Rewind the tape of life to the Cambrian, run it forward again, and what comes out?
Stephen Jay Gould’s answer was: something unrecognizable. He argued that the early history of animals was dominated by contingency, that small perturbations at the start would compound into radically different outcomes, and that nothing about the current cast of characters was inevitable. Simon Conway Morris, working on the same fossils, drew the opposite conclusion. He pointed at the enormous catalog of features that evolved independently in unrelated lineages and argued that the destinations are limited even when the routes are many, that selection keeps finding the same small set of good answers, and that something like us was always going to show up eventually.
That argument is thirty years old, both principals have been criticized, and it remains the framing question underneath every comparative claim in neuroscience. Because the convergent evolution of intelligence is either the strongest evidence available that minds are a solution physics forces, or it is a catalog of superficial resemblances that has been asked to carry far more weight than it can hold. Which of those it is depends on distinctions the popular version of the argument almost never makes, and getting them right is the difference between a comparative science and a list of coincidences.
What convergence has to mean to be evidence of anything
Start with the taxonomy, because the entire inferential value of a convergence claim depends on which category it falls into and the categories get blurred constantly.
Homology is similarity from shared ancestry. Your forelimb and a bat’s wing have the same bones in the same arrangement because the common ancestor had them. That tells you about descent and nothing about the problem being solved.
Convergence proper is independent origin of similar features in lineages whose common ancestor lacked the feature. The camera eye in vertebrates and cephalopods is the standard case, and it is a strong one because the last common ancestor had at most a light-sensitive patch and the two lineages built lens, iris, and retina separately, arriving at different retinal orientations that betray the independence.
Parallelism sits between them, where lineages independently evolve similar features starting from similar ancestral states and often using the same underlying genetic machinery. Whether parallelism counts as evidence of a forced solution is exactly what people argue about, and the distinction between convergent evolution and parallel evolution has been drawn in enough incompatible ways that some researchers have argued the terms should be abandoned in favour of describing the actual starting states.
Deep homology is the category that ruins most casual claims. Distantly related lineages frequently build convergent structures using the same conserved regulatory genes, because those genes were sitting in the ancestral toolkit doing something related. Pax6 is involved in eye development across an enormous range of animals with structurally unrelated eyes. That does not make the eyes homologous, and it does not make the convergence fake, but it does mean the two lineages were not starting from a blank sheet.
The philosophers have pushed back on the deflationary use of this. The argument that shared regulatory genes undermine independence only works if those genes are causally responsible for the features on which the convergence judgment rests, and for eyes they are not: the specific morphology that makes a camera eye a camera eye is not specified by the shared upstream gene. Deep homology means the lineages had similar tools available. It does not mean the outcome was determined by the tools, and collapsing those two claims is the most common error in this literature.
The philosophical work on this is sharper than the biological summaries. The argument that convergent evolution can function as a natural experiment for the contingency question makes the point that proponents of convergence have lumped causally heterogeneous phenomena into one basket, treating every convergent event as equivalent evidence. It is not. A convergence produced by a genuinely open developmental space is strong evidence that the solution is forced. A convergence produced by two lineages having no other option available because of shared developmental constraint is evidence of the constraint, which is a different claim.
The strongest convergent evolution cases, and why they hold
Some convergences survive that scrutiny. The ones that do share a feature: the lineages are far apart, the starting materials are genuinely different, and the functional demand is specific.
Camera eyes are the durable example precisely because the anatomies betray independent construction. Vertebrate retinas are inverted, with photoreceptors pointing away from incoming light and the wiring in front of them, which produces a blind spot where the optic nerve exits. Cephalopod retinas are the sensible way around, photoreceptors facing the light, no blind spot. Two lineages solved the same optical problem and made opposite errors doing it, which is what independence looks like. The count is higher than two, since camera eyes have arisen separately in some cnidarians and compound eyes independently in several arthropod and annelid lineages, and estimates of how many times eyes of some kind have evolved run into the dozens.
Echolocation in bats and toothed whales is the second, and the confirmed distribution of the trait across mammalian lineages shows it appearing separately in laryngeal-echolocating bats, in tongue-clicking fruit bats, and in odontocetes. Different anatomy, same physics problem, same computational solution. The bat side alone contains an internal convergence, since laryngeal echolocation and tongue-click echolocation in fruit bats appear to be separate acquisitions within a single order, which means the trait arose more than once even among animals that were already flying nocturnal insectivores with the same ears.
