What Are Animals Talking About—and Can AI Help Us Understand Them Better?

Niloofar

September 30, 2026

My closest encounters with what looks like animals “talking” usually happen in our garden. It is also where animal communication first became something I wanted to understand. When I work outside, I sit quietly enough to fade into the background, happily spying on the animals.

Then come the chattiest of our tree-dwelling neighbors: the magpie family.

One day I was on the hammock with my son, hugging him while he told me about his day at kindergarten. The magpies arrived and started one of their noisy exchanges, repeatedly cutting into his story. Irritated, he finally asked, “Can you tell the magpies to stop talking? It’s very rude.”

Apparently kindergarten had taught him that interrupting is bad manners.

But it was the word talking that caught my attention. He does not normally use it for our cat or dogs. For the magpies, though, it came naturally. To him, they were talking—and, rather annoyingly, talking over him.

That left me with some questions: When animals pass information to one another, when does that count as communication? And when, if ever, does communication become language? How much of animal communication have scientists really decoded?

There is another reason I became curious. For years my husband, who is fascinated by new technologies, has been convinced that AI may eventually help us understand animal communication.

So another big question behind this article is: Can AI really help us understand animal communication?

Here, I focus on animals communicating naturally with one another. I don’t cover attempts to teach animals human language or artificial communication systems, which are a different subject.

My son’s sweet, naive comment was what finally triggered my curiosity enough to turn these questions into this article.

How does information pass between animals — and when is an animal really "saying something"?

Information moves between animals almost constantly, but only part of that movement is communication. A useful way to sort out what counts as communication is to picture a ladder with five rungs. On Rung 1, information simply leaks out of an animal that is busy doing something else. On Rung 5 sits human language. As you climb, the animal has more and more control over what gets sent — and it looks more and more like the animal is saying something. The climb is a gradual slope rather than a single step, which is why the point where “saying something” begins is genuinely blurry.

Rung 1: information that just leaks out

Start with the kind of event that is not communication at all. A mouse rustling through dry leaves as it forages at night is not communicating with an owl. That owl is hunting the mouse by listening for exactly that rustling sound. The mouse is making a noise as it goes, and the owl is simply listening in. A leak like this is called a cue: information that escapes incidentally, as a by-product of some activity other than communication. Nothing has been sent. Cues still matter, because they are often the raw material for signals. A signal can get started out of a cue. Animals are already watching some body feature, or some action, of another animal as a cue. Then that other animal develops a slightly exaggerated version of the feature, or performs a slightly enhanced version of the action. That exaggerated feature, or that enhanced action, is where a signal begins.

Rung 2: communication, and the signal inside it

Communication is the whole event, not just the thing sent. It is the transfer of information from one animal to another in a way that affects behavior and the survival or breeding success of one or both of them. It has three parts: a sender, a signal, and a receiver who does something in response. The signal is only the middle part. Unlike a cue, a signal is produced specifically to give information to receivers.

Ants show the whole event plainly. A worker that finds food lays a chemical trail along the ground on her way back to the nest. Other workers then follow that marked trail out to the food on their own. Sender, signal, receiver, response — the full loop.

A signal also has to travel from the sender to the receiver, and it can travel by any of several routes. Each route is called a channel. A channel is not a step up the ladder — it is a property of the signal, one of the physical ways a signal can cross the gap. Animals monitor cues carried by sound, light, chemicals in their surroundings, electric fields, movement in the surrounding water or air, and touch. Every one of those six has also been recruited to carry signals by at least some animals. One example of each is enough here:

Sound. Male Cassin’s finches sing to attract females.
Chemicals. The ant trail just described is a chemical signal, laid on a surface rather than released into the air
Light and color. Male fireflies emit repeated flashes of light to attract mates, and the pattern of long and short flashes differs from species to species.
Touch. Many birds and mammals groom or preen one another, which serves to confirm friendly bonds or relative rank within a group.
Movement of water or air. In some crickets that do not chirp, a puff of air is the signal itself: they fan air at each other with their wings during courtship and fights, and other crickets of the same species feel the puffs.
Electric fields. Knifefishes and elephantfishes vary how fast they pulse and the shape of each electric pulse. They use these changes to handle territory defense, fights, rank, courtship, and keeping a foraging school together. Air conducts electricity so poorly that electric signaling among animals is limited to water.

Many signals also combine channels, so an animal may be sending on two of them at once.

One oddity is worth a mention. A surprising number of animals can detect the Earth’s magnetic field, yet as of 2011 no animal had been shown to use magnetism to signal to another.

Rung 3: meaning

Rung 3 is referential meaning: the signal gives listeners information about a particular thing out in the world, rather than only broadcasting the sender’s own mood or condition.

One of the best-studied cases is the vervet monkey (Chlorocebus pygerythrus) of the Kenyan savanna, which has distinctive alarm calls for different kinds of danger. When researchers played a recording of the leopard alarm with no predator anywhere near, the monkeys still ran into the trees.

Two things need saying about what a result like that does and does not establish. The leopard alarm goes with a leopard, and listeners react as though a leopard were there. Scientists are deliberately not claiming that the monkey means “leopard” when it calls.

Rung 4: signals the animal varies

Rung 4 is the flexible signal: a signal the animal changes in order to convey different information.

Rung 3 and Rung 4 are easy to confuse, so it is worth being blunt about the difference. They are two separate properties of a signal, and they answer two different questions. Referential asks one question: does the signal carry information about a specific thing out in the world? Flexible asks a different one: can the animal vary the signal to convey different things? The vervet monkey is the example for the first question. It has separate calls for a leopard, for an eagle, and for a snake, and the meaning comes from which of those calls it gives. The honeybee is the example for the second question. She performs one kind of dance and alters it, and the alterations are what carry the different information. So a flexible signal is not a more advanced grade of referential signal. The two are different properties, and the question each one settles is different.

The honeybee waggle dance is a clear example of a flexible signal. A returning forager dances on the honeycomb. The direction of her waggling run, and how long the run lasts, both shift according to where the food she found is sitting. One dance form, many different messages.

Ants have a less obvious flexible signal of their own. Workers that follow a trail and find plenty of food reinforce the trail chemical on the way home. Workers that find little food, or find the food source used up, mark less or not at all, and the trail fades away.

The thread running through the whole ladder: control rises, gradually

Notice what has been creeping upward alongside meaning. At the bottom, nothing is chosen at all — the mouse’s rustle is not the mouse’s decision. A little higher, some signals are fixed. The red, white, and blue coloration on the head of an adult male mandrill (Mandrillus sphinx) stays fixed and on display for his entire life. A male wild turkey carries a similar set of colors on fleshy ornaments on his head. He can deflate many of those ornaments when he is not displaying to females or confronting a rival. The two animals wear nearly the same colors, and their control over those colors is the exact opposite. The turkey can deflate his ornaments and turn his display down. The mandrill has no such off switch — his colors are on for his whole life.

One clear piece of evidence that an animal has such a switch is the audience effect: senders adjusting their signaling depending on who is likely to be around. Roosters are less likely to give food calls when hens are absent. Yellow mongooses give alarm calls only when others of their own kind are nearby. Fork-tailed drongos feed in the canopy and routinely call when a predator appears overhead. For predators on the ground they rarely call at all — unless ground-feeding pied babblers are in their mixed flock of several bird species. When those babblers are there, the drongos do call. All three cases show the same pattern: the animal is more likely to signal in some company and less likely to signal in other company. That is control, and control is something you can watch and count.

Now put the vervet monkeys beside those three cases, because the contrast between the vervets and those three cases is the most important idea on the whole ladder, and the easiest one to blur.

What we can see is the listeners’ side. The vervet result described at Rung 3 is a good example. The monkeys’ reaction can be watched, filmed and counted, and it shows the listeners drawing a correct conclusion about the world.

What we cannot see is the caller’s side. The caller may be trying to inform the others, or may simply be making a sound that the others have learned to read. From the outside those two possibilities look identical: one monkey calls, and the rest go up the trees. Watching the listeners does not separate them, and neither does playing the call back, however carefully.

So control and intent are not equally knowable. Control is observable: you can watch whether an animal sends a signal or withholds it, whether it varies that signal, and you can count what it does. Intent is not observable in the same way. Evidence from watching behavior shows what animals do, not what they understand while doing it, so it cannot settle what the calls mean to the animals themselves.

Rung 5: language, in one paragraph

Language sits at the top, and it is built differently. Three of its hallmarks are enough to show why. The first is that its units carry specific meanings: a word means a particular thing. The second is that those words are arbitrary. A word generally does not resemble the thing it stands for. The word for a dog is dog in English, Hund in German, and chien in French, and none of those words sounds like a dog. The third is displacement: meanings about the past, the future, or distant places can be encoded and understood, not just meanings about what is here and now. You can describe a meal you ate a decade ago, or a mountain you have never visited. No other species has been shown to possess language as a whole, although many share important components of it. So, on present evidence, language appears to be uniquely human.

