There is no such thing as a fish.
That sounds like a provocation and it is a taxonomic fact. The category contains ray-finned fishes, lobe-finned fishes, sharks and rays, lampreys, and hagfish, spanning something over half of all vertebrate species and a range of evolutionary distance far exceeding the gap between a mouse and an ostrich. A tuna is more closely related to you than it is to a shark. A coelacanth is more closely related to you than it is to a salmon. Grouping all of them as fish is like grouping everything that is not a beetle and calling it a non-beetle: the category is defined by what it excludes rather than by what it contains. Any general claim about fish cognition is therefore a claim about more evolutionary diversity than any comparable statement about mammals or birds could possibly cover.
That matters for fish cognition because a paraphyletic grab-bag encourages a single answer to a question that has no single answer. And the single answer the culture settled on is that fish are simple, forgettable, and probably insensate, a belief that survives on three legs: fish do not have faces we read, they do not make sounds we hear, and they live in a medium we cannot enter without equipment. None of those are claims about the animal.
The last twenty years have made that position expensive to hold. Fish have been shown to use tools, recognize individual humans, cooperate across species, transmit local traditions, count, plan, and pass a mark test that most mammals fail. Whether any of it means what the popular coverage says it means is a separate question, and this is a field where both the enthusiasm and the skepticism are unusually loud.
Fish cognition and the three-second memory that never existed
Start with the audit, because the goldfish myth is doing more work than any actual finding.
The claim that goldfish have a three-second memory has no identifiable source in the scientific literature. It appears to be pure folklore, propagated because it was useful: an animal that forgets everything cannot be bored in a small bowl, cannot suffer meaningfully, and does not require thinking about.
The evidence runs entirely the other way. Goldfish form associations that persist for months. They learn to navigate mazes and retain the solution. They can be trained to press a lever for food and, when the lever is made functional only during a specific hour, learn to show up during that hour. Carp caught and released in a fishery become markedly harder to catch again, and the avoidance persists for a year or more, which is a memory of a specific aversive event with a duration measured in seasons.
Other fish do considerably better. Salmon imprint on the chemical signature of their natal stream as juveniles and use that memory to return years later across ocean distances. Cleaner wrasse maintain individual client relationships across hundreds of partners, tracking which clients have already been serviced that day and which are still waiting. Frillfin gobies memorize the topography of a tide pool at high tide and, when stranded, jump accurately between pools they cannot see into, using a spatial map acquired hours earlier. That is a stored spatial representation being consulted without the animal being able to check it against current perception, which is the operational signature the cognitive map literature treats as the demanding case.
The interesting question is not whether the myth is false but why it was so durable. It persisted because nobody had an incentive to check and because the alternative is inconvenient. Roughly a hundred billion farmed fish are killed annually, plus a trillion or so wild-caught, and there is no other vertebrate group where the number is that large and the welfare regulation that thin.
The myth also has a structural cousin worth naming, which is the assumption that a fish in a tank is a fish in its habitat. Much early work on fish behavior was done in bare aquaria on animals with nothing to do, and the resulting picture of a listless, undifferentiated animal was a description of the housing. Enriched environments change fish behavior substantially, improve learning performance, and alter brain gene expression, which means a substantial fraction of the older literature was measuring deprivation.
The pain argument, presented fairly
This is the most contested question in fish cognition and it deserves both sides at full strength, because a great deal of the popular coverage presents one of them as settled.
The case for fish pain runs as follows. Fish possess nociceptors, first demonstrated in rainbow trout, including polymodal receptors on the head and face responsive to mechanical, thermal, and chemical stimuli, some of them more sensitive than comparable human skin receptors. Trout injected in the lips with acetic acid show elevated opercular beat rate, rocking behavior while resting, rubbing the affected area against surfaces, and cessation of feeding, with a time lag between the stimulus and the behavior that argues against pure reflex. Administering analgesics reduces those behaviors. Fish will pay a cost, entering a normally aversive environment, to obtain pain relief. They show conditioned place avoidance for locations where noxious events occurred, which requires the experience to be aversive rather than merely detected.