Neural architecture supplies the cases that matter most here. The expansion-then-convergence design, where a modest input fans out onto an enormous population of small cells and then converges onto few outputs, appears in the vertebrate cerebellum, in the insect mushroom body, in the octopus vertical lobe, and in the avian pallium. Four lineages with no shared ancestral structure, one architecture, because pattern separation is a problem with a good solution. The functional requirement is specific: take inputs that overlap heavily, project them into a much higher-dimensional space where they no longer overlap, and then read out from that space with modifiable connections. Any system that needs to learn associations between similar things will benefit from that arrangement, and four lineages found it without comparing notes.
Heading representation is the same story. Vertebrate head direction cells and the insect central complex both encode orientation as activity moving around a ring, arrived at independently, in structures with no anatomical correspondence. The machinery of spatial navigation appears to have a small number of viable implementations.
And the 2025 developmental work on bird and mammal pallium is the sharpest recent case, because it established that neurons occupying equivalent positions in functionally comparable circuits are generated at different developmental times from different progenitor regions. Same circuit, different construction, demonstrated at the resolution of individual cell types. That result is the template for what a modern convergence claim should look like: not a resemblance noticed at low magnification, but a developmental trajectory traced cell by cell and found to differ where a homology claim would require it to match.
The prestin story, and how a convergence claim gets audited
The molecular level is where convergence claims have been tested most rigorously, and the resulting correction is the best worked example available of how this field actually works.
Prestin is the motor protein in cochlear outer hair cells that amplifies and tunes hearing. In 2008 and 2010, researchers reported that prestin sequences from echolocating bats and toothed whales cluster together phylogenetically, despite the animals being nowhere near each other on the species tree, because of parallel amino acid substitutions. Later work identified the specific parallel sites and demonstrated functionally that the substituted variants alter the protein’s electrophysiological properties in ways correlated with high-frequency hearing. That is convergence with a mechanism attached, and it holds up.
Then came the overreach. A 2013 genome-wide analysis reported that convergence in echolocating mammals was not restricted to a handful of loci but widespread across the genome, which was covered as a major result about how pervasive molecular convergence is.
It did not survive. Subsequent analyses compared the rate of convergent amino acid substitutions in echolocating mammals against non-echolocating control outgroups and found the genome-wide frequency to be comparable, meaning the apparent signal was the background rate of convergence that any two lineages produce by chance across a whole genome. There is no genome-wide protein sequence convergence specific to echolocation.
What survived the audit is more interesting than what was claimed. A functional enrichment test asking not how much convergence there is but where it sits evaluated more than four thousand tissue-affecting gene sets and found that convergent substitutions in echolocators are most significantly overrepresented in the set of genes regulating development of the cochlear ganglion, affecting eighteen genes. The total amount of convergence is unremarkable. Its distribution is not.
That sequence is worth holding onto as a template. A striking claim, a control comparison that dissolves it, and a reformulated question that recovers a real and narrower result. Anyone citing molecular convergence as evidence for anything should be able to say which of those three stages their example is at.
A related nuance has arrived more recently and it complicates the picture usefully. Comparative work across toothed whale species has found that the same hearing genes carry different signatures of selection depending on habitat, with accelerated evolution in coastal and riverine lineages relative to deep-ocean ones. The convergence, in other words, is not a single event producing a single answer. It is an ongoing process being tuned differently by different acoustic environments, which means treating echolocation as one trait that evolved twice understates how much variation sits inside each instance.
The experiments you can actually run
Gould called replaying the tape a thought experiment we cannot possibly perform, and that has stopped being true, which is the largest change in this argument since he framed it.
Experimental evolution runs the replay directly. Populations of microbes founded from a single ancestor, propagated in parallel under identical conditions for tens of thousands of generations, provide replicate tapes with the initial conditions controlled. The results are mixed in an informative way: many adaptations recur across replicates, sometimes down to the same gene and occasionally the same mutation, while some innovations appear in one lineage and never in the others despite identical selection. Parallel outcomes are common. They are not universal.
