not what is being tested.
This is why macro-level causal talk needs no second channel of causation running downward into the microphysics. If the underlying physical dynamics are complete, every macro-level change is implemented through lower-level interactions. But macro-level causal claims remain legitimate when they capture stable dependencies under intervention. Saying that raising a thermostat setting causes a furnace to turn on does not posit a new force operating above electrical and molecular dynamics; it identifies a higher-level variable whose manipulation reliably changes another. “Downward causation” is harmless if it means causal description across scales. It becomes a substantive metaphysical claim only if it means new causal powers not contained in the underlying dynamics.
Something close to this is Mark Bedau’s weak emergence, on which a macrostate is emergent when it is derivable from the microdynamics only by stepping through them. That captures the computational case exactly and leaves the structural one out, since renormalization is a shortcut. The two branches need separate treatment.
The Whole Can Be More Compressible Than the Parts
A gas containing roughly (10^{23}) molecules occupies a microscopic state space beyond enumeration, yet much of its macroscopic behavior can be described with a handful of variables: pressure, volume, temperature, and entropy. What is remarkable is not the appearance of some spooky property once enough molecules gather together, but that a system this large admits such a radical reduction in descriptive dimensionality while preserving reliable dynamics.
The usual slogan that “the whole is greater than the sum of its parts” obscures what is happening. The whole need not contain anything beyond its parts and their interactions. What it can be is more compressible than the sum of its parts.
Compression alone, however, is too cheap. Any dataset can be compressed relative to some choice of representation, and almost any complicated object can be summarized if we are willing to discard enough information. Scientific emergence requires compression that preserves dynamics under intervention.
Many physically distinct microstates must map onto the same macrostate, and those microstates must behave alike under the interventions that define the domain. Temperature is useful because countless molecular configurations instantiate one temperature and obey the same thermodynamic regularities. An algorithm can be implemented by different instruction sets, circuits, or physical substrates while its computational organization is preserved. This is multiple realizability, and it is often why the higher-level description has scientific value: the emergent variable identifies what remains invariant while implementation details vary.
If macrovariables are abstractions chosen by modelers, does emergence exist only in the eye of the observer? Could we invent arbitrary coarse-grainings and declare whatever patterns they reveal to be emergent?
We can invent them, but the world determines which ones work.
Dennett’s real patterns make the first half of this case: a pattern’s reality turns on whether the data admit a description more efficient than the bit map, not on whether anyone finds it interesting. The standing objection to that account is that compressibility is defined relative to a chooser’s purposes, which reintroduces exactly the observer-relativity it was meant to dispel. The intervention criterion closes the gap. A macrovariable earns its status when the information discarded by the coarse-graining stays predictively irrelevant under the interventions the domain admits, and that is a fact about the system rather than about the modeler. Arbitrary classifications rarely survive it.
Temperature is an observer-defined variable in the mundane sense that humans invented the concept and notation. Whether systems grouped by temperature obey stable thermodynamic relations is not a matter of convention. The center of mass is likewise a constructed coordinate, but whether its motion satisfies a compact dynamical law is an objective feature of the system.
The coordinate system is chosen. The invariance it reveals is not.
This gives the compression idea a stronger form. Scientific emergence is compression constrained by dynamics: a coarse-graining succeeds because the underlying world permits large amounts of microscopic information to be discarded without destroying the regularities we want to predict.
Universality Shows Why Microscopic Detail Can Become Irrelevant
Critical phenomena show that higher-level regularities can be objectively insensitive to microscopic detail. Near a continuous phase transition, systems made from very different microscopic constituents exhibit the same scaling behavior and critical exponents.
Renormalization-group theory explains why. Under repeated coarse-graining, many microscopic differences become progressively less important, while a much smaller set of large-scale parameters controls the observable behavior. Systems with different microphysics converge on the same effective description and fall into the same universality class. Batterman’s reading of this is that the renormalization group does not merely describe the insensitivity but demonstrates it, by showing which perturbations die out under rescaling.
Saying that critical behavior “emerges” explains nothing; showing why microscopic differences wash out under renormalization explains a great deal. Nothing violates microphysical law and no additional causal substance appears, yet the higher-level structure is not a reflection of our ignorance. Computational cost is a property of the process, and dynamical invariance is a property of the system. Calling either of them merely epistemological understates what has been established.