The case against is not an argument that fish do not respond. It is an argument about what responding means. The skeptical position, argued most prominently by Brian Key and James Rose, holds that nociception and pain are different things: nociception is detection and reflex, pain is a felt experience, and the second requires neural architecture fish do not have. The specific claim, developed at length in the argument that fish do not feel pain and what that implies for phenomenal consciousness, is that phenomenal consciousness in vertebrates depends on identifiable properties of neural organization that mammals and birds possess and fish do not, and that inferring subjective suffering from behavioral and physiological responses is the same error as inferring it from a withdrawal reflex in a decerebrate preparation.
The counter-argument from the other camp is that the neocortical requirement is an assumption rather than a finding, and that it was already falsified by birds performing cortex-grade cognition in a forebrain with no layers at all. If a crow can hold an abstract rule without a neocortex, the inference from missing neocortex to missing experience does not go through.
Where this actually sits: nociception in fish is established and not disputed by anyone. Behavioral responses consistent with an aversive state are well documented. Whether there is something it is like to be the fish having them is not resolved, cannot currently be measured, and depends on a theory of consciousness the field does not have. The 2024 New York Declaration on Animal Consciousness placed all vertebrates including fishes in the category of realistic possibility rather than strong support, and noted that where a realistic possibility exists, ignoring it in decisions affecting the animal is irresponsible. That is the defensible position, and it is neither camp’s preferred headline.
Worth being explicit about the asymmetry in what the two errors cost. If fish do not have experiences and we act as though they do, the cost is economic and procedural. If fish do have experiences and we act as though they do not, the cost is roughly a trillion animals a year. That asymmetry does not settle the scientific question and it does bear on what to do while the question stays open, which is the distinction the declaration was written to make.
The mirror, and the fight it started
In 2019 a research group reported that bluestreak cleaner wrasse passed the mark test, and the resulting argument has been more productive than the result itself.
The finding: wrasse given a mirror progressed through the standard sequence, from apparent social response to atypical contingency-testing behavior to self-directed action. Marked with a colored spot on the throat, visible only via reflection, they scraped the marked area against surfaces, and did so in the presence of a mirror and not in its absence. A follow-up examining the capacity with ecologically relevant marks found the effect held with a larger sample, seventeen of eighteen fish passing, run by an independent generation of students, and established that the marks eliciting scraping resembled ectoparasites, which is precisely what a cleaner fish is professionally interested in. Later studies reported that wrasse recognize their own faces in photographs, and that they check their body size in a mirror before deciding whether to attack a rival.
The objections came fast and some are good. Frans de Waal argued the spontaneous behaviors were ambiguous and the marks physically irritating, meaning the scraping could be a response to sensation rather than to seeing. Gordon Gallup, who devised the test, argued the fish may be interpreting the reflection as another individual carrying a parasite and attempting to inspect it. Others noted that a test designed around a visually-guided primate hand may not transfer.
The reply from the original group is the part worth extracting, and it turns the whole thing into an argument about methodology rather than about fish. The marks were chosen to be ecologically meaningful precisely because an arbitrary dye spot is meaningless to an animal that does not groom for aesthetics. Green and blue marks produced no scraping; brown ones did. That specificity is hard to explain as irritation. And the deeper point is that if a test only counts when it is administered in a form the subject has no reason to care about, the test is measuring motivation rather than capacity, which is the failure mode that ran through decades of ape false-belief results and through experiments cats declined to complete.
A 2025 follow-up reported faster mark-directed responses than earlier work and documented wrasse using bits of food to test the mirror’s contingency, manipulating the reflection to check whether it tracked them. Contingency testing is the behavior that precedes self-recognition in the species that pass.