So when is an animal really saying something? The ladder gives no single rung where saying something starts. What it gives instead is a rising line: from a leak the animal cannot help, to a signal it cannot switch off, to a signal it can withhold when nobody useful is listening, to a signal it reshapes to fit what it has found. By the time you reach the honeybee’s dance or the drongo’s selective alarm, “saying something” feels like a fair description of what the animal is doing. Whether it feels that way to the animal is a separate question, and one that watching behavior does not settle.

Sources

What are some examples of animals communicating with sound?

Elephants, crickets, spiders and baleen whales all send messages with sound: through air, through water and through solid ground. Baleen whales are one of the two main whale groups; the other is the toothed whales. Sound is pressure moving through matter. Its speed depends on what it is moving through: about 344 meters per second (1,129 feet per second) in air, about 1,500 meters per second (4,900 feet per second) in water, and about 5,100 meters per second (16,700 feet per second) along a bar of steel. Whether the sound travels through air, water or solid material, the delay between sender and receiver is usually milliseconds.

Elephants are one long-distance case. Female elephants can recognize up to 100 close and distant relatives by the low “rumble” contact calls those relatives make — contact calls being the kind animals use to keep in touch with their group. They can still recognize these relatives at ranges of up to 3 kilometers (1.9 miles).

Even one widespread species can sound different from place to place. The Eurasian magpie (Pica pica) has a rattling “chatter” call — the call it makes most often throughout the year. A 2011 study compared the four-syllable chatter calls of sericea, the magpie subspecies found in Korea and Japan. The study covered seven locations: five on the South Korean mainland, plus two island sites, Jeju in Korea and Saga in Japan. The calls differed measurably across those seven locations. Island birds differed from mainland birds. On the mainland, northern birds differed from southern ones. Overall, the magpies at Saga were the most distinctive of the seven groups of birds. What these call differences mean to the magpies themselves has not been demonstrated.

Sound also travels through solid ground, and a surprising range of animals send signals that way. Fiddler crabs, snakes, rodents and elephants all signal at the surface of the earth. Ground signals usually travel as ripples running along the surface. Low frequencies fade more slowly than high ones. Most animals signaling at a surface favor the lowest frequencies their bodies can detect. In dry sand, the band of frequencies that carries best is 300 to 400 hertz, meaning 300 to 400 vibrations per second. Elephants appear able to pick up these very low-frequency surface ripples from kilometers away.

Some animals build themselves an instrument. A calling cricket vibrates its wings, but wings on their own make a weak low call. The air the wings push on one side slips around the body and cancels most of the low notes on the other side. That matters, because low notes are the ones that travel farthest. Tree crickets fix that problem by chewing a hole in a leaf and calling with their bodies in the hole. The leaf blocks the short way around, so far less of the sound cancels itself, and the cricket’s low notes come out at a reasonable volume.

Other crickets dig their instrument instead. The short-tailed cricket (Anurogryllus muticus) digs a horn-shaped hollow in the ground and sits inside it. Echoes off the walls turn the hollow into a directional sound source. Mole crickets go further. They build a horn at the ground surface, and a narrow tunnel connects that horn to a bulb-shaped chamber underneath.

Spiders signal on a structure they built themselves. Signals travel along spider webs, and spiders can build a web whose sound-carrying properties suit them. In a typical orb web, signals seem to travel best as lengthwise vibrations in the spoke threads. Carried this way, all frequencies travel about equally well up to 9 to 10 kilohertz, meaning 9,000 to 10,000 vibrations per second.

The deep ocean holds a layer called the SOFAR channel — short for sound fixing and ranging — where sound travels more slowly than in the water above and below it. Sound that starts to stray out of the layer is bent back into it instead of escaping, so it can carry extraordinarily far. In tropical and warmer temperate oceans the channel typically begins 600 to 1,200 meters (about 2,000 to 3,900 feet) down. Whether any animal uses it to communicate is unsettled. It has been suggested that baleen whales could use the SOFAR channel, but as of 2011 most baleen whales appeared to call from much shallower water, 20 to 40 meters (66 to 131 feet) down.

Sound feels like the obvious way to send a message, because people are born into a world where sound communication is a dominant activity. But producing a useful sound signal presents significant hurdles. Most animals do without sound communication altogether.

Sources

What are some examples of animals communicating with visual signals?

Jumping spiders, peacocks, quail, fiddler crabs, fireflies and squid all signal by being looked at. Animals use three very different mechanisms to generate light signals.

Pigments. A pigment soaks up some wavelengths of light and lets the rest come back. The color you see is whatever the pigment did not absorb.
Structured surfaces. Microscopic structures reflect certain wavelengths back rather than absorbing them.
Light the animal makes itself. Some animals generate their own light, which is called bioluminescence.

On top of those three ways of making color and light, animals also combine body orientations, postures, gestures and movements into visual signals.

Fixed markings are one kind of visual signal. Males of different species of jumping spiders in the genus Habronattus wear strikingly different head and face markings. The metallic color of the eyespots in a peacock’s tail comes from a square array of microscopic rods of melanin, a group of dark pigments, in the feather. That array reflects light back rather than soaking it up. Males with brighter eyespots have higher mating success. Those precisely built arrays are easily disrupted by poor condition or a heavy parasite load.

Posture carries messages without any color change at all. Male Gambel’s quail (Callipepla gambelii) raise the feathered crest on their heads during aggressive encounters. Dominant birds hold the crest erect and tipped forward. Subordinate males flatten it down and point it backward.

Movement can be a whole conversation. At low tide, each male fiddler crab (Uca annulipes) advertises himself and his burrow to females by waving his single, greatly enlarged claw. When a female goes to inspect one male’s burrow, the other males cluster around and match both the rate and the timing of their waving to his. This matched waving brings the males no advantage in avoiding predators or in attracting extra females. Instead, it emerges from the competition between neighboring males.

Then there is self-made light. In firefly beetles, the males emit repeated flashes to attract mates. The pattern of long and short flashes is specific to each species, which keeps mating within the species. Males of one firefly species, Photinus carolinus, go further and synchronize their flashes with each other. Crustaceans called ostracods court with light too. Males of the Caribbean species Vargula annecohenae display to females at sunset by squirting small packets of light-producing chemical into the water. In Vargula ostracods, a courting male swims a corkscrew path and then spirals upward. As he swims, he puffs the light-producing chemical out through nozzles near his mouth. Females that are not lit up watch him and follow.

Squid and octopus show the most sophisticated color change used for signaling. Squid and octopus change color with chromatophores, which are small many-celled organs. Each chromatophore is a pouch of brown, red or yellow pigment. Muscle fibers radiating out from the pouch pull it open, under the animal’s own nervous control. By expanding different-colored chromatophores in combination, the animal runs through a variety of colors in rapid succession. It uses those color patterns in courtship, in aggression and in holding a group together.

Visual signals are limited by range and by line of sight. In clear air, a large object that contrasts with its background can be picked out at about 6 kilometers (3.7 miles). In dusty, humid or foggy air, that distance drops to roughly 50 meters (165 feet). In very clear ocean water, the maximum viewing distance is around 60 meters (200 feet), and it is usually much less in fresh water. Anything opaque between the sender and the receiver blocks the signal outright, because light cannot bend around objects the way sound can.

Sources

What are some examples of animals communicating with chemicals and scent?

Elephants, moths, rabbits, sea anemones and ants all send chemical messages. Chemical signaling is the oldest method of communication there is. A chemical that carries a message between members of the same species is called a pheromone. Some pheromones, called primers, change the receiver’s body chemistry slowly over time. Others, called releasers, produce an immediate response: approach, retreat, freezing, or some other recognizable movement or posture.

The same molecule can turn up in wildly unrelated animals. A chemical called (Z)-7-dodecenyl acetate is a main sex-attractant ingredient for the Asian elephant. It is also a main sex-attractant ingredient for about 140 species of owlet moth.

Some chemical messages are aimed at the very young. Mother rabbits (Oryctolagus cuniculus) produce a mammary pheromone that attracts their pups and guides them to suckle.

Some chemical messages are aimed at the neighbors. When the colonial sea anemone Anthopleura elegantissima is wounded, it releases a small, quickly evaporating alarm pheromone from its tissue. The anemones next to it pull back sharply.

Ants run some of the most elaborate chemical signaling systems. Ant trail pheromones can be a single chemical, a simple blend of two or three chemicals, or a complex mixture of up to 14 components. How long a trail pheromone lasts is matched to what the species needs. Pharaoh ants (Monomorium pharaonis) mark a trail branch that no longer leads anywhere useful with a long-lived repellent. They mark a branch that does lead somewhere useful with a short-lived attractant. Pharaoh ant trails also branch off the main trail at a standard 60-degree angle. Because of that fixed angle, a worker who loses the trail and finds it again can tell which way is home.

Chemical messages carried through air or water share one limitation: they are slow. Odor molecules have to travel the whole distance themselves. Either a current of air or water carries them along, or they drift on their own from where they are concentrated toward where they are not. So the typical delay between sender and receiver runs to seconds, minutes or even days. Molecules left to drift on their own move through water roughly ten thousand times more slowly than through air. So a moth signaling for a mate releases its scent slowly and keeps releasing it for several hours. That gives distant moths time to catch the scent and fly toward the source.

