A honeybee has roughly a million neurons. You have about eighty-six billion. That is a ratio of about eighty-six thousand to one, and it is the number people reach for when they want to explain why bees are simple.
Here is what a million neurons does. It navigates several kilometers from a hive and returns by a straight line, correcting for the sun’s movement across the sky using an internal clock. It communicates the distance and bearing of a food source to other individuals through a symbolic dance. It learns to associate arbitrary colors, shapes, and scents with reward, and it learns abstract relational concepts like same and different, and above and below, which transfer to novel stimuli. It counts, to about four, and shows something resembling a zero concept, treating an empty set as lying at the low end of a numerical continuum rather than as a separate category. It recognizes human faces well enough to discriminate a trained target from novel ones. And when researchers gave bumblebees small wooden balls they had no reason to interact with, the bees rolled them, repeatedly, with no reward attached, more often when young, which is the operational definition of play.
That is not a simple animal running a small program. That is an extremely compressed implementation of most of the things a vertebrate brain does, and the compression ratio is the thing worth explaining. Insect cognition is the field where the relationship between neural hardware and behavioral capability breaks down most visibly, and it breaks in the direction nobody expected.
What neuron count predicts about insect cognition
The intuition that more neurons means more capability is not baseless. It is just a weak predictor with enormous residuals, and arthropods are where the residuals are largest.
Run the numbers across the group. A fruit fly runs about a hundred and forty thousand neurons. A honeybee around a million. An ant somewhere between a quarter and a million depending on species. A jumping spider, with visual capabilities rivaling small vertebrates, has a central nervous system of a few hundred thousand neurons packed into a body a few millimeters long, with so much of the animal given over to brain that in some species neural tissue extends down into the legs because there is nowhere else to put it.
Against that, a mouse has seventy million and a crow manages primate-grade cognition on well under two billion. The elephant carries 257 billion and puts almost all of them in the cerebellum. If neuron count were the variable, a bee should be somewhere below a nematode in capability, and it is not.
What appears to matter more is organization, and specifically two things. First, how much of the count is dedicated to computation rather than to sensory transduction or motor output. Insects put an enormous fraction of their neurons into the optic lobes, which is why the residual for central processing is smaller than the headline figure suggests, and yet the central processing is what does the impressive work. Second, how efficiently the circuitry is arranged for the specific problems the animal actually faces.
The general lesson is that neuron count is a budget rather than a capability, and what an animal buys with the budget varies enormously. Insects spend theirs on a small number of tightly optimized circuits that solve exactly the problems an insect has, with essentially nothing left over. Vertebrates spend theirs on generality, which is expensive and pays off in flexibility. Neither strategy dominates, and the group that built comparable capability on a completely different body plan made a third set of choices again.
Two structures doing most of the work
Insect brains are not miniature vertebrate brains. They are organized around a different plan, and two neuropils carry most of the load.
The mushroom bodies are paired structures handling learning and memory, and their architecture is one of the recurring designs in neuroscience. Sensory input, historically olfactory but including visual input in ants and bees, arrives at a region called the calyx and fans out onto a very large population of small intrinsic neurons called Kenyon cells, which then converge onto a small number of output neurons. That expansion-then-convergence arrangement produces sparse, high-dimensional representations in which similar inputs become easy to distinguish, and modification of the connections between Kenyon cells and output neurons is where associations get stored, with dopaminergic neurons delivering the reinforcement signal that determines which connections change.
That same many-to-few architecture appears in the vertebrate cerebellum, in the octopus vertical lobe, and in the avian pallium. Four lineages, no shared ancestral structure, one design. It keeps getting rebuilt because pattern separation is a problem with a good solution and the solution is findable.
The central complex is the other structure and it handles spatial orientation. It contains a ring of neurons that encodes heading direction, functioning as a compass, with activity that shifts around the ring as the animal turns. It integrates polarized skylight information, tracks self-motion, and is now well established as the site of path integration, the running vector calculation that lets an ant return home in a straight line after a meandering outbound trip.
That heading representation is the same computational object as vertebrate head direction cells, arrived at independently, in a structure with no anatomical correspondence to anything in a vertebrate brain. The machinery of spatial navigation turns out to have a small number of viable implementations, and insects found one of them first. Modelling work has shown that the central complex and the mushroom bodies can be combined into a single decentralised architecture that reproduces how real insects weight path integration against visual memory depending on which is currently more reliable, which is the same reliability-weighted arbitration larger brains perform across sensory modalities.