Heil’s own second example makes the point better than critical phenomena do. He observes that the particles of the Standard Model appear to have arisen from the radiation field of the early universe, while their properties and governing laws are not extractable from what we know of that earlier state, and he treats the non-extractability as suspicious. But this is the situation effective field theory was built to describe. Ultraviolet detail decouples: low-energy behavior is controlled by a small number of effective operators, and physics at higher energies enters only through traces suppressed by powers of the scale separation. The failure of extraction is not a gap in the world. It is what a decoupled description looks like from below.
That reply concedes something real. Decoupling explains why low-energy physics is insensitive to high-energy detail; it does not derive the Standard Model, whose particle content and parameter values remain inputs. Heil’s inference does not survive the concession: non-extractability is what a decoupled effective description looks like from below, and it is evidence about scales rather than about ontology.
The deductive characterization Heil reports treats failure to extract the higher-level description from lower-level knowledge as the mark of emergence, and often that failure does conceal an explanatory gap. But the scientific achievement can consist in discovering a transformation under which most lower-level information becomes irrelevant. The missing step may be a scale transformation, a coarse-graining procedure, a computational process, or the right collective variables. Sometimes the science consists in finding out what can be discarded.
Where Emergence Becomes an Excuse
The fourth sense is the one Heil is really after. In its strong metaphysical form, emergence claims that even after the underlying physical state and laws are completely specified, some additional fact or causal power remains unaccounted for by them. This is no longer a matter of computational difficulty, coarse-graining, or explanatory convenience. It is an ontological addition.
Strong emergence may be true, and nothing here shows it impossible. But calling a property strongly emergent does not explain it.
The usual route to it runs through the supposition that two physically identical universes might differ in whether anyone in them is conscious. I have argued in Hard to Explain Is Not Free to Vary that this supposition smuggles in a causally inert extra bit rather than discovering one. Grant it anyway. Introducing the variable creates a new explanatory burden, and it belongs to the account that posits the addition rather than to the physics it is added to: what states instantiate it, what determines its character, whether it enters causal dynamics, and what laws connect it to the rest of physics. If consciousness is genuinely fundamental, those questions replace the demand for reduction; they do not disappear.
Physics already contains fundamental entities and quantities that are not explained by reduction to something deeper. Their scientific legitimacy does not come from the word fundamental. It comes from being embedded in theories that specify mathematical relations, dynamics, constraints, and empirical consequences. Strong emergence should be held to the same standard.
Heil sets consciousness aside as a distraction, which is defensible in a piece about mechanics and unfortunate here, because consciousness is where the four senses are hardest to keep apart. “Consciousness emerges from sufficiently complex neural computation” sounds explanatory while leaving the central questions untouched. What organization matters? What physical or computational invariant corresponds to a particular conscious state? Why does one organization correspond to one structure of experience rather than another? The word cannot occupy the place where the theory should be. But declaring consciousness irreducible solves nothing either. If irreducibility is true, it tells us what kind of explanation we need, not that explanation has become unnecessary.
Heil’s first example makes the same point in a less loaded form. Saying that life emerges from chemistry is not an explanation. A satisfactory account must show how chemical systems acquire replication, compartmentalization, metabolism, inheritance, error correction, and eventually Darwinian evolution. Once those mechanisms are understood, describing life as an emergent regime of chemistry is appropriate. The emergence is the phenomenon explained.
Science often begins by naming stable phenomena before their mechanisms are understood. The mistake is treating the name as the conclusion.
Postscript
Heil is attacking a genuine pathology. Scientists and philosophers invoke emergence where the causal account becomes difficult, and there the word functions as an intellectual stop sign.
A diagnostic rule follows:
If emergence is being offered as the answer to “how?”, the explanation is incomplete.
But the disunity Heil identifies is a reason to sort the word’s uses rather than discard it. Computational and structural emergence survive the sorting. Complex systems frequently admit higher-level descriptions that discard nearly all microscopic detail while preserving dynamics that hold up under intervention, and those descriptions can be multiply realizable, insensitive to large classes of implementation detail, and far more tractable than the microstates that instantiate them.
The ontology can therefore remain fully reductionist without making prediction or explanation single-level. Fundamental physics constrains everything that can happen, and it does not follow that the vocabulary of fundamental physics is the vocabulary in which everything is most efficiently predicted or best understood.
Explaining why higher-level variables exist, why they are robust, and how they arise is the scientific work. Calling them emergent does not perform that work; it identifies the phenomenon the work must explain.
Emergence is the result.