Sources

Bradbury, J.W. & Vehrencamp, S.L., Principles of Animal Communication, 2nd ed., Sinauer Associates, 2011 — Chs 6, 14

What are some examples of animals communicating by touch?

Zebras, buffalo, chimpanzees, flies and ants all send messages by making contact. Touch is probably the most widespread sense there is. But touch signals work only when the sender and the receiver are actually touching. So they show up in the situations where animals are already close: fighting, courting, greeting and grooming. Touch also registers faster than sight or smell. A touch deforms the sense receptor directly. It does not need the chain of chemical reactions that vision and smell require. So in most animals, nerve responses to touch fire 10 to 100 times faster than responses to light or to chemicals.

Contests are one place touch shows up. The early phases of a fight are usually handled by displays at a distance: visual, sound, scent or electrical displays. Touch becomes common as the fight escalates. Mountain zebras (Equus zebra) challenge each other by laying their heads on the opponent’s shoulders and pressing down. Once a fight is settled, the loser may make a submissive gesture that also uses touch. A subordinate African buffalo (Syncerus caffer) lowers its head and places it gently under the belly, or between the rear legs, of a dominant animal. A chimpanzee (Pan troglodytes) calms a dominant animal by gently touching the dominant animal’s genitals.

Courtship runs the same way. It opens with signals that work at a distance and then moves to contact. Mediterranean fruit flies, or medflies (Ceratitis capitata), court with the fine filaments dangling from their antennae. The male faces the female and rocks his body. As he rocks, he uses his own antenna filaments to spread hers outward. If she accepts this, he moves behind her to mate. Female harvester ants (Pogonomyrmex) send the opposite message, just as physically. When a male mounts a female that is not receptive to mating, she produces a buzzing vibration. The vibration travels straight through the points where her body touches his, and he dismounts.

Greeting uses touch as well. Ants, bees, wasps and termites rhythmically tap a new arrival’s antennae or body with their own antennae. The tapping tells the insects doing it who has turned up. It also tells the newcomer that it has been noticed and assessed. An elephant greets a familiar elephant by putting the tip of its trunk into the other elephant’s mouth.

Then there is grooming one another. Grooming is perhaps the most widespread of all the touch behaviors that hold a group together. Grooming strips parasites and debris off the groomed animal. It also reaffirms a friendly bond, or a difference in rank, between the two animals. And it gives the groomer a chance to inspect the scent on the other animal’s body.

Touch can even carry directions. Some ants recruit nestmates to newly found food by tapping them with their antennae and forelegs. These ants then lead the recruits to the food. On the way, the leading ants stay close enough to the recruits to keep physical contact.

Sources

Bradbury, J.W. & Vehrencamp, S.L., Principles of Animal Communication, 2nd ed., Sinauer Associates, 2011 — Ch. 7 “Short Range Modalities”

How do animals communicate through movements in water and air?

By shoving the water or air around them hard enough that a nearby animal feels the shove. A fin beat or a wing flick does not just make a sound. It also pushes the water or air itself, and an animal close by can feel that push directly. Signaling by moving water or air works only over a very short range. For a typical fish the useful reach is only one to two body lengths.

What the receiving animal picks up is not a steady push. It is a difference in movement between two separate points in the water or air. One patch of water or air moves, or speeds up, differently from another patch nearby. To pick up that difference, an animal has to sample the water or air in two or more places at the same time.

Fish and water-dwelling amphibians do this with the lateral line. The lateral line is a straight or curved row of movement detectors running along the body. Each detector, called a neuromast, is a cluster of sensory cells capped by a blob of jelly that moves when the water moves past. Squid and cuttlefish carry similar rows of detectors. Some insects, crickets among them, use hairs instead. A cricket has two cerci, a pair of projections at its rear end. Each cercus carries about 1,000 hairs that respond to air currents. Different hairs respond to currents coming from different directions.

Some of what the lateral line and the cricket’s hairs pick up is a genuine signal. In certain crickets that do not chirp, the moving air is itself the signal. Those crickets fan small puffs of air with their wings during courtship and during fights, and other crickets of the same species feel those puffs with their cerci. Several kinds of fish use tail beating or fin fanning as movement signals when they fight and when they court. Schooling fish, squid, and cuttlefish read the movements of their neighbors this way. They use that information to hold their spacing and keep the group together.

Signaling by moving water or air has one strange feature. A sound is gone from any given spot the moment it passes. But a moving animal leaves a wake, and a wake lingers. Swirls can stay in the water for several minutes after even a small fish has swum by.

Sources

Bradbury, J.W. & Vehrencamp, S.L., Principles of Animal Communication, 2nd ed., Sinauer Associates, 2011 — Ch. 7 “Short Range Modalities”

Which animals communicate using electric fields, and how does that work?

A small set of fish. Electric fields do not travel well enough between land animals to be useful for signaling. So electric signaling happens in water. In fresh water, the fish that signal this way include the South American knifefishes, the African elephantfishes, a single African species called the frankfish, and several kinds of catfish. In the sea, skates appear to signal electrically too.

The body equipment for electric signaling has two halves. The sending half is an electric organ. In fish, the electric organ is usually built from modified muscle cells. In skates and elephantfishes, the electric organ sits in the narrow stretch of tail between the body and the tail fin. The frankfish and most knifefishes use body muscles running from just behind the gills to the tip of the tail. The receiving half is a set of electroreceptors. Knifefishes, elephantfishes and the frankfish all ended up with a similar receptor type tuned to the fast, high-frequency part of an electric discharge. Elephantfishes went further and grew a separate receptor for communicating with each other, called the knollenorgan. The knollenorgan is usually 10 to 20 times more sensitive than the receptor elephantfishes use for finding objects.

One fact shapes everything else about how these fish signal. Knifefishes (apart from electric eels), elephantfishes and the frankfish give off electric discharges continuously. They do this because they use those discharges to feel out the objects around them in dark or muddy water. So the only way these fish can communicate with electric discharges is to alter the pattern they are already producing — the rate of the pulses, the shape of the pulses, or both. They do exactly that, in several ways: sudden brief speed-ups called chirps, gradual speed-ups and slow-downs, repeated rises and falls called warbles, and short silences. These changes in the discharge pattern are used to defend territory, to fight, to settle dominance, to court, and to keep a foraging school together.

Identity information is often important to these fish. Territorial species benefit from knowing an intruder’s species, sex and individual identity. More social species benefit from knowing age, sex and rank. Knifefishes, elephantfishes and the frankfish encode that information by varying the discharge rate, the pulse shape (called the waveform), or both. Where many species live side by side, differences in waveform help sort species and sex apart. In some species, the waveform appears to be distinctive enough to identify the individual fish.

One example shows the trade-off involved. In some pulse knifefishes, such as Brachyhypopomus pinnicaudatus, and in elephantfishes such as Marcusenius macrolepidotus, males remodel their electric organs as they mature. As a result, each male discharge lasts much longer than a female’s. The long pulse carries lower frequencies. So females can recognize a male from farther away. The long pulse also makes the male easier for predatory electric eels and catfish to find and attack.

There is a crowding problem too, and a fix for it. When two of these fish are close, their discharges can smear together. So both wave species and pulse species shift their discharge rates to reduce the overlap. This adjustment is called the jamming avoidance response.

Sources

Bradbury, J.W. & Vehrencamp, S.L., Principles of Animal Communication, 2nd ed., Sinauer Associates, 2011 — Ch. 7 “Short Range Modalities” and Ch. 1

Which animals use more than one kind of signal at the same time?

A great many — more than was once realized. A channel is one of the routes a signal can travel by, such as sound, sight or smell. When an animal recruits more than one channel into a single display, that display is called a multimodal signal. Combining channels increases the number of distinct messages the animal can build. Two broad patterns show up. Courtship displays in many backboned animals mix a visual part with a sound part. The signals of insects, spiders and water-dwelling animals without backbones often mix a visual part with a smell part. Some of these signals add sound as well.

Peacock spiders are a good place to see two channels working at once. A male of the Australian peacock spider Maratus volans unfurls a fan from under his abdomen. The fan looks like the tail of a male peacock. He aims it at a watching female and waves his legs in an intricate, coordinated dance. While he is doing all that, he also shakes his abdomen back and forth to produce vibrations. In a series of lab experiments, a research group led by Damian Elias found that both the visual part and the vibration part mattered to whether a female chose a male and whether he mated. Even so, females were more likely to mate with males that spent more of the display time on the visual part. One proposed explanation is that combining channels lets a female decide faster and more reliably. Deciding faster and more reliably may matter because many peacock spider species often turn up in the same places.

The barn swallow (Hirundo rustica) combines a different pair of channels. Male barn swallows communicate with both males and females using song and the patch of red color on the chest. One channel is heard and one is seen, and both are aimed at both sexes.