The complete synaptic wiring diagram of an adult fruit fly brain is now available, all hundred and forty thousand neurons and their connections mapped, which makes Drosophila the only animal above a nematode where the full circuit is known. That is a resource with no equivalent anywhere else in neuroscience, and it converts questions that were previously arguments about anatomy into questions that can be traced through actual wiring.
The optic lobes deserve their own note because they consume so much of the budget. In a fly, a large share of the total neuron count sits in visual processing, arranged into repeating columns that each handle one point in the visual field, with motion detection implemented in a circuit whose logic was worked out from behavior decades before the neurons were identified. That columnar arrangement is a different solution to the same problem vertebrate visual cortex solves with a different anatomy, and it delivers the temporal resolution that makes a fly effectively impossible to swat.
Bees, and the results that broke the ceiling
The bumblebee work of the last decade has done more to move assumptions about insect cognition than anything else in the field, and it culminated in a 2024 result that is genuinely difficult to accommodate.
The setup is a two-step puzzle box, deliberately designed after a lockbox task built for cockatoos. Opening it requires performing one action that produces no reward, followed by a second action that does. That temporal and spatial separation between the first step and the payoff is exactly what makes a behavior hard to acquire by trial and error, because nothing reinforces the first move.
Naive bees could not solve it. The experimenters could not even train demonstrators without temporarily inserting an intermediate reward for the first step, which was then removed. But once trained demonstrators existed, roughly a third of naive observer bees learned to open the two-step box by watching, without ever being rewarded for the first step themselves.
That is social acquisition of a behavior that lies beyond individual innovation, which is the technical threshold some researchers have used to define cumulative culture and to argue it separates humans from everything else. Finding it in an insect does not mean bees have culture in the full human sense. It means the threshold does not do the work it was asked to do, and that insect cognition sits on the wrong side of a line that was drawn specifically to keep it out.
Earlier work in the same program had already established the components. Bumblebees learn string-pulling from demonstrators and the behavior spreads through colonies. They learn to roll a ball to a target for reward, and observers improve on the demonstrated technique rather than copying it exactly, choosing a closer ball when the demonstrator used a distant one. Colonies seeded with a demonstrator using one of two possible box-opening techniques converge on the demonstrated variant and keep preferring it even after discovering the alternative, which is conformity.
The implication the researchers themselves drew is the interesting one: the elaborate nest architectures of bees and wasps, and the fungus-farming and aphid-herding of ants, may have originated as innovations that spread by copying before becoming fixed in the species repertoire.
Ants, and intelligence that is not in any individual
Ant colonies do things no ant does, and the mechanism is the reason the group belongs in a discussion of collective cognition rather than individual cognition.
Leafcutter ants run agriculture. They cut vegetation they cannot digest, transport it underground, and cultivate a fungus on it, maintaining monoculture gardens with antibiotic-producing bacteria on their bodies to suppress competing molds, removing contaminated material, and adjusting which plants they harvest based on how the fungus responds. That relationship is tens of millions of years old and it is farming by any functional definition, predating human agriculture by a factor of roughly a million.
Army ants build bridges out of themselves, with individuals locking together across gaps and the structure adjusting its position and size according to traffic, dissolving when the cost of the ants tied up in the bridge exceeds the benefit of the shortcut. Nobody is measuring that tradeoff. It emerges from each ant following local rules about when to hold still and when to move on, and the same principle produces the rafts that fire ants form during floods, where the colony becomes a buoyant structure with a viscoelastic response to stress that no individual could produce or perceive.
Weaver ants pull leaves together in chains and use their own larvae as tools, squeezing them to produce silk and stitching the leaf edges together, which is tool use with the tool being a live family member.
Foraging optimization runs on pheromone trails, which is memory stored in the environment rather than in any head, with shorter routes accumulating reinforcement faster and outcompeting longer ones. The colony finds the shortest path without any ant knowing what a path is.
Honeybee swarms run the same logic on a harder problem. When a colony divides, scouts survey candidate nest sites, return, and advertise them with dances whose vigor reflects site quality, with scouts also recruiting to other scouts’ sites and eventually producing a quorum at one location. The decision is genuinely distributed, it weighs multiple criteria including cavity volume and entrance size, and it reliably picks well. Nothing in the swarm has a representation of the alternatives.
