A forager comes back to the pitch-dark hive and runs a figure-of-eight on the vertical comb. A few minutes later, several bees that have never been to those flowers fly out and head straight for a point a few hundred metres away.
That looks more like speaking than anything else an animal does: it hands over a location the listener has never seen and that is not in front of anyone. But set the criterion for "language" at exactly that, and bees pass comfortably while a chimpanzee raised like a child and taught gestures by hand for years does not. On the same test paper, the ranking can flip.
So the effort goes into stating the question, not into collecting evidence. A dance, a videotape, and a string of bird calls played backwards each push the question along a stretch.
First, what the dance encodes. The hive is dark and the comb hangs vertical. The bee runs a figure-of-eight, and in the straight middle stretch she walks forward while waggling side to side — that stretch is the waggle run. However many degrees the waggle run tilts away from straight-up-against-gravity, that is how many degrees the flowers lie away from the sun's azimuth: vertical stands in for the sun. The longer the waggle run lasts, the further the food, at roughly a kilometre per extra second, with the conversion varying by population and region.
This was doubted for decades: might the recruits simply be following flower scent carried on the dancer? In 2005 Riley, Menzel and colleagues glued tiny radar transponders onto bees and plotted, track by track, the flights of recruits leaving after watching a dance. They flew straight out on the bearing and for the distance the dance indicated, correcting along the way for crosswind drift — to a place they had never visited. On arrival they usually had to circle and search before finding the flowers: the dance gives directions, not coordinates.
Now what it cannot encode. There are no words in this code. Five degrees off in the angle puts the meaning five degrees off; it does not thereby become a different meaning. Human speech is the reverse: swap the initial consonant of 妈 (mā, "mother") for that of 趴 (pā, "to lie face down") and you do not get a slightly-off 妈 (mā, "mother"), you get another word, with no transitional zone in between — the categorical perception from the topic on the grid the brain draws over sound.
So the bee's displacement — being able to talk about what is not present — is real, and on that one criterion it beats many mammals. But it will only ever say this one kind of thing: direction, distance, and how much the trip is worth, since a longer, more vigorous dance recruits more bees. A swarm hunting for a new nest site uses the same dance. It cannot say that yesterday's flowers were mown down, and it cannot put "that direction" together with "there is danger over there" — there are no parts to put together.
From 1966 the Gardners raised a chimpanzee named Washoe like a child and taught her manual gestures. Later came the gorilla Koko, the chimpanzee Nim, and the bonobo Kanzi — Kanzi working not with gestures but with a board covered in printed symbols, one symbol per word, each symbol called a lexigram. These projects made magazine covers at the time: apes producing hundreds of signs, and combining them.
In 1979 Terrace and colleagues went back through Nim's videotapes stretch by stretch, and the whole thing turned. Most of Nim's "sentences" imitated a gesture the teacher had just finished making, or cut in before the teacher had finished. He almost never opened a topic himself. His longest string ran sixteen words: "give orange me give eat orange me eat orange give me eat orange give me you". The length is stacked, not nested — human sentences grow longer by embedding one thing inside another, this one just appends.
None of which means apes are stupid. Kanzi's comprehension test looks quite different: 660 spoken English instructions, many of them asking for an object to be used in a way he had never seen, such as putting the toothbrush in the soda. Kanzi got 74% right; the same items given to a two-year-old child, Alia, got 65%. But that is one individual against one individual, with no way to replicate it, and the evidence is exactly that weak.
The real fault line is who reads the tape. Over a century ago a German horse named Hans appeared to do arithmetic, and was later found to be reading posture cues his questioner leaked without knowing it; the case gave us a term of art, the Clever Hans effect. Studies like these have to be run blind. Pinker quotes the recollection of the only native signer on the Washoe team — a deaf member asked to log every sign the chimp made: "they were always complaining because my log didn't show enough signs." One person's memory is not a measurement, but the problem it points at is entirely real: the same tape, scored a little more loosely or a little more strictly, gives a different conclusion.
One more thing has to be said plainly. Sign languages are full languages, with their own grammar and their own level of minimal units — the subject of the topic on sign language being a complete language. What these apes picked up was a handful of gestures borrowed from a sign language, not its grammar. Blurring the two both inflates the projects and shortchanges the languages.
Start with one that passes cleanly. In 1980 Seyfarth, Cheney and Marler hid loudspeakers in the Kenyan grass and played vervet monkeys their own alarm calls with no predator anywhere in sight. On the leopard call, monkeys on the ground ran for the trees; on the eagle call, monkeys in the canopy dropped down and into the bushes; on the snake call, they stood up and scanned the ground around them. Call matched predator class, and not by watching where a caller was looking, since the speaker was hidden. That is hard evidence of reference.
It is less tidy than the textbook picture. In 2015 the recordings were reanalysed with quantitative clustering: the three predator calls do separate, but once calls given during aggression are added in, the categories begin to overlap — females' snake, eagle and aggression calls run into each other, and males' barks mix across the leopard and the fight contexts. These calls carry both what was judged and how aroused the caller was.
Now one that goes further. In 2016 Suzuki and colleagues played Japanese great tits their own calls in the field: the ABC notes make a bird scan warily around, the D note makes it approach the source, and played together as ABC-D the bird does both. The crucial part: reversed to D-ABC, it mostly does neither. The order-sensitivity is solid and has since been run on other birds too.
Whether that amounts to composed meaning is argued over. Critics point out that the bird might simply have memorised ABC-D whole, as a third fixed signal, or be hearing two utterances back to back. Human compositionality is a procedure applied freely, computing a new meaning from any two suitable parts, and that much has not been shown in tits. The dispute is live.
So the shape of the criterion becomes clear: it is not a gate but a checklist, and something passes on every line. Displacement — bees pass. Reference — vervets pass. Order-sensitivity — Japanese great tits pass. Only openness is still passed by humans alone: the new meanings a person can build have no ceiling, because sound is first broken into a few dozen reusable parts and only then assembled into words and into sentences, which is the duality of patterning from the topic on language not being a dictionary. No non-human species has been found to have that layer.
Which is why asking whether animals "have language" slides so easily into ranking species. Asking instead what a given system can express, and what stops it, is the question that produced every experiment above.
Probably only the uses: asking questions, describing what never happened, talking about the system itself. Notably none of those live in the sounds — they can only be seen in how the thing is used, so the criterion would move from structure to use.
The question needs taking apart first. A continuous code is cheaper when one quantity stands for another: a bearing need not be quantised into words. What discreteness buys is recombination and relayability. They solve different problems, so "better" depends on which problem is on the table.
This is the part of the history that should least be skipped. An individual raised like a child and taught hundreds of gestures can neither return to a wild group nor remain suitable for a laboratory; later placement has depended on a few sanctuaries and long-term funding, and it was not in the project design from the start.