Why use two channels instead of one? There are two current explanations, and no settled choice between them. In the first explanation, each channel may carry different information, so the receiver learns more from the pair than from either alone. In the second explanation, each channel may carry the same information independently. Then one channel still gets through when the other is drowned out by noise or garbled by a mistake. Combining channels is not free, either. To start with, the animals involved need the right equipment for sending and detecting signals in more than one channel. And coordinating the parts of a display may take nervous-system wiring that a single-channel signal never needs.

Sources

What did the vervet monkey alarm call experiment show, and why was it revised?

The experiment showed that three different alarm calls set off three different responses, with no predator present to prompt those responses. It was revised because no other team reproduced the original finding for decades, and when researchers went back to it in 2013 and 2015, it did not hold up in the form it was first reported.

The study was published in 1980. Robert Seyfarth, Dorothy Cheney and Peter Marler recorded the alarm calls that vervet monkeys in the Kenyan savanna give when leopards, eagles and snakes are present. They then played those recordings back with no predator anywhere near, so nothing but the sound could account for what the monkeys did. The monkeys responded differently to each call. A leopard alarm sent them into the trees. An eagle alarm made them look up or run for cover. A snake alarm made them look down, or walk toward the source of the sound. Each call carried something different about the kind of predator, and the listeners acted on the difference. The 1980 study became the textbook example of an animal call carrying a meaning.

The revision came much later. No other study reproduced the 1980 findings for decades. Later work reported in 2013 and 2015 went back to them and revised them. Seyfarth and Cheney, two of the original researchers, later concluded with colleagues that the meaning of these alarm calls bears little resemblance to the meaning of human words. What lets a listening monkey pick a fitting response, they concluded, is the difference in sound between the calls, possibly together with the surrounding situation. That is a smaller claim than the one the 1980 report was taken to make. The calls still guide what listeners do, but as sounds a monkey interprets in the situation at hand, rather than as words for leopard, eagle and snake.

The correction has a great deal of ground to cover. By the end of 2020, the 1980 paper had been cited almost 1,500 times. It was cited at a higher rate after 2000 than before. So the original story kept spreading while the evidence behind it was being reworked. It could be some time before news of the revision reaches all the different fields where the original version took hold.

Sources

Do any animals have language?

No animal communication system studied so far has been shown to be language in the full sense. Several animals have individual pieces of language. Researchers still disagree about exactly where the line between communication and language should be drawn. One thing a full language lets you do is make up brand-new sentences and understand ones you have never heard before. So a speaker can say things nobody has ever said.

Since 1960, one of the most influential ways to draw the line between communication and language has been a checklist. In 1960 the linguist Charles Hockett listed 13 “design features” of language. They include having words with specific meanings, and being learned rather than inherited. Many of Hockett’s 13 features now need major revision.

One ability, complex vocal learning, shows how unevenly the pieces of language are spread among animals. Complex vocal learning means changing the sounds you produce because of what you have heard, rather than only using the calls you were born with. Copying a sound you hear is one way to do it. Humans depend on complex vocal learning. A child learns the words of the language spoken around her by listening and copying. Many birds, whales and seals share this ability. As of 2010, complex vocal learning had not been found in chimpanzees, or in any other non-human primate. No non-human primate has been observed spontaneously producing even a word of the local human language. Some parrots, by contrast, pick up spoken words without being taught. Chimpanzees are our nearest cousins. So on complex vocal learning, we have more in common with a parrot than with chimpanzees.

Humpback whales show vocal learning at work in the wild. Male humpbacks sing during the breeding season, and at any one time and place every breeding male sings the same song. That shared song keeps changing, typically by about 40 percent in a year. The males change with it together, so the population’s song this year is not the song it sang last year. Humpback song is one of the least disputed cases of a learned song: the whales are copying one another rather than singing something they were born knowing. Off eastern Australia, the humpbacks’ song once changed quickly to match a very different song from Australia’s west coast. The change apparently came after a few whales moved from one coast to the other. As of 2010, changes like these in humpback song were among the clearest evidence available that animals have cultures.

Other species hold other pieces, and the pieces are design features from Hockett’s list. These are features animals turn out to share with human language. Two of them show how scattered the sharing is.

One is duality of patterning, which means building meaningful units out of meaningless ones. Chickadees and chestnut-crowned babblers combine sound elements that mean nothing on their own, building calls that do specific jobs. Human languages build words out of meaningless sounds in much the same way.

Another is displacement, communicating about something that is not here and now. A honeybee dances to tell her hive-mates where to find food that none of them can see.

Marmosets and elephants add a less settled case: they use calls that seem to pick out particular members of their group.

No animal other than humans has been shown to have language in its full human form, and different species turn out to hold different pieces of it. No other animal has been shown to do what humans do: fold all of these pieces into one system flexible enough to express any thought at all.

Sources

What are Hockett's design features of language, and how many of them do animals have?

Hockett’s list has 13 features. Most of them turn out to be far more common among animals than he expected. Charles Hockett was an American linguist. In 1960 he broke language into a list of parts, so that other species could be checked against the list.

Here are the 13 design features:

Vocal-auditory channel — language travels by voice and ear.
Broadcast transmission — anyone close enough can hear what is said.
Rapid fading — the sound fades fast, so it does not clog the airwaves.
Interchangeability — any speaker can also be a listener.
Total feedback — speakers hear everything they say.
Specialization — the signal works on the listener’s senses and mind, not by raw physical force like a push or a bite.
Semanticity — some units, such as words, have specific meanings.
Arbitrariness — those meanings are mostly arbitrary: a word does not have to resemble what it means.
Discreteness — each thing said differs from every other by at least one distinct feature.
Displacement — you can talk about the past, the future, or things far away.
Productivity — new sentences can be made up and understood.
Duality of patterning — meaningless sounds combine into meaningful words, and words combine into sentences.
Cultural transmission — languages are learned, not inherited.

Hockett expected the first nine features to be shared with other animals. He treated the last four as the key steps toward full human language.

Some of the features have since needed correcting, because of what is now known about human language itself. Hockett treated language as spoken. But sign languages are full languages too. And words that resemble what they mean turn out to be common in languages, both spoken and signed.

So how useful was the checklist? The checklist shaped the study of animal communication far more than it shaped linguistics. Researchers are still revising the checklist more than six decades later. As for how many of the features animals have, most of the 13 have now been found in one animal or another. But no animal has been shown to have language in its full human form. The features are spread across different species rather than gathered in any one.

Sources

If animal sounds aren't language, why do animals make them?

Because a sound that changes what another animal does can pay for itself. A signal lasts in evolution when it usually helps both sides: the sender sends because the receiver responds, and the receiver responds because responding is usually worth it.

The work those signals do sorts into a few broad kinds.

Settling disputes without a fight. Most species carry a range of signals for conflicts over food, mates and space, usually with more signals for aggression than for backing down. In many cases those signals settle the dispute, and the fight never happens.
Finding and courting a mate. Both sexes presumably want signals that show the species and sex of a possible mate, and that honestly show readiness to breed, so mating can go ahead efficiently.
Holding a group together. Animals that live in groups signal to stay together, to coordinate movement, and to organize the things they do as a group.
Warning about a predator. An alarm signal is given by one animal to warn the others nearby that it has detected a predator.
Announcing food. Food signals tell other foragers that food has been found, and sometimes where it is and how good it is.

Notice how that list is built. Every item is named for the job the signal does — the situation it is used in, and what it achieves there. That is one natural way a species’ signals get sorted, and it works well, because both the situation and the listener’s response can be watched, counted and tested. Play the sound, and see what the listeners do.

But that gives you the job a call does, not the thought behind it. Some alarm calls send listeners fleeing for cover — that tells you what the call is for. It does not tell you whether the caller meant to warn anyone, or what the sound stands for in the animal’s own mind. What a call is for can often be measured. What it stands for in the animal’s own mind, mostly, cannot.

Sources

How has the technology for studying animal communication changed over time — and what was the biggest breakthrough?

The tools for studying animal communication fall into two groups. Group A records the raw material. Group B makes sense of what was recorded. The biggest breakthrough, the sound spectrograph, came from Group B.

Group A — tools that record the raw material. These capture a signal and keep it. Working out what it means is a separate job.

Writing it down. Before sound recording, the only record was a description. Bird songs were written as words that imitated the sound, or as notes on a musical staff. One 1925 rendering of the corn bunting’s song reads “Tic-tic-tic-tic-a-tee’ze.” A written version works as a memory aid, but it is not suitable as a scientific description.
Sound recording. Recording kept the sound itself instead of a description. Portable magnetic tape recorders, from the 1950s, made it possible to collect wildlife recordings all over the world in a systematic way.
On-animal recording tags. A recorder small enough to ride on an animal’s body records the animal’s calls and movements at the same time. So each sound comes with a record of what the animal was doing when it called

Group B — tools that make sense of what was recorded. These turn a recorded signal into something that can be measured, tested or sorted.