What nobody in the argument disputes is that something specific is happening in front of the mirror that does not happen otherwise. The dispute is entirely about what the fish is representing, and the honest reading is that the mark test was never a clean instrument. It has produced negative results in gorillas that are now attributed to eye-contact aversion rather than to absent capacity, a widely cited magpie positive that failed replication, and passes in species nobody expected. A test that behaves that way is measuring several things at once, and reading it as a binary about self-awareness was always more than it could bear.
Tool use, cooperation, and the reef as a workplace
Fish do things the classical definitions were not written to accommodate.
Several wrasse species carry bivalves to a chosen rock and strike them against it repeatedly to break them open, returning to the same anvil. Under the strict definition the fish is arguably not a tool user, since the rock stays put and the clam does the moving, which is a ruling that says more about the definition than about the animal. Archerfish shoot jets of water at insects above the surface, adjusting for refraction at the air-water boundary, compensating for prey distance, and hitting moving targets, which is a ballistics problem solved without hands.
The cooperation work is where fish cognition gets genuinely hard to dismiss. Groupers and coral trout hunt cooperatively with moray eels and octopuses, and the coordination is not incidental. The grouper performs a headstand signal over a crevice where prey has hidden, recruiting a moray that can enter spaces the grouper cannot, and the two hunt as a unit with each taking prey it could not otherwise reach. In controlled tests, coral trout chose the more effective of two collaborators after limited experience, and performed comparably to chimpanzees on a task requiring partner selection. That is interspecies cooperative hunting with partner choice, in a fish, and the reef systems where this has been documented in detail keep producing behavior nobody had a category for.
Cleaner wrasse run something closer to a service economy. They remove ectoparasites from client fish at established cleaning stations, and clients are of two kinds: residents with no alternative and visitors that can go elsewhere. Wrasse prioritize visitors, because a resident cannot leave. They cheat by taking client mucus, which they prefer to parasites, and clients respond by fleeing or by chasing. Wrasse behave better when observed by bystanders, which is reputation management, and clients watch cleaning interactions before choosing a station, which means the audience effect is responding to a real market. Pairs of wrasse that cheat get punished by their partners. The behavioral repertoire is the same one that shows up in the primate cooperation literature, in an animal with a brain that weighs a fraction of a gram. Wrasse also perform better on a task requiring them to choose the plate that will be removed first, prioritizing the ephemeral option over the permanent one, than apes and capuchins tested on the equivalent problem, which is the kind of result that gets explained away rather than absorbed.
The archerfish case deserves one more line because the physics is underrated. Refraction at the air-water interface displaces the apparent position of an aerial target by an amount that varies with viewing angle, so a fish aiming from directly below faces no correction and one aiming obliquely faces a large one. Archerfish compensate across a wide range of angles, adjust jet volume and velocity for target distance and size, and juveniles improve with practice and by watching others shoot. That is a learned ballistic solution to an optical problem, and the tool-use definitions written for primates do not have a category for a projectile made of water.
The lateral line, and a sense with no terrestrial version
The sensory equipment deserves treatment because it is genuinely alien and because it explains a great deal of behavior that looks impossible.
The lateral line is a system of mechanoreceptive organs, neuromasts, distributed along the flanks and head, some exposed on the skin and some recessed in fluid-filled canals. Each contains hair cells structurally similar to those in the vertebrate inner ear, and they respond to water movement relative to the body. What the system delivers is a continuous readout of the flow field around the animal, which means a fish detects the wake of a passing object, the pressure disturbance of an approaching predator, and the reflection of its own movement off nearby surfaces.
That last capability is the important one. A fish swimming near a wall generates a flow pattern altered by the wall’s presence, which allows a form of hydrodynamic imaging: blind cave fish navigate complex environments and can characterize the shape and distance of obstacles from flow alone. It is the closest thing in biology to touching at a distance, and it has no analogue in any terrestrial sense. Air is too thin to carry the equivalent information, which is why the system exists in fish and amphibian larvae and disappears in the lineages that left the water. Any account of what it is like to be a fish has to accommodate a channel that reports the shape of the surrounding water continuously, in every direction at once, with no attention required.