The pattern that emerged from those studies is worth stating because it maps onto the comparative data. Replicates founded from the same ancestor converge more often than replicates founded from different ancestors, and the extent to which convergence involves the same gene is greater when the taxa are closely related. That is the same relationship visible across animals, where bat and whale echolocation shares more molecular detail than vertebrate and cephalopod eyes do, and for the same reason: shared starting material narrows the search space before the search begins.
The limitation is that microbial populations replay a shallow tape. They test whether a given genome under a given pressure finds the same solution, which is a real question and not the one Gould was asking. Nobody can replay the Cambrian, and the deep replay remains a thought experiment. Comparative anatomy across living lineages is the nearest substitute, which is why the distribution of nervous system architectures across the animal tree carries the weight it does. What the laboratory work establishes is that shallow replays are substantially repeatable, which is consistent with both camps and decisive for neither.
Convergence in things that are not alive
The strongest argument that certain solutions are forced comes from a source nobody planned as evidence, which is that engineered systems keep landing on the same designs.
Artificial neural networks trained on navigation tasks develop grid-like representations in their hidden layers without anybody building them in. Networks trained on image classification develop early layers that look like edge detectors, which is what the first stages of biological visual systems do. Systems trained to localize sound develop something functionally like coincidence detection. None of these systems has an evolutionary history, a metabolism, or a body, and they converge on solutions biology found first.
That is a different kind of evidence than comparative anatomy provides, and it is worth being careful about what it shows. It does not show that biological and artificial systems work the same way, and the differences are substantial. What it shows is that the solution is being forced by the structure of the task rather than by anything about neurons, which is exactly the claim convergence arguments are trying to establish and which cannot be established from biology alone because every biological example shares chemistry, energetics, and a planet.
Engineering supplies the same argument from another direction. Sonar was invented by people who did not know bats existed, and arrived at pulse-echo ranging with frequency modulation for the same physical reasons. Cameras have lenses. The convergence between what engineers build under a constraint and what evolution builds under the same constraint is the cleanest available demonstration that the constraint is doing the work.
The systems being built to reason without any biological substrate at all are therefore a live test of the whole framework, and the features they converge on with biological minds are the best current candidates for being forced rather than inherited.
Where convergence stops
The failures are the part that makes the successes meaningful, and Gould’s side of the argument has better evidence than the popular version admits.
Language is the obvious case. Vocal production learning has evolved several times, in songbirds, parrots, hummingbirds, cetaceans, bats, elephants, and humans, which is genuine convergence on a capacity. Syntax has evolved once. There is no second lineage with recursive grammar, despite hundreds of millions of years and enormous numbers of social animals with sophisticated communication. Some of the components have converged: bonobo call combinations show compositionality, the honeybee dance is genuinely symbolic, and vocal labeling for individuals has appeared separately in elephants, dolphins, parrots, and marmosets. The assembly of those components into open-ended recursive structure has not.
Cumulative culture is similar. Social learning is everywhere. Traditions are widespread, in chimpanzee tool repertoires, in cetacean foraging techniques, in bird dialects, and now in bumblebees acquiring behaviors beyond individual innovation. Open-ended ratcheting accumulation has happened once. The demographic argument that ape culture is capped by population structure rather than by ape minds is worth taking seriously here, because if it holds then the single occurrence reflects an ecological accident rather than a cognitive threshold, and the convergence question changes shape entirely.
Body plans are the deepest case. The major animal body plans were established in the Cambrian and essentially nothing new has appeared since. That is not selection finding the same answers repeatedly; it is developmental entrenchment, where the regulatory networks specifying gross morphology became so deeply interconnected that changes to them stopped being survivable. Gould’s actual argument applied specifically to that early window, and he acknowledged the entrenchment that followed, which is a nuance the popular version of the debate loses entirely.
The distinction that emerges is between shallow and deep replays. Rerun the tape from a hundred million years ago and you probably get eyes, wings, echolocation, and something with a big forebrain, because the developmental architecture was already fixed and the selective problems recur. Rerun it from the Cambrian and the entrenched body plans themselves are up for grabs, and there is much less reason to expect anything familiar.
That distinction dissolves a large part of the argument. Conway Morris assembles shallow replays and concludes evolution is predictable. Gould was arguing about the deep replay and conceded entrenchment afterward. Both are substantially correct about the depth they were addressing, and the appearance of disagreement comes from a shared vocabulary applied to different questions.
