Since the 1970‘s, when the first whale songs were released to the public, we’ve been fascinated with whale communication networks. Songs of the Humpback Whale, the soundtrack to so many acid trips and massage sessions, was a massive bestseller, and used by the Judy Collinses of the world as background for folk songs. In 1979, National Geographic released a “flex-disc” of whale song with the magazine that further spread both its popularity and peoples’ connection to it. The “Save the Whales” movement soon followed: if something can sing, it deserves to live. Whale songs were an essential feature of the rising environmental movement; they created compassion for other species. The sounds of whales felt meaningful, even if the meaning wasn’t immediately intelligible.

The communication between sperm whales has never been as popular as the songs of the humpback. Sperm whales don’t sound as nice. They make a series of clicks, called codas, which are sometimes compared to an underwater zipper or to firecrackers. To me, they sound sometimes like a bowl of the world’s loudest Rice Krispies and sometimes like scraping a metal bar loudly against a railing at different speeds. It is a profoundly inhuman sound, deeply unfamiliar.

But for researchers who listen closely and analyze the codas, it is possible to identify separated, repetitive patterns in sperm whale communication. In the 70 years since researchers began this work, two things have become clear: the whales are speaking to one another — and we have only the most limited idea of what they’re saying.

The sperm whale has the world’s most powerful biological sonar system and the world’s largest brain, about five times the size of ours. Sperm whale codas, at the very least, appear like language.

But their specific functions — what one sound or series of sounds signifies — remain largely a mystery. Some codas carry information about the speaker, its family and its clan. You could say that’s a whale equivalent to “Here I am.” You could even call that a name. Sperm whales form matrilineal groups in which they care for and even nurse each other’s calves. Groups of sperm whales, or clans, have dialects with at least 20 unique codas. Calves learn to sing over a two-year period, which comes with a period of babbling, imitating the patterns and cadences of adult speech but with nonsense syllables. Sperm whales have what could be perceived as cultures — learned traits of speech shared within groups but not between groups. Some whales meet with their clans every day to click to each other, often for decades. 

It is nearly impossible not to anthropomorphize sperm whales. One researcher recently compared a mother whale clicking to younger female whales before leaving her calf in their care to a mom exchanging a few words with the babysitter before going out. The functions of whale codas and human small talk seem to be the same — at least through our human lens of understanding — even if correlations between codas and words are not exact, or even identifiable. 

All this brings us back to artificial intelligence. AI is now being used to analyze all kinds of natural sounds and decode them in a way that makes sense to people. Elephant trumpets, fish rumblings, orangutan howls and bat vocalizations are all being input into AI systems. Even tomato plants and wheat ears give off distinctive sounds that can be revealing when machine learning is applied to them. But whales are special, not only because of their enormous remove from us, their being in another oceanic world of unimaginable pressure and darkness, but also because so much of their lives — the gathering, the chatter, the society — seems so familiar.

“One of the things that blows people’s minds is that a sperm whale’s grandmother is important to them, and that where they come from and how they live their life is important to them,” says Shane Gero, scientist-in-residence at Carleton University in Ottawa and whale biologist for Project CETI. “But when we talk about similarity and sameness, we also do a slight disservice to both humans and whales.”

The problem of describing and figuring out the significance of sperm whale codas is complicated for our human brains. Human languages are broken up into words and phrases. But communication between non-human animals doesn’t seem to have the same basic structure. Trying to identify the elements that make up non-human communication, for example in chimps, has historically consisted of manual annotation by experts. It’s hard to know what parts of the whale sounds matter, especially when human prejudices about language are unavoidable. That’s why machine learning, which detects patterns human beings cannot detect, at a scale incomprehensible to the human brain, could be such a potent tool in figuring out what these clicks and pops mean. 

Project CETI is emblematic of the current moment in artificial intelligence, both in its ludicrous promise and its sudden arrival at moments of impenetrable limit and profound mystery. CETI combines machine learning with marine biology, robotics, acoustics and linguistics in an attempt not just to understand whales but to speak back to them. “We are on the way there,” says Gašper Beguš, the Berkeley linguist guiding the machine learning towards interspecies communication. “Of course, you never know before you’re there.”