The lateral line is also what makes schooling possible. A school of thousands turns as a unit with latencies faster than visual reaction time would permit, and the coordination runs through each fish tracking the flow signature of its immediate neighbors. Disable the lateral line and schooling degrades badly while vision remains intact. A school is therefore a distributed sensor as much as a defensive formation, since a disturbance detected by any individual propagates through the group as a wave of hydrodynamic information, which makes the collective-behavior framing more literal here than in most social animals.
Add to this: electroreception in sharks and rays through the ampullae of Lorenzini, sensitive enough to detect a buried flatfish by its bioelectric field; magnetic sensing implicated in salmon homing; and in some lineages the active electrolocation that runs an entirely separate perceptual channel, covered as its own case in the physics of animals that generate their own probe signal. The umwelt of a fish is assembled from channels that mostly do not exist on land.
Brains without a cortex, doing cortex work
The neuroanatomy is where the skeptical argument lives, and it repays attention rather than assertion.
Fish have a pallium, the forebrain region that in mammals develops into cortex, but the developmental process differs fundamentally. Mammalian cortex forms by evagination, with the neural tube walls folding outward. The ray-finned fish pallium forms by eversion, folding the other way, which means the topology is inverted relative to the mammalian arrangement and the correspondence between regions is genuinely hard to establish.
Despite that, functional homologies are reasonably well supported. The lateral pallium appears to serve hippocampal functions, with lesions producing spatial learning deficits comparable to hippocampal damage in mammals. The medial pallium appears to serve amygdala-like functions in emotional learning and avoidance. Fish show conditioned fear responses, stress hormone systems homologous to ours, and behavioral responses to anxiolytic drugs that parallel mammalian effects. Zebrafish became one of the standard vertebrate models in neuroscience partly for that reason and partly for practical ones, since the larvae are transparent, which allows whole-brain imaging at cellular resolution in a behaving vertebrate. A substantial fraction of what is known about vertebrate neural circuits generally has been worked out in a fish, which sits oddly beside the assumption that fish brains are too simple to support anything interesting.
Relative brain size varies enormously across fishes, and some cartilaginous fishes have brain-to-body ratios comparable to birds and mammals. Manta rays have the largest brains of any fish and have shown behavior in front of mirrors that has been interpreted as contingency checking, though the sample is tiny. Cartilaginous fishes generally have been understudied relative to their brain investment, and the comparison of neuron counts and cortical allocation across species has almost no fish data in it, which is a gap rather than a finding.
The structural argument that matters is the one the independent construction of executive function in bird forebrains already made: functional equivalence does not require anatomical homology, and demanding a mammalian structure before granting a mammalian capacity is a bet on a specific theory of how brains work rather than a finding. Fish are the group where that bet is currently being cashed, and the cephalopods running comparable capacities on a nervous system with no vertebrate correspondence at all are the extreme version of the same point. Zebrafish also turn out to run two-stage sleep with signatures analogous to slow-wave and rapid eye movement states, in a brain with no cortex, which is one more capacity the architecture was supposed to preclude.
Culture, tradition, and what moves between fish
Social learning in fish is well documented and it produces exactly what the definition of culture requires.
Guppies learn escape routes and foraging routes from conspecifics, and the learned route persists in the population after the original demonstrators are removed, which is transmission rather than individual learning. French grunts follow traditional migration paths between resting and feeding sites; transplant naive individuals into a population and they acquire the local route; remove the residents and the route vanishes. Bluehead wrasse mating sites persist across generations and are maintained by tradition rather than by any property of the site, which was demonstrated by removing entire populations and finding the new occupants established different sites that then persisted in turn.
That last result is the strongest culture demonstration in fishes, because it rules out ecological determinism directly. The site was not special. The knowledge that the site was the site was what persisted.
Migration adds the largest-scale version. Many species run traditional routes between spawning, feeding, and overwintering grounds, and in several cases the route appears to be maintained by naive individuals following experienced ones rather than by any inherited program. Where that is true, the population’s spatial knowledge is stored in the animals rather than in their genes, with all the fragility that implies. It is the same structure as the matriarch holding a family’s map of water sources, differing only in that nobody thinks of a herring as a knowledge repository.