The critical distinction is that none of this requires an individual ant to be smart. A single ant runs a modest behavioral repertoire triggered by local cues. The colony’s capability is a property of the interactions, which is why ant behavior has been productive as a source of optimization algorithms and why the same organizational logic appears in vertebrate groups whose collective vigilance exceeds any individual’s.
Individual ants are not negligible either. Desert ants run path integration accurate enough to return directly across featureless terrain, they count steps with an odometer that can be experimentally corrupted by gluing on stilts, and they learn visual panoramas. That is substantial individual navigation in an animal with a brain smaller than a grain of salt.
Jumping spiders, and vision that should not fit
Salticids are the arthropods that break the pattern hardest, because they run visual cognition at a level normally associated with small vertebrates on a nervous system that fits in a pinhead.
The eyes are the starting point. Eight of them, with the two forward-facing principal eyes having a narrow field of view, a long focal length, and a layered retina, mounted on muscles that move the retina behind a fixed lens to scan a scene. Spatial acuity in some species exceeds that of animals thousands of times larger. Some salticids have color vision extending into ultraviolet, achieved in one genus through a filter arrangement that produces a green-sensitive channel from a depth-dependent blur comparison.
The behavior that matters is detour planning. Portia is a spider that hunts other spiders, including species that would readily eat it, and it approaches by routes that require moving away from the prey and losing sight of it entirely. Placed on a platform with a view of the target, the spider scans, selects a route, and then executes it, sometimes over the course of an hour, arriving above the prey and dropping onto it on a silk line.
The experimental work on route selection during the locomotory phase of a detour established that the decision is made during a scanning phase before movement begins, that spiders select the route with an unbroken path even when a shorter but discontinuous option is available, that they scan again en route and select intermediate goals as they go, and that they will take a detour when necessary and a direct route when not. Later work found they choose whether a detour is needed before setting off. That is a plan formed from visual inspection and held in working memory through a period when the goal is not visible.
The neural resource allocation is what makes it remarkable. The principal eyes have a retina only a few receptors wide, which is why the retinal scanning is necessary: the spider builds a picture by sweeping a narrow high-resolution strip across a scene, which is a deliberate trade of instantaneous field of view for acuity in an animal that cannot afford both. Everything downstream runs on a few hundred thousand neurons.
Portia also improvises. Against web-building spiders it plucks the web to generate vibrations, and against some prey it varies the pattern until it finds one that produces an approach, which is trial-and-error signal generation rather than a fixed script.
The comparison worth making is to the tool-using birds and mammals whose planning gets described in mentalistic language by default. A spider doing the same thing gets an associative explanation first, and the asymmetry is about the observer’s expectations rather than the data.
The sentience question, and where it currently sits
Insects moved from clearly-not-conscious to actively-argued within about a decade, and the trajectory is worth tracking because the evidence accumulated in a specific direction.
The behavioral findings that shifted things: bumblebee ball-rolling meeting the criteria for play, with no reward and more of it in young animals. Honeybees showing pessimistic judgment biases after being shaken, treating ambiguous stimuli as more likely to be negative, which is the standard operationalization of an affective state in animals. Bees showing something resembling optimism after unexpected reward, with the effect blocked by dopamine antagonists. Crayfish and other arthropods showing anxiety-like states responsive to anxiolytic drugs. Injured insects showing altered nociceptive thresholds resembling sensitization, with work in fruit flies identifying a persistent hypersensitivity after nerve injury that outlasts the injury itself and that depends on a specific descending inhibitory pathway, which is the closest thing to a mechanism anyone has produced on this question.
The anatomical argument runs through the central complex, which some researchers have proposed as an integrating structure functionally analogous to the vertebrate midbrain and capable of supporting a unified egocentric model of the animal’s position and state. That is a specific and testable proposal rather than a general appeal to complexity.
The 2024 New York Declaration on Animal Consciousness placed insects in the category of realistic possibility rather than strong support, alongside other invertebrates, and its authors were explicit that they were not asserting insects obviously are conscious. The careful reading is the correct one, and a great deal of coverage flattened it into something stronger.
The skeptical position deserves its full weight. Behavioral analogies to vertebrate affective states can be produced by simpler mechanisms, the drug results demonstrate conserved neurochemistry rather than conserved experience, and the entire inference chain runs through a theory of consciousness the field does not possess. The same impasse appears with fish, with cephalopods, and with sleep and dreaming generally, and insects are simply where it bites hardest because the numbers are so large.