The sound spectrograph. It turns a recorded call into a kind of graph called a spectrogram. Time runs from left to right along the bottom. Frequency — how high or low the sound is — runs from bottom to top. Differences in color or darkness show where the sound’s energy sits at each moment.
The playback experiment. Researchers record an animal’s call. Then they play it through a loudspeaker to other animals. They do this at a moment when whatever normally causes the call, such as a predator, is not there. If the listeners still react as though the predator were there, the reaction comes from the sound alone.
Computer analysis. This is where AI enters the story. Earlier computer methods simply measured sound, for example by breaking a call into its separate frequencies. Machine learning, a form of AI, came later. In its most common form, machine learning learns to recognize calls from example recordings that people have labeled; other methods find patterns in recordings nobody has labeled. It now identifies species by their calls with close to human accuracy. The spectrograph and the playback experiment do not use AI; only machine learning does. So far, the main job of machine-learning software has been finding who is calling. That job is now changing toward working out what the callers are saying

So which mattered most? Sound recording, the spectrograph and the playback experiment were all real turning points. Playback made meaning testable. But the spectrograph was the real breakthrough. It was the first instrument that let researchers plot a communication signal continuously, frequency against time, and measure it. Together with the tape recorder, it opened the study of bird song much as the first microscope opened a new world of organisms.

Sources

How did sound recording change the study of animal communication?

Sound recording made the sound itself the record. Once a call could be stored, it could be replayed and examined long after the animal had gone quiet.

The earliest known animal recording was made by Ludwig Koch in Germany in 1889. It was of a captive bird called a shama, and it was made on a wax cylinder. The earliest recordings of wild birds were probably of a song thrush and a nightingale. Cherry Kearton made them in England in 1900, also on wax cylinder. The pioneers who followed cut their recordings directly onto discs. They were Koch himself, working in Germany, Belgium and Britain; Carl Weismann in Denmark; and A. R. Brand, Paul Kellogg and A. A. Allen in the United States. Recording onto discs needed bulky, awkward equipment, which greatly held the work back.

Albert R. Brand made his money on the New York Stock Exchange through a kind of trading called arbitrage. He quit that trading months before the 1929 stock-market crash. In 1930 he took up ornithology, the study of birds, at Cornell University in the United States, under the supervision of Arthur A. Allen. Brand recorded bird songs onto movie film and transferred the sounds to phonograph records. He had a Fox-Movietone film recorder altered so it could pick up the higher frequencies of bird song. Reading a film recording back was tedious work. The sound sat on the film as tiny lines running across the edge of the strip. Brand examined these lines under a low-power microscope. The speed of the film was known. So counting the lines per unit of time gave a figure in cycles per second, meaning vibrations per second. From those figures he could draw a graph of a song. Pitch ran up one axis, and the length of notes and silences ran along the other. In 1935 he published the first visualizations of songs taken from a film soundtrack. He eventually recorded the voices of more than 300 species. Brand was persistent. Once, working with A. O. Gross of Bowdoin College in Maine, he used short-wave radio to send the calls of Leach’s petrel (Oceanodroma leucorhoa) from Kent Island. The calls went to his sound truck on Grand Manan Island, 3 miles (about 4.8 kilometers) away.

Magnetic tape recording devices were invented during those same years. Fritz Pfleumer took out the first patent on audiotape in 1929. In 1935 a German electronics company produced a tape recorder and player called the magnetophon. The following year, the magnetophon was used to make the first public tape recording, of the London Philharmonic orchestra at Ludwigshafen. What changed the scale of field work was the portable magnetic tape recorder of the 1950s. It made it possible to collect wildlife recordings all over the world in a systematic way. Two other inventions helped: the parabolic (dish-shaped) reflector and the gun microphone.

Recordings then began to pile up into collections open to anyone with a serious use for them. Paul Kellogg and his colleagues built a collection at Cornell that became the Library of Natural Sounds, the first wildlife sound archive in the world. By 1980 the Library of Natural Sounds held about 40,000 tape recordings of some 4,000 species, mostly birds, from all over the world.

Sources

What are on-animal recording tags, and how are they attached to wild animals?

On-animal recording tags are recorders that travel on an animal’s body, rather than sitting in the water nearby. On whales, the tags are held on by suction cups. A tag is either placed from a boat with a long pole, or pressed on by a drone. Putting the microphone on the animal matters. An underwater microphone in open water tells you that a call happened. A tag helps tell you which animal made the call and what that animal was doing at the time. Tags like these are not only for animals at sea. Miniature microphones have been fitted to collars and small backpacks on small land mammals. One study used collar microphones to record whole days of calling by wild chipmunks.

One such tag is the DTAG, short for digital acoustic recording tag. Mark Johnson and Peter Tyack developed it in 1999 at the Woods Hole Oceanographic Institution in the United States. The DTAG has a built-in hydrophone, which is an underwater microphone. It also carries sensors that track which way the animal’s body is pointing. The whole DTAG is about a liter (roughly a quart) in volume, and it weighs 500 grams (1.1 pounds) dry. The DTAG’s sensors track the animal’s orientation in three dimensions. They do it quickly and finely enough to capture individual tail strokes. The sound and the sensors record in step with each other. So the timing between a call and a movement can be worked out precisely. Each time a DTAG is put on an animal, it records for up to 12 hours.

The DTAG’s attachment was deliberately made non-invasive, to disturb the tagged animal as little as possible. At least two suction cups are needed to hold the tag rigidly enough for its sensors to track the animal’s orientation accurately. The tag was placed on the whale from a 12-meter (about 39-foot) carbon-fiber pole. The pole rested in an oarlock on the front (bow) of a boat and stuck out over the water. This setup let the operator reach the whale’s back without leaning out over its tail. Suction cups tend to leak and let go within hours. So the attachment carries a small pump that keeps the suction in the cups. The pump is driven by the pressure changes of the whale’s own dives. The attachment also carries a deliberate release. A nickel-chromium wire seals a valve in each air line. When a clock circuit signals, the wire corrodes away in seawater.

A newer open-source tag for sperm whales was described in 2025. It sticks on with four soft silicone suction cups, which do not pierce the skin. It carries three underwater microphones that record in step with one another. Three microphones matter rather than two. With only two microphones, the timing method used to find the direction of a sound cannot tell apart two possible angles. The two angles sit mirrored on either side of the line joining the two microphones. A third microphone removes that left-or-right confusion. The three-microphone design is intended to help pull apart the calls of whales calling over each other. In the tests reported in 2025, the direction-finding was tried on a loudspeaker rather than on whales. The 2025 tag can be put on from a boat with a long pole, or flown out by drone. The drone is a small racing drone, waterproofed and made to float. It comes down and presses the tag onto the whale’s back in a single tap.

Sources

What is a sound spectrograph, and how did it let scientists see sound?

A sound spectrograph is a machine that turns a sound into a picture you can measure. Scientists and engineers at Bell Telephone Laboratories developed it around the time of the Second World War. One of its first purposes was to help with wartime code breaking.

The picture the spectrograph draws is called a spectrogram. A note that slides from a higher frequency to a lower one appears as a band sloping down the page. A rough, noisy burst, with many frequencies sounding at once, appears as a smear. The whole sound is laid out where the eye can compare one part with another.

Seeing sound this way changed what researchers could do. Bird song had been described in words and written in musical notes for centuries. The spectrograph let researchers measure the fine structure of a signal instead of relying on the ear alone. Ornithologists, the scientists who study birds, took up the spectrograph fast. In 1954, W. H. Thorpe used spectrograms to lay out the stages by which a young chaffinch learns its song.

Author’s note: When you identify a bird by its sound in the Merlin Bird ID app, you are watching a spectrograph at work — the app shows the sound as a spectrogram while it listens.

Sources

What is a playback experiment, and how does it work?

A playback experiment tests what a call means to the animals that hear it. Researchers have used playback for a long time to explore what animal signals do and mean. The playback study that became the textbook example of meaning in an animal call was published in 1980.

The signal played back is usually a sound recorded from an animal, such as a call. Researchers play the recording through a loudspeaker to other members of the same species, in the place where those animals live. Then they watch how the animals respond.

The key is timing. The call is played at a moment when whatever normally causes the call is absent. An alarm call, for instance, is normally given when a predator appears. Suppose researchers play the alarm call with no predator around. If the listeners still react as though a predator were there, the reaction comes from the sound alone. The reaction shows what the call tells the listeners.

The most common kind of playback experiment is called passive playback. It uses fixed recordings and broadcasts them at random times set in advance. Researchers then measure how the animals move, where they look, and whether they call back. Passive playback is also called non-interactive, because the recording never changes in response to what the animals do. In interactive playback, by contrast, the broadcast is adjusted to the animal’s replies. In a 2023 study in Southeast Alaska, researchers played a single recorded humpback whale contact call, known as a “whup.” A contact call is the kind animals use to keep in touch with their group. They timed the broadcasts to answer the calls of an adult female humpback known as Twain. For 20 minutes, Twain called back, taking turns with the broadcasts. The researchers describe their results as preliminary.

The textbook example is the 1980 vervet monkey experiment, which used exactly that timing. The alarm calls were recorded while a leopard, an eagle or a snake was present. They were then played back with no predator anywhere near, so any reaction had to come from the sound of the call.

Sources

How do computers analyze animal sound at a scale people cannot?