Fish also learn socially about predators, acquire food preferences from others, and in some species show conformity, matching the majority behavior even against private information. The song traditions in birds and the foraging traditions in cetaceans get called culture without controversy. The same phenomena in fish get called behavior, which is a vocabulary difference rather than an empirical one.
The conservation implication is sharp and underappreciated. If migration routes and spawning sites are culturally transmitted, then a population reduced below the point where knowledgeable individuals persist loses the information permanently, and the habitat can recover without the behavior returning. That is a plausible component of why some collapsed fish stocks have not recovered their historical distributions despite decades of reduced fishing pressure.
Numbers, faces, and the individual recognition problem
The cognitive test results are worth listing because their existence is the argument.
Fish discriminate quantities, distinguishing larger from smaller groups with a ratio-dependent accuracy profile matching that of human infants and other vertebrates, and some species do it on small numbers with near-perfect accuracy. Archerfish learn to discriminate human faces, distinguishing a familiar face from dozens of novel ones and continuing to do so when the images are converted to grayscale and standardized for shape. That is a fish performing a task once considered to require specialized primate face machinery, in an animal with no evolutionary reason to care about human faces at all, which is the detail that makes it a capacity result rather than an adaptation result.
Cleaner wrasse pass transitive inference tests, inferring that if A beats B and B beats C then A beats C, which requires representing relations rather than memorizing pairs. Medaka show a face inversion effect, the same signature of configural processing found in humans and chimpanzees, which is the marker that distinguishes specialized face processing from general pattern recognition and which had been considered a hallmark of animals with substantial visual cortex.
Individual recognition is the underlying capacity and it is more demanding than it sounds. A cleaner wrasse maintaining differentiated relationships with over a hundred clients has to recognize each, remember its history, and act accordingly, which is a social memory load comparable to a primate group. Damselfish recognize individual conspecifics by facial ultraviolet patterns invisible to us and, apparently, to their predators, which is a private signaling channel operating in a band outside the human sensory range entirely.
Personality, stress, and the individual fish
The assumption that fish within a species are interchangeable has failed as thoroughly as the assumption that they are simple, and it failed for the same reason: nobody had checked.
Fish show consistent individual differences in behavior that persist across contexts and over time, which is the working definition of personality in behavioral ecology. The best-characterized axis is boldness, with individuals reliably falling along a spectrum from bold to shy in how quickly they explore novel environments, approach unfamiliar objects, and resume feeding after a disturbance. Those differences correlate with metabolic rate, growth, stress hormone profiles, and survival, and they are heritable in several species.
Stress physiology is homologous to ours in the parts that matter. Fish run a hypothalamic-pituitary-interrenal axis functionally equivalent to the mammalian adrenal system, releasing cortisol in response to stressors, with chronic elevation producing immune suppression, reduced growth, and reproductive impairment. Anxiolytic drugs developed for humans produce the expected behavioral changes in zebrafish, which is why zebrafish became a standard model for screening them.
The more contested findings concern positive states. Fish show behavioral fever, voluntarily moving to warmer water when infected, which requires the animal to act on an internal state rather than a stimulus. Some species show what has been characterized as emotional fever, a small rise in body temperature after a stressful handling event, which had been considered a marker restricted to amniotes. And cleaner wrasse interactions appear to reduce client stress hormone levels, which is a tactile interaction producing a measurable physiological benefit.
None of this settles the consciousness question and all of it constrains it. An animal with individual temperament, a homologous stress axis, drug responses that match ours, and state-dependent behavioral thermoregulation is not a reflex machine, whatever else it is or is not.
The claims that do not hold up
An audit, running in both directions since this field has committed both errors.
The three-second goldfish memory is false, has no source, and served a purpose. Variants putting it at seven seconds or thirty are equally unfounded and equally useful.