The practical consequence is the part that is moving. Insect farming for protein has scaled into the trillions of individuals annually, and it is being built with essentially no welfare framework, on an assumption about insect experience that the relevant scientists now describe as unresolved. The cephalopods and decapods that acquired legal recognition on the strength of an evidence review demonstrate that the framework can move when somebody assembles the evidence, and nobody has assembled it here at anything like the scale the numbers warrant.
The body as part of the computation
Arthropods offload an unusual amount of their control problem into anatomy, which is part of how the neuron budget stretches so far.
The exoskeleton is the clearest case. Leg joints have mechanical properties that produce stable behavior without neural correction: passive elasticity absorbs perturbations, and the geometry of the linkage means a leg encountering an obstacle deflects in a useful direction automatically. A cockroach running over rough terrain is being stabilized substantially by its own mechanics rather than by a controller computing corrections, which is why cockroach locomotion degrades gracefully rather than failing when the terrain gets unpredictable.
Sensory hairs distributed across the cuticle deliver airflow, vibration, and contact information directly to local circuits, with escape responses running through short reflex arcs that bypass the brain entirely. Segmental ganglia handle a great deal of local business independently, which is why a decapitated cockroach walks and why the distributed-with-oversight arrangement recurs across so many body plans. A cockroach begins turning away from an approaching predator within a few tens of milliseconds of detecting the air disturbance, and the conduction delays that make that speed remarkable are exactly why the circuit is short.
The web is the extreme version and it has generated a real argument. An orb weaver’s web is a sensory apparatus: the spider reads prey type, size, and position from vibration patterns, and it adjusts web tension, which changes the transmission properties. Some researchers have argued this constitutes genuinely extended cognition, with the web functioning as an externalized part of the perceptual and computational system rather than as a tool the spider uses. Whether that is a substantive claim or a redescription is contested, and the phenomenon is the same one that shows up whenever an animal incorporates an external object into its own operation.
Metamorphosis, and the brain that gets rebuilt
Holometabolous insects do something with their nervous systems that has no vertebrate parallel and that raises a question nobody has fully answered.
A caterpillar and the moth it becomes are the same individual with different bodies, and the transition is not a gradual reshaping. Inside the pupa, much of the larval tissue is broken down, imaginal discs that have been carried since embryogenesis expand into adult structures, and the animal is substantially reconstructed. The nervous system is remodeled heavily: some larval neurons die, others are pruned back and regrow with entirely different connections, and new neurons are added for structures the larva did not have, including flight muscles, compound eyes, and reproductive organs.
The question is whether anything survives that. It does. Moths trained as caterpillars to avoid an odor paired with a mild shock retained the aversion as adults, which means information encoded before metamorphosis persisted through the reorganization. The mushroom bodies appear to be the relevant structure, with some of their neurons surviving and being incorporated into the adult circuit rather than being replaced.
That result belongs alongside the planarians that retain learning through decapitation and regeneration of a new brain as a case where memory outlasts substantial destruction of the structure that formed it, and both raise the same question about where the trace is actually held. Neither has a clean answer, and both suggest the trace is distributed across tissue rather than localized to the circuit that appears to be doing the work.
The functional logic of metamorphosis is worth stating too, because it is a strategy rather than a quirk. Larva and adult occupy entirely different ecological niches, eat different food, and do not compete with each other, which lets a single species exploit two ways of making a living. The cost is the reconstruction. The insect nervous system evidently supports being taken apart and reassembled in a way no vertebrate brain would tolerate.
Where the ceiling actually is
Insects do not do everything, and the failures are as diagnostic as the successes.
Individual learning capacity is genuinely limited. A bee learns associations quickly and forgets them on a schedule, and the number of independent items it can hold is small. Long-term memory exists but the capacity is nothing like a corvid’s cache inventory. Flexible reversal of learned rules is harder for insects than for vertebrates of comparable ecological complexity, and behavior tends to be more strongly channeled by species-typical routines.
Generalization outside the trained domain is where the gap is clearest. A bee that has learned an abstract same-different rule transfers it impressively, and that transfer is narrower than what a primate or a corvid manages with an equivalent rule. The circuits are specialized, and specialization buys efficiency at the cost of range.