Computers analyze animal sound at this scale with machine learning, a form of AI. In the usual approach, the software learns what a call looks like from example recordings that people have labeled. Then it scans hours of audio for the same pattern.

Computers became necessary because of the sheer amount of sound now being recorded. Recorders can run on their own for very long stretches. Self-running recorders, called autonomous recording units, now sit in habitats from forests to coral reefs. They record continuously with nobody present. In the ocean, recorders anchored to the seafloor are left in place for months or even years before being collected. People also upload their own recordings to citizen-science platforms, websites where members of the public share data for research. Vast amounts of this recorded sound come unlabeled. Machine-learning software can now work through thousands of hours of these recordings with little human help. It can pick out which species are calling, how many calls there are, and even clues about behavior. Software that finds the start and end of each sound automatically now exists. Even so, as of 2025, cutting a field recording into its separate sounds is commonly still done by hand. A person draws boxes around each sound on a spectrogram.

One such tool is a type of machine learning called a convolutional neural network. It treats a spectrogram, which is a picture of a sound, the way image software treats a photo. It learns patterns straight from the labeled examples, instead of from rules someone wrote in advance. Convolutional neural networks now identify species by their calls with close to human accuracy. Across many studies, their accuracy is between 89 and 97 percent. The exact figure depends on how the spectrogram is drawn.

A trained network can also be reused instead of built from scratch. A system already trained on a large collection of bird sound can be adjusted with a small amount of new material. It can then be put to work on a new habitat, on new species, or even on a different group of animals. In practice, as few as 20 to 30 labeled examples per category may be enough for a respectable result. BirdNET, free software that identifies birds by their calls, is built on a convolutional neural network.

So far, these computer methods have mainly found and labeled calls. In other words, they have worked out who is calling. Working out what the animals are saying is where the field is heading.

Sources

How much of animal communication have scientists actually decoded — including what AI has newly revealed?

Less than you might expect. For this comparison, animal communication can be examined at four related levels: recurring patterns, broader structure, meaning, and syntax. These are not four separate and equal parts of language: pattern is a general descriptive term, while syntax is one particular kind of linguistic structure. Progress across them is very uneven. Scientists have identified many patterns and mapped substantial structure in a few species. Far fewer signals have been experimentally connected to specific meanings, and evidence for human-like syntax remains sparse and contested. As of 2026, computer models had therefore decoded mainly patterns and structure, not meaning or syntax.

Patterns are where much of the recent progress lies. Software has revealed previously undescribed features in sperm-whale codas, evidence that mouse ultrasonic calls may form a continuum rather than separate types, receiver-specific information in marmoset and elephant calls, and previously unclassified orca calls.

Structure has been mapped in considerable detail for a few species. Sperm-whale codas, for example, vary along four features that combine to produce a much larger repertoire than previously recognized. However, the researchers explicitly studied how the codas are constructed, not what they mean.

Meaning has been decoded in far fewer cases. The honeybee waggle dance conveys information about the direction and distance of a resource. Vervet monkey alarm calls produce different responses associated with different dangers, although their original interpretation as word-like signals was later scaled back. Under one proposed set of criteria, the dolphin signature whistle currently provides the strongest case for a name-like call. Claims about meaning generally require behavioral evidence, such as controlled playback experiments, rather than recordings alone.

Syntax remains the least securely established of the four. A critical 2021 review concluded that evidence for word-like semantics and syntax in animals’ natural communication was sparse and weak.

So scientists have made substantial progress in finding patterns and describing structure. A limited number of signals have been connected experimentally to particular information or responses, but AI has not yet translated a natural animal communication system, and evidence for word-like meaning and human-like syntax remains limited.

Sources

What do "meaning," "pattern," "structure" and "syntax" mean in language?

In short: a pattern is anything that recurs in recordings, structure is how a signal system is built, meaning is what a signal does to the animal that hears it, and syntax is combining words into larger meaningful wholes. Mixing these four words up makes the research sound either more or less impressive than it is.

A pattern is the loosest of the four: something that recurs in the recordings, a sound or sequence that turns up again and again. Software can find a pattern without knowing what it is for. Finding patterns is what these programs are good at, and what they have produced is mostly patterns rather than meanings.

Structure is more specific. Working out the structure of a signal system means three things: what the separate units are, how those units combine into longer sequences, and which animal answers which other animal. Sperm whales, for example, exchange click patterns, both taking turns and overlapping one another. That observation is structure.

Meaning is the hard one. The meaning of a signal is what it does to the animal that hears it. A recording on its own does not show that effect. Meaning has to be worked out by observation and experiment. Researchers broadcast a signal and check whether the animal responds differently, and fittingly, to signals from different situations. For example, a listener flees from an alarm call, or approaches or calls back to a contact call, the kind animals use to keep in touch with their group.

Here is the distinction the whole subject turns on. Picture a page of writing in a script you cannot read. You can count the marks, note which marks follow which, and see which writer answers which other writer. All of that is structure. You can have every bit of that structure and still know nothing about what the writing says. What the writing says is meaning.

Syntax is narrower again: combining separate words into larger meaningful wholes. As of 2010, whether animals combine words this way remained highly contested.

Sources

How does the honeybee waggle dance work, and how was it decoded?

The waggle dance is how a honeybee tells her hive-mates where food is: how far away it lies and in which direction. It was decoded by watching bees dance through the glass wall of a hive, and then checking which feeder newly arriving bees flew to.

A forager that has found a good patch of flowers dances on the vertical surface of the honeycomb, in the complete darkness of the hive. Other workers gather around and follow her. If the food is fairly close — say within 50 meters (55 yards) of the hive — she performs a round dance. If it lies farther off, she performs a waggle dance instead. The part that carries the information is the waggle run: a straight walk across the comb with her abdomen shaking from side to side. How long the waggle run lasts tells the other bees how far to fly. Roughly every extra 75 milliseconds of waggling stands for another 100 meters (about 330 feet) from the hive. Distances up to about 10,000 meters (6.2 miles) are encoded this way. Direction comes from the angle of the run. Outside the hive, a bee steers by the sun. Inside the dark hive, the dancer uses gravity instead. The angle between her waggle run and straight up the comb matches the angle between the food and a line from the hive to the sun. A run pointing straight up means fly toward the sun. A run 20 degrees to the right of vertical means fly 20 degrees to the right of the sun.

It took Karl von Frisch years of experiments to decode the dance. He marked individual bees with dots of paint and trained them to visit feeding stations stocked with strong sugar solution. Then he watched their dances through the glass walls of an observation hive. The dances changed in predictable ways with the distance and direction of the feeder. The key test came next. He trained scout bees to one feeding station. Seven feeders were spread around it, all stocked with equally attractive sugar water, and he collected the new bees that turned up at them. Most of the new bees arrived at the feeder in line with the station the scouts had been visiting. A matching test did the same for distance. Scouts were trained to a feeder 750 meters (about 820 yards) from the hive. Far more new bees were caught at the two feeders nearest that distance than at any other feeder. Von Frisch shared the 1973 Nobel Prize with Konrad Lorenz and Niko Tinbergen.

The waggle dance is a well-worked-out case of meaning in an animal signal. Its parts map onto the direction and distance of food. The dance is not the whole message, though. There is some evidence that the bees following a dancer also learn the odor of the food from her body, touching her where that odor is most intense: at her mouthparts and hind legs.

Sources

Do dolphins use signature whistles as names?

Possibly. Of all the animal calls proposed as names, the dolphin signature whistle makes the strongest case, but the question is not closed. A 2026 review applied one recently proposed set of tests for what counts as a name. Under those tests, it concluded that bottlenose dolphins (Tursiops species) currently provide the strongest case to date.

What the evidence shows:

Each bottlenose dolphin develops its own signature whistle, set apart by the whistle’s pattern of rising and falling pitch.
People can match whistles to the right dolphin by eye on a spectrogram (a picture of a sound), with an average accuracy of 94.5% when choosing among 20 possible dolphins.
Dolphins produce not only their own signature whistles but other dolphins’ whistles too.
A dolphin that hears a copy of its own whistle played back calls straight back, and does not respond that way to other whistles. Computer-made whistles give the same pattern of results, even with the cues of the calling dolphin’s voice stripped out. So the pitch pattern itself appears to be what matters.

What is less certain: the evidence is clearer on two points, that a whistle picks out one individual and that caller and listener share its meaning. On whether the whole dolphin community shares that meaning, the evidence is only suggestive.

Sources

Jaakkola, K., King, S.L. & Bruck, J.N., “Do animals use names? Conceptual and empirical criteria”, Trends in Cognitive Sciences, Opinion, online 22 July 2026 (open access, CC BY-NC-ND) — “Conceptual criteria for names”; “Bottlenose dolphins”; “Concluding remarks”; Table 1 and footnote d

What have scientists worked out about the structure of sperm whale clicks?

A lot about how sperm whale calls are built, and very little yet about what they mean. Sperm whales communicate in codas. A coda is a short burst of clicks, generally under two seconds long, that whales in a group trade with one another.