The two-experiment argument
The most useful thing convergence does for comparative neuroscience is provide independent replicates, and the value depends entirely on how independent they actually are.
A single instance of anything cannot distinguish necessity from accident. If nervous systems had evolved once and every animal inherited the same design, there would be no way to tell which features are forced by the problem of being an animal and which are frozen accidents from one ancestor. Every claim about what cognition requires would be a claim about a sample of one.
Convergence breaks that. Where two lineages arrive at the same solution from different starting material, the shared features are candidates for being forced and the divergent features are candidates for being contingent. Birds and mammals both built dopamine-rich association areas downstream of sensory convergence; that looks forced. Mammals layered their pallium and birds did not; that looks contingent.
The inferential move is worth stating precisely because it is the engine of the whole subject. Shared features across independent origins are candidates for necessity. Divergent features across independent origins are candidates for accident. Neither classification is secure from a single comparison, which is why the value of the method scales with the number of genuinely independent replicates, and why the group that built complex cognition on a body plan with no vertebrate correspondence is worth more to this argument than any number of additional mammals.
The independent origin question at the base of the animal tree is the extreme version. If ctenophores built neurons separately from everything else, there are two experiments in neural organization running on this planet, and comparing them would be the single most informative dataset in the subject. If neurons evolved once, the sample is one, and every generalization about nervous systems is a generalization about a lineage.
The honest caveat is that independence is a matter of degree rather than a binary. Bats and whales are both mammals with the same inner ear, the same cortical plan, and the same developmental toolkit, which makes their echolocation convergence considerably less independent than the vertebrate and cephalopod eyes. The deeper the shared ancestry, the more of the answer was already determined before either lineage started.
What actually constrains the solutions
If convergence is real, something must be doing the constraining, and the constraints sort into three kinds worth keeping separate.
Physics sets the outer boundary. Only certain quantities are detectable at biological scales with biological materials, which is why nothing detects radio waves and why electroreception is a water capability. Optics dictates that a lens focuses light, so any animal wanting a high-resolution image is going to build something lens-shaped. Acoustics dictates that resolution scales with frequency, so any animal wanting fine acoustic detail is going to end up in the ultrasonic. These are not evolutionary tendencies; they are boundary conditions.
Chemistry restricts the toolkit. There are only so many ways to build a transducer, and biology uses three: photopigments that change shape when hit, ion channels that open under deformation or heat, and receptors that bind molecules. The same protein families keep appearing in unrelated animals because the number of workable options is small.
Computation restricts the algorithms. Pattern separation, path integration, coincidence detection for timing, and predictive extrapolation for dealing with conduction delay are problems with a limited number of good solutions, and a nervous system that needs to solve them will land on one of them regardless of the tissue it is made of.
Development supplies a fourth constraint that is easy to overlook and cuts the other way. An organism cannot evolve any structure a selective pressure might favour; it can only evolve structures reachable by modifying an existing developmental program. That restricts the options in a way that produces convergence for reasons having nothing to do with the problem being solved, since two lineages with similar developmental architecture will find similar solutions because those are the ones available. Distinguishing convergence-from-constraint from convergence-from-optimality is the hardest problem in the field and it is frequently not attempted.
Underneath all three is the constraint that governs everything in this subject: energy. Neural tissue is the most expensive tissue an animal can run, which means every design is an optimization under a hard budget, and optimization under identical constraints produces similar answers. That is why the budget framing rather than the capability framing keeps turning out to be the productive one across the whole comparative literature. That is the mechanism behind convergence, stated plainly. It is not that evolution is aiming at anything. It is that the search space has a small number of low-cost regions and enough independent searches will find them.
Why convergence is easy to claim and hard to establish
The methodological problems are severe enough that a large fraction of published convergence claims would not survive proper testing, and knowing the failure modes is most of the skill in reading this literature.
Phylogenetic uncertainty comes first. Whether a trait is convergent depends entirely on the tree, and if the tree is wrong the classification flips. The ctenophore-versus-sponge argument is the clearest case, where the same anatomical facts imply either one origin of neurons or two depending on a branching order that took fifteen years and a switch to chromosome-scale synteny to resolve, and that is still contested.
Superficial resemblance is the second and it is rampant. Two structures can look alike at low resolution and be doing entirely different things, and comparative neuroanatomy has a long history of naming a region in one animal after a region in another and then reasoning as though the shared name were a finding. The century-long misnaming of the avian forebrain as striatum is the standard cautionary tale.