Lifespan is the other hard constraint. A worker bee lives weeks in the foraging season. Whatever it learns dies with it, and the colony’s persistence across seasons runs through the queen and through comb structure rather than through accumulated individual knowledge. That forecloses the kind of long-horizon accumulation that makes decades-long social memory worth building, and it is the same constraint that shapes cephalopod cognition from the other direction, where a solitary animal with no generational overlap has to build everything within a single short life.
None of that makes the capabilities less real. It locates them. Insects are extraordinarily good at a specific set of problems within a short life on a tiny energy budget, and the honest framing is that they are not scaled-down vertebrates any more than they are automata.
The claims that do not hold up
An audit, because this group attracts both dismissal and inflation.
Insect cognition does not exist, in the sense that insects are simple reflex machines, was the twentieth-century default and it does not survive the learning literature. Bees form abstract concepts, transfer them to novel stimuli, and acquire behaviors socially that they cannot innovate individually.
Insects definitely feel pain overstates the evidence in the opposite direction. Nociception is established. Affective experience is not.
Insects cannot feel anything because they have too few neurons assumes a threshold nobody has identified, and the neuron-count argument has failed everywhere else it has been applied.
The waggle dance is not really communication was argued for decades and was settled by experiments using robotic bees and by tracking recruited foragers, which confirmed that receivers extract distance and direction and fly to the indicated location. The dance is one of the few genuinely symbolic communication systems outside human language, encoding a vector in a body movement performed in the dark on a vertical surface with the sun’s position substituted by gravity.
A queen ant commands the colony inverts the structure. Queens lay eggs. Colony decisions emerge from worker interactions and pheromone signals with no central authority anywhere.
Spiders are insects is a taxonomic error worth correcting since it appears constantly. Arachnids have eight legs, two body segments, no antennae, and diverged from insects well over four hundred million years ago. Crustaceans are closer to insects than spiders are, and the current picture places insects inside the crustacean radiation, which makes a shrimp and a bee closer relatives than a bee and a spider.
Insects are individually mindless and only colonies are impressive undersells solitary species. Jumping spiders, solitary wasps, and dung beetles that navigate by the Milky Way, taking a snapshot of the sky and rolling in a straight line against it, are not running on collective intelligence. Most arthropod species are solitary, and the social insects that dominate popular attention are a small and unrepresentative slice.
Bees can see ultraviolet so flowers look completely different to them is true in direction and frequently overstated in specifics, since the reconstructions circulating are approximations built from photographic filters rather than measurements of bee perception.
What insect cognition is actually evidence for
Assemble it and the group makes three arguments that the rest of comparative neuroscience keeps needing.
The first is about compression. A million neurons is enough for navigation with path integration, symbolic communication, abstract concept learning, social transmission of behaviors beyond individual innovation, and something meeting the criteria for play. Whatever the relationship between neural hardware and capability is, it is not linear and it is not close to linear. That does not mean bees are secretly vertebrates. It means capability per neuron varies by orders of magnitude depending on how specialized the circuitry is, and generality is the expensive thing rather than capability as such. That reframing matters well beyond insects, because every comparative ranking built on brain size or neuron count is implicitly assuming the opposite, and the comparison of cortical neuron counts across species is where the assumption gets tested most directly.
The second is about convergence at the circuit level. Mushroom bodies and cerebellum and octopus vertical lobe implement the same expansion-and-convergence design. The central complex and vertebrate head direction cells implement the same heading representation. The camera eyes and echolocation systems that keep getting rebuilt make the same point at the level of sensors, and the independent construction of nervous systems at the base of the animal tree may make it at the root. There appear to be a small number of good solutions to the recurring problems, and evolution finds them repeatedly. That is a claim with teeth, because it means the features shared across independent nervous systems are the ones forced by the problem, and the features unique to one lineage are historical accident.
The third is about scale, and it is uncomfortable. Arthropods are most of animal life by species count and by individual count, they include the animals whose collective behavior underwrites global agriculture, and they are the group about which the least is known relative to their abundance. The insect sentience question is unresolved and involves numbers that make every other welfare debate look small.
The 24-lecture Neurozoology course runs 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.
A bumblebee cannot work out the two-step box on its own. Watch another bee do it, and a third of them can. That is a fact about a creature with a million neurons, and it was published in 2024, which is a reasonable estimate of how recently the ceiling on insect cognition was where everyone assumed it had always been.
Leave a Reply