What could be decoded is how the calls are built. A 2024 study worked through 8,719 codas recorded off the Caribbean island of Dominica. The study found that a coda varies along four features, and that the four features combine freely:

Rhythm — the pattern of spacing between the clicks.
Tempo — how long the whole coda lasts.
Rubato — a gradual stretching or shrinking of coda length across a run of codas from one whale.
Ornamentation — an extra click added at the end of a coda.

Rhythm and tempo were already known. Rubato and ornamentation were newly identified. Whales trading codas were shown to sense and act on all four features. Because the four features combine, the number of distinct codas comes out at nearly ten times what was previously believed. Before the 2024 study, researchers already knew that some codas reveal which whale is calling. Almost everything else about the coda system was unknown.

What could not be decoded is what the codas mean. The 2024 study looked at how codas are built, not at what they mean. It offered no account of what any coda means. It also ran no playback experiments, in which recorded calls are played to animals to see how they respond. The researchers say that working out meaning will need interactive playback experiments with whales in the wild, in which the broadcast is adjusted to what the animal does.

Sources

What did AI reveal about the ultrasonic calls of mice?

That mouse squeaks may not come in separate types at all. Mice make ultrasonic calls: squeaks too high-pitched for human ears to hear.

In a 2021 study, researchers gave spectrograms of the calls (pictures of each sound) to a machine-learning program. The program worked out its own description of each call, instead of using a short list of measurements chosen by people. The program’s description captured more of the information in each call than the hand-picked measurements did.

When the calls were arranged by the program’s description, they did not appear to fall into separate types. Instead, the mouse calls spread across a continuous range. In this respect they sat closer to random noise than to zebra finch song, which does form clear groups. This result challenges a common assumption that mouse calls come in distinct types. That assumption is built into software used to sort mouse calls.

The authors also noted limits to their conclusion. Whether a set of sounds truly falls into groups cannot be answered definitively. And mice might still hear the range of calls as separate categories. One of the scientists who reviewed the study before publication disagreed that the calls form a continuous range.

Sources

Do marmosets and elephants use name-like calls for each other?

Maybe. Marmosets and elephants both appear to use calls aimed at particular individuals. But as of 2026 the evidence is mixed, and in neither species has anyone yet found the part of the call that would work as the name.

In the marmoset case, a machine-learning program matched contact calls, known as phee calls, to the marmoset each call was aimed at. It did better than chance, though only modestly. Marmosets were also more likely to respond to calls previously aimed at them than to calls aimed at other marmosets.

The elephant case runs the same way. A machine-learning program matched African elephant rumbles to the elephant each rumble was aimed at. It did better than chance, though with relatively low accuracy. Elephants responded more strongly to rumbles previously aimed at them than to rumbles aimed at other elephants.

Three things keep the marmoset and elephant cases from counting as names. First, software did the sorting, and nobody has yet identified which features of the call carry the “name.” Second, no evidence has been put forward that the call stands for the individual in the animal’s mind. Third, another explanation fits, at least for marmosets. A caller that partly copies a call it has heard would produce the same results.

So these are name-like calls, not established names.

Sources

Jaakkola, K., King, S.L. & Bruck, J.N., “Do animals use names? Conceptual and empirical criteria”, Trends in Cognitive Sciences, Opinion, online 22 July 2026 (open access, CC BY-NC-ND) — Highlights; “Common marmosets”; “African elephants”; “Partial vocal matching”; Table 1 and footnotes b and c; “Concluding remarks”

What are the limits of using AI to decode animal communication — and could some signals remain untranslatable?

AI is very good at one part of decoding animal communication: finding patterns in huge amounts of recordings. It can pick out which species or which individual is calling, count calls, catch clues about behavior, and even generate new calls in the same style. The patterns it finds give researchers ideas about what signals might do, and those ideas can then be tested in experiments. What AI cannot do on its own is tell what the patterns mean. Five limits stand in the way.

Naming the situation. To connect a call with a meaning, people label the situation each call was made in. They can only label situations that humans have names for. This means the AI can only link calls to the kinds of meaning people already thought to look for.
Checking the results. The software latches onto anything that happens to line up in the data. Many times it has picked up background noise instead of the signal. So every group of calls the software finds is only a hypothesis until that group is tested on the animals, for example with a playback experiment. In many cases the animals give no clear response that can be measured. A finding may then never be confirmed or ruled out.
Missing channels. Animals signal through sound, sight, touch, smell and electric fields, often several at once. Some of that signaling lies beyond human senses. Software fed only on sound misses the rest. If a played-back call draws no reply, the missing piece might be a scent that was never recorded. Even the honeybee waggle dance probably carries touch and sound cues about the food that the eye misses. So AI may be working on only part of a message.
Uneven data. Most of this software remains heavily dependent on recordings people have already labeled. These labeled recordings are spread unevenly across animal groups and regions. Software trained in one place can also struggle in a different habitat, on a different recorder, or against a different background of noise. This means the software works best for well-recorded species and places, and less well for the rest.
What animals do not talk about. In 2023, the researchers Yossi Yovel and Oded Rechavi argued that we probably cannot ask about anything a species does not signal about. For example, if cats do not discuss their feelings with each other, we could never ask a cat how it feels. They add that animal signals probably also cover things wholly alien to us. If they are right, some meanings lie outside anything humans can label.

These limits are not all of one kind. Uneven data is the practical one: more recordings, from more places, would ease it. Missing channels look practical too, since more kinds of data could be recorded. But it is not yet clear which kinds would be needed. Yovel and Rechavi count that gap among the obstacles they consider built in, because what we fail to record follows from the limits of our own senses. As Yovel and Rechavi see it, the rest are built in: naming the situation a call was made in, checking what the software finds, and the fact that animals may not signal about the things we would want to ask them. And a single signal can run into several of these limits at once — made in a situation people have no name for, drawing no response anyone can measure, carried partly through a channel nobody records, and coming from a species with few recordings.

So could some signals remain untranslatable? Possibly. Yovel and Rechavi hold that the built-in obstacles would not vanish even if the power of AI grew a millionfold. They make that argument about talking with animals. The argument applies to decoding only if the same obstacles carry across. Still, much is unknown, and some species may yet prove to communicate in richer ways than anything shown so far.

Sources

How are AI, machine learning and deep learning related — and how do they work with this information to understand what animals say?

AI, machine learning and deep learning nest inside one another. AI is the broadest idea: getting computers to behave intelligently. Some AI involves no learning at all. People type facts in by hand, and the computer reasons from those facts with logical rules. Machine learning is the kind of AI that gets its own knowledge by finding patterns in raw data, instead of being handed the facts. Deep learning is a kind of machine learning that builds complicated ideas out of simpler ones, layer on layer. For example, a system can recognize a person in a photo by combining edges into corners and outlines, and then combining those corners and outlines into the figure of a person.

Given enough recordings of animal sounds, these methods are put to work in a few ways:

Learning from labeled examples. People label recordings: this is a robin, this is an alarm call. A machine-learning program then learns which features of the sound predict each label. This is how species-identification tools work.
Finding patterns without labels. Other programs sort unlabeled recordings on their own. The aim is to discover which call types an individual animal, a group or a population uses. The mouse-call study by Jack Goffinet and colleagues worked this way.
Learning from huge mixed collections. Self-supervised deep learning needs neither labeled recordings nor anyone deciding in advance which features of a sound matter. It is the basis of foundation models: large models trained once and then used for many tasks. NatureLM-audio, from the Earth Species Project, is one.
Linking sound to context. Some methods combine different kinds of data: the sounds, the animals’ movements, who is calling, and what is happening around the callers.

Which of these methods is most useful? Machine learning is now central to computer analysis of animal sound, especially for identifying species. For understanding what animals tell each other, researchers writing in 2023 pointed to self-supervised deep learning as of interest. They also judged that methods linking sounds to behavior and context seem especially promising. Whichever method is used, what the software produces is patterns. Those patterns still have to be tested on the animals themselves.

Sources

Who is leading the research on AI and animal communication — and what have they actually achieved?

Research on AI and animal communication is spread across several organizations, including these:

Project CETI (the Cetacean Translation Initiative) works on what sperm whales say to one another. So far it has mapped how the whales’ click calls are built. It has also built an open-source recording tag that rides on a whale’s back. It has not yet worked out what the calls mean. Finding that meaning is what the project is working toward
The Earth Species Project builds AI meant to work across many kinds of animal at once, rather than building a separate program for each group of animals. Its program NatureLM-audio listens to an animal recording and answers plain-language questions about that recording. A person can ask, for example, which species is calling, or whether more than one animal can be heard
The Cornell Lab of Ornithology, with Chemnitz University of Technology in Germany, builds BirdNET, which names the bird species behind a recorded call and is free to use.
The Coller Dolittle Prize, run by Tel Aviv University in Israel with the Jeremy Coller Foundation, offers an award of US$100,000 for work on understanding and communicating with animals, and awarded it in 2025 and again in 2026.

Sources

What does Project CETI do, in more detail?