Fish are stupid fails on the taxonomy before it fails on the evidence. There is no cognitive claim that can be true of a category containing hagfish and manta rays.
Fish do not feel pain is stated with more confidence than the evidence permits, and so is fish definitely feel pain. Nociception is established. The subjective component is unresolved and currently unmeasurable.
Fish have no memory beyond seconds is refuted by year-long hook avoidance and multi-year natal homing.
Sharks must swim constantly or die is true of some species using ram ventilation and false of many others that pump water over their gills and rest on the bottom, and recent metabolic work supports genuine sleep-like states in at least one species.
Fish cannot recognize individuals is refuted by wrasse client tracking and archerfish face discrimination.
Fish felt no pain because they lack a neocortex assumes the conclusion. The neocortical requirement is a hypothesis about consciousness, not an established constraint, and the avian evidence is a live problem for it.
Catch and release is harmless is not supported. Post-release mortality varies widely with species, hooking location, handling time, and water temperature, and is frequently substantial. Individual cutthroat trout in heavily fished stretches have been recorded caught and released nearly ten times in a single season, which is a different welfare question from the mortality one.
Fish are conscious and self-aware, the enthusiastic version, outruns what a contested mark test on one specialized species establishes.
What fish cognition is actually evidence for
Assemble it and fish cognition turns out to be less a story about fish than a controlled test of how comparative claims get made.
Every one of the capacities above was assumed absent until somebody designed a test the animal had a reason to take. The wrasse mark test worked with an ecologically meaningful mark and failed with an arbitrary one. The archerfish face task worked because shooting at things is what archerfish do. The grouper partner-choice result worked because hunting with a moray is a real problem the animal already solves. The failures follow the same rule in reverse. Fish do poorly on tasks requiring them to manipulate objects with appendages they do not have, to attend to human pointing gestures they have no reason to read, or to persist at problems in bare tanks under bright light with an observer leaning over them. Every time the task fit the animal, the capacity appeared, and the pattern is identical to the one that runs through the ape literature, the cat literature, and every case where a negative result turned out to be a statement about the experiment.
The second lesson is about anatomical prejudice. The argument that fish cannot have experiences because they lack a neocortex is the same argument that held bird forebrains were basal ganglia, and it failed there for reasons that apply here. Whether it fails here is not yet settled, and the honest position is that we are running a theory of consciousness we cannot test on an animal we cannot ask. The same impasse appears wherever the question comes up, and it is not a gap that better instruments will obviously close.
The third is scale, and it is the one with consequences. Fish are the largest vertebrate group, the most heavily exploited, and the least protected, and the gap between the evidence and the regulation is wider than for any other class of animal. Roughly a trillion individuals a year pass through a system built on the assumption that nothing much is happening inside them, and that assumption was never a finding. Cephalopods and decapod crustaceans acquired legal recognition as sentient in the United Kingdom on the strength of a systematic evidence review; fish, which are vertebrates with nociceptors and homologous stress systems, remain outside most welfare frameworks in most jurisdictions. The inconsistency is not defended on scientific grounds because it is not defensible on scientific grounds. It is a legacy of which animals people had already decided to think about.
None of that is a claim that fish are secretly primates. Most fish are small, short-lived, and running behavioral repertoires narrower than a crow’s. The claim is that the category was never coherent, that the tests were built for other animals, and that the confident negative was doing work nobody had earned.
There is no such thing as a fish, and the animals in that non-category include one that carries a clam to a specific rock, one that recruits an eel by pointing at a crevice, one that inspects a parasite it can only see in a reflection, and one that has been remembering the shape of a tide pool since the tide went out. Nothing on that list required a cortex, and the group they belong to was never a group. The 24-lecture Neurozoology course runs the whole tree of life on that basis, alongside the study of how knowledge moves between animals, the first edition’s survey of nervous systems, and the working animals whose capacities got discovered by people who needed something from them.
The goldfish was never the problem. The bowl was, and so was the story we told about why it was fine.