Absence of a proper control is the third, and it is what the prestin correction exposed. Without asking how much similarity two lineages would show by chance, any similarity looks like evidence.
Publication incentive is the fourth. Convergence findings are more publishable than the absence of convergence, which means the literature systematically overrepresents them, and nobody writes a paper reporting that two lineages solved a problem differently. Divergence under identical pressure is arguably more informative about the shape of the search space than convergence is, and it is close to unpublishable.
And selective attention is the fifth. It is trivially easy to construct a list of convergences and conclude that convergence is pervasive, because the list is assembled by looking for convergences. The denominator, meaning all the traits that did not converge, is never counted.
What convergent evolution licenses about minds
Bringing it back to nervous systems, the convergent evolution of intelligence supports a specific and limited set of conclusions, and it is worth being precise about which.
It supports the claim that certain computational solutions are forced, and this is the part with the most weight behind it. Pattern separation architecture, heading representation, camera eyes, and active sensing keep reappearing in unrelated lineages, and the most economical explanation is that the problems have few good answers.
It supports the claim that no particular anatomy is required. Layered cortex is not necessary for abstract rule representation, since birds do it without one. A centralized brain is not necessary for associative learning, since cnidarians manage it on a nerve net. Neurons are not necessary for memory, since organisms without them regulate and rebound. Every anatomical prerequisite that has been proposed for a cognitive capacity has eventually been falsified by an animal lacking the anatomy and possessing the capacity. That is a strong enough pattern to function as a prior: when somebody proposes that structure X is required for capacity Y, the historical base rate says an animal will turn up without X and with Y.
It does not support the claim that human-like intelligence is inevitable. Syntax happened once. Cumulative culture happened once. Those are the features that actually distinguish human cognition, and they are the ones with no independent replicate anywhere in the tree.
And it does not settle the contingency argument. Convergence at the level of eyes and echolocation is compatible with radical contingency at the level of body plans, because those are different depths of replay. Gould and Conway Morris were, to a substantial extent, talking about different questions and treating the answers as though they competed.
The reasonable position is that evolution is highly repeatable at the level of solutions to well-defined physical problems and much less repeatable at the level of which lineages exist to solve them. Eyes are inevitable. Vertebrates were not.
Which has a consequence for the question everyone actually cares about, which is whether minds like ours are likely anywhere else. Convergent evolution of intelligence supports the expectation that any planet with animals will produce sensing, learning, memory, and navigation, because those are solutions to problems any mobile organism has. It supports nothing about language, cumulative technology, or civilization, because those have a sample size of one on the only planet anyone has checked, and a sample of one cannot distinguish inevitable from fluke.
Where this leaves the comparison
The practical value of all this is a rule for reading any claim that two animals do the same thing.
Ask how far apart the lineages are, because the deeper the shared ancestry the less the convergence establishes. Ask whether the shared machinery was in the ancestral toolkit, because deep homology means the search was not free. Ask what the control is, because similarity has a background rate. Ask whether the resemblance survives increased resolution, because most do not. And ask what did not converge, because a list of hits without the misses is not evidence about a rate. Those five questions are most of what separates a convergence claim that constrains a theory from one that decorates it. Those five questions are most of what separates a convergence claim that constrains a theory from one that decorates it.
Applied consistently, that filter thins the catalog considerably and leaves a residue that is genuinely load-bearing: a small number of computational and optical and acoustic solutions that keep being found, by animals with nothing in common but a problem.
The 24-lecture Neurozoology course works the tree of life on that basis throughout, alongside the study of how knowledge moves between animals, the first edition’s survey of nervous systems, and the working animals whose capacities were discovered by people who needed something from them. The elephants and cetaceans that arrived at long-horizon social memory separately and the fish running individual recognition on a brain weighing a fraction of a gram are the kind of case the filter passes.
The tape has been replayed, repeatedly, in parallel, on the same planet. Every lineage that independently built an eye or a sonar or a memory system was running the experiment, and the results are sitting in the anatomy waiting for somebody to ask the question with a control group attached.
The eye evolved dozens of times. Syntax evolved once. Any account of minds that cannot explain both numbers is not finished.
Leave a Reply