Project CETI is a nonprofit working to understand what sperm whales say to one another. It is based in New York, in the United States, and on the Caribbean island of Dominica, where its field work happens. Sperm whales communicate in codas: short bursts of clicks that they trade back and forth. Those codas are what the project studies.

The project has two published results worth setting out. The first is a map of how codas are built. Researchers working with the project found that codas vary along four features, and that those features combine. Because the features combine, the number of distinct codas is far larger than previously believed. That map covers structure, not meaning. The second result is a recording tag worn by whales, called the CETI Bio-Logger. The tag’s designs, build details and code are published free, so any research group can build one or improve on it. The tag has already ridden on sperm whales off Dominica, staying attached through deep dives.

The project is now working on the meaning of the codas, which the map leaves open. The researchers who made the map say that finding meaning will need playback experiments with wild whales.

Sources

What does the Earth Species Project do, in more detail?

The Earth Species Project is a research group in Berkeley, California, in the United States. It builds AI tools meant to study animal communication across many species at once. Its main tool is NatureLM-audio, released in 2024. NatureLM-audio is a foundation model: one large model trained on a broad mix of data and then used for many different jobs, instead of a separate small model for each job. It was trained on animal sounds, human speech and music together. The hope is that what it learns from speech and music carries over to animal sounds, where recordings that people have already labeled with what is in them are much scarcer.

Using NatureLM-audio means giving it a recording and typing a question in plain language, and the model types an answer back. The question can be which species is making this sound, whether this is a song or a plain call, or whether more than one animal can be heard.

Its results so far are mixed. On a set of standard tests it had not been trained for, NatureLM-audio outperformed the four other systems it was compared with on most of those tests. It was also given species it had never heard before, and asked to answer with each species’ full scientific classification. It named the right species about a third of the time: far better than chance, and still wrong two times out of three. The model, its code and its test set are all released openly, so anyone can run the model, check its results, or try to beat them.

Sources

What does BirdNET do, in more detail?

BirdNET is free software that identifies which bird species is calling from its sound. Two groups make it jointly: the K. Lisa Yang Center for Conservation Bioacoustics at the Cornell Lab of Ornithology in the United States, and the Chair of Media Informatics at Chemnitz University of Technology in Germany.

On the phone app, you record a few seconds of sound, and the app sends the clip to a BirdNET server along with the date, time and place. A species name comes back, with a score for how sure the software is. You can then tell the app whether it named the bird correctly. The recordings are kept for research, and anything that identifies the person who made them is removed.

BirdNET began with birds of North America and Europe and now covers more than 6,000 species worldwide. Because it asks for nothing but a phone and a few seconds of sound, it drew far more people than a project built for experienced birdwatchers. In 2020 the app had more than three times as many participants as eBird, the Cornell Lab’s older checklist project.

Sources

What is the Coller Dolittle Prize, in more detail?

The Coller Dolittle Prize is an annual award of US$100,000 for work toward understanding and communicating with other species. Tel Aviv University in Israel runs it together with the Jeremy Coller Foundation. Its organizers drew on the Turing test, asking the same kind of question about other species: can software hold an exchange with one?

The 2025 challenge prize went to Laela Sayigh, for finding that wild dolphins share whistle types other than each dolphin’s own signature whistle. The 2026 challenge prize went to Julie Elie, for work on decoding the “language” of zebra finches, which are songbirds.

Above the annual award sits a much larger prize: either a US$10 million investment or a US$500,000 cash prize. That larger prize is reserved for a team that achieves two-way communication with another species. No winner of that larger prize has been announced. The organizers’ own account of the timing is not consistent: the prize pages name winners of the 2025 and 2026 challenges, while the prize’s home page says the first annual prize will be awarded in 2027. Entries for the 2027 challenge close on September 30, 2026.

Sources

What is the future of AI and animal communication — how might it help us understand animals?

There is no settled forecast. What can be said is what researchers are aiming at, and what they say stands in the way.

The aim has shifted. For years, AI software was aimed at a sorting job: working out which species, or which individual, is making a sound. That software is now described as moving on from asking “who is calling” toward uncovering what the callers are saying and why that matters. Two predicted advances follow from this shift. Decades of stored recordings could be used to track how a population’s calls and songs change over time, drifting slowly, converging on a shared form, or in some cases being replaced wholesale. And the sounds an animal makes could be compared with its genes, to find genes linked to song learning.

Leading bird-call software, such as BirdNET, performs well. But it needs a dedicated classifier trained for each target group of animals. A classifier is a program that sorts calls into categories. The Earth Species Project’s NatureLM-audio is a model built to work across many species at once. With animals it had never heard, it has shown what has been described as promising generalization. That is a step in recognizing those animals, not in reading what they say.

What would prove that an animal signal has really been decoded? Two tests are already in use. The first is replication: a finding has to hold when another team tests it again. The 1980 vervet monkey alarm-call result, long the textbook case, did not pass that test. The second test is strict, and was set out in 1993. A call counts as pointing at something in the world only if it draws appropriate responses from listeners who have no supporting context to help them. That bar appears to have been applied more loosely since. Analyzing large archives of recordings, on its own, is judged unlikely to achieve breakthroughs in decoding what signals do. What machine-learning methods can do is reveal patterns. Those patterns may inform hypotheses about what signals do. Those hypotheses may in turn inform the design of controlled experiments. The software narrows the search; the test still has to be run on the animals themselves.

And meaning may not emerge from this pattern-finding. As of 2026, models of animal signals have identified patterns rather than meanings. A viable translation of any nonhuman animal’s natural communication system is still described as far off. Even so, a 2023 assessment expects the leap in AI to yield a better understanding of animal communication. The same assessment holds that some obstacles are built in: naming the situation a signal belongs to, and deciding what counts as a response. Those obstacles would not vanish even if the power of AI grew a millionfold — an argument the assessment’s authors make about talking with animals.

Sources

Is it ethical to record animals and play sounds back to them — and are there rules for how far researchers should go?

There is no single yes or no, and no one global rulebook. Recording animals and playing signals back to them are ordinary parts of animal-behavior science. Written rules cover both activities: national laws, which differ sharply between countries, and guidelines from scientific societies. The sources reviewed turned up no rules written specifically for AI-assisted work with animal signals.

Listening is the gentler of the two activities. A recorder left in a forest or lowered into the ocean collects sound with no person present. Even so, round-the-clock recording has been described as a kind of acoustic surveillance. Current guidance on such recorders asks researchers to keep disturbance to a minimum. The guidance also asks them to test two things before putting a recorder out: whether the recorder is really needed, and whether the intrusion it causes is in proportion to what the study needs.

Playback carries more risk. In a playback experiment, a recorded signal is broadcast to living animals. Three risks of playback have long been flagged. The sound can distress the listeners. Playback can also wear away the signal’s meaning. Animals that keep hearing a sound may get used to it, and then fail to respond when the real signal comes. And playback can damage the standing of the animal whose voice is played. If one individual’s recordings are played too often, its group may start to ignore that animal.

AI makes all three risks sharper. An AI model may produce a call that scientists cannot confidently interpret, so nobody knows what is being broadcast. The model may produce a voice that belongs to no known individual animal, so the broadcast is not really a playback of anything. And testing an AI model means broadcasting to animals far more often, and in a wider range of situations, than a careful traditional study would. One worry was raised in 2023: broadcasting calls to wild humpback whales before those calls are understood could accidentally change how whales sing across a whole ocean basin.

Two scientific societies — the Association for the Study of Animal Behaviour (ASAB) and the Animal Behavior Society — jointly publish Guidelines for the ethical treatment of nonhuman animals in behavioural research and teaching, revised in 2026. The guidelines add to the laws of the country where the work is done rather than replacing those laws. The journal Animal Behaviour uses the guidelines to judge the papers it receives, and university and institute ethics committees use them before approving a project.

So how far may researchers go? The guidelines are general in scope, because behavioral research spans so many species and methods. Researchers must weigh the knowledge a study is likely to gain against the harm it could do — to the animals studied, to their wider population, and to the ecosystem around them. Playback can itself be the gentler choice. Where researchers would otherwise stage an encounter between live animals, the guidelines ask them to consider models of animals or recorded sound instead. In 2023, researchers working on AI decoding called for best-practice guidelines and laws, saying the ethical issues must be dealt with directly, not as an afterthought.

Sources

So was my son wrong? Less than adults usually assume. The magpies outside were passing real information to one another, and once a signal is shaped for a listener, “talking” is not a foolish word for it. Language is the higher bar, and no animal has been shown to clear it: no other animal folds all the pieces into one flexible system.

How much has been decoded? A few signals, in a few species. What has grown much faster is the map of how signals are built, which is not the same as knowing what they mean.

And my husband’s optimism? Half earned. The machines have found real structure nobody had described before, in whale clicks, in mouse squeaks, in calls that seem aimed at one animal. Meaning is the part they cannot reach alone, and that still takes an experiment with the animals themselves.

The magpies are still interrupting. We still do not know what they are saying.

This article was created through research, curiosity, and a deep love for understanding animals by Niloofar Moharrami for Nested Questions.