One of the strangest features of artificial intelligence is that its ultimate outputs are unknowable. Nobody — not its inventors, not its most sophisticated practitioners — knows exactly how it will respond to data at large scale. Its powers are always surprising. Which is why capacities like the ability to speak with whales are even conceivable. The surprise of artificial intelligence goes both ways, too. When AI was used in the attempts to understand and contain the spread of COVID-19, it didn’t live up to its imagined potential, even though the problem of COVID-19 — which required rapidly processing vast amounts of information and making quick decisions — seemed, in theory, an ideal situation for the application of AI. It wasn’t. There were even some concerns that AI might have done more harm than good. 

At the same time, AI has replicated phenomena that verge on the bizarre, that transcend what we think of as the technological. The ability of large language models to imitate deceased people based on just a few snippets of text messages they sent in life has become so sophisticated that the European Union recently introduced legislation to curb AI practices it deems unacceptably risky including anything intended to manipulate or deceive. In artificial intelligence, the craziest things sometimes are the most likely to come true. 

Artificial intelligence, in its powerful strangeness, could be a way forward when it comes to animal communication. (It is also a dry run for what it might be like to communicate with alien species if they ever arrive from space.) Human brains are just not capable of making comparisons between animal sounds and human language. “The problem of constructing an inventory of phonetic, lexical, or grammatical units is much harder for whale communication compared to human languages,” a recent piece of CETI research explained. “Not only do we not know which acoustic features are meaningful, how they vary systematically, or which units are correlated with behavior, [but] it is also not trivial to probe for and test and verify meaningful units in their communication.” In other words, trying to derive meaning from sperm whale clicking is flummoxing. Where even to begin? What humans define as the inventory of meaningful units is a shorthand that whales may or may not agree with. 

The first need for any machine learning system is data, and so far, CETI doesn’t have enough. GPT-3 — one of the most well-known large language models trained using internet data to generate text — was trained on 175 billion parameters, or data points. GPT-3 now seems pitifully tiny; its successor, GPT-4, which came out last year, is said to have 1.76 trillion parameters. The CETI team are hoping to amass 400 million recordings of sperm whale clicks. So far, they have collected 8,719 codas from what researchers call the Eastern Caribbean 1 clan.

Which is why CETI researchers are now gathering more whale songs off the coast of Dominica with robotic fish and hybrid aerial-aquatic drones. AI is making inroads already, even with the scanty data. “We confirmed stuff that humans thought was meaningful,” says Beguš, “but we also found some acoustical properties that are meaningful that humans have not found yet.” Their most recent paper demonstrated that the number of clicks and the intervals between clicks is significant and also that spectral regularity and spectral mean — the frequency of their clicks and the consistency of their sequences — matter. “It’s a very classic case where AI is not giving you the final answer but it’s guiding you towards the clues, to where to look for an answer,” Beguš says. 

The deeper problem in understanding whale communication is not computational, but foundational. “We do not know what is meaningful in whale communication systems,” Beguš says. “What are those properties that whales are encoding their meaning into?”

The poverty of language to describe language itself is inherent. The mirror cannot reflect itself. The word “language” carries with it a host of assumptions about what constitutes sense and nonsense. The shorthand terms we use to describe how we make meaning — words, sentences, grammar — were always descriptions we knew to be shallow attempts to get a handle on unfathomably profound realities. Our languages, our ways of meaning, are intimately bound up in our biology, the specifically human way of being in the world, and that frame of reference is inescapable 

As Gero, the whale biologist at Project CETI, puts it, “The experience of being a whale is fundamentally different to the experience of being a human.” Whales are so different from us that the difference may literally be unimaginable. “The complex, multi-party nature of sperm whale vocalization, and especially the presence of vocal learning and chorusing behaviors with no obvious analog in human communication, suggests that this discourse-level structure is as important as the utterance-level structure for understanding whale communication,” the CETI study says. Is it possible to imagine a language that starts with discourse and proceeds, from there, to speaking? 

In these extensive gatherings of sperm whales off the coast of Dominica, what if meaning isn’t between them but among them? What if it’s not that an individual whale has an idea it expresses, but that there’s an idea among them that the whales join themselves to? All of this is pure hypothesis, fumbling to describe a way of communicating that is not our own. If discourse comes before grammar, is it language? Is it expressivity? What is it?

But whale language may not be the problem. Whale meaning may simply be inaccessible to us. The German philosopher Ludwig Wittgenstein famously said that if a lion could speak, we could not understand it. Artificial intelligence could prove him right.

The most prominent linguist in the world, Noam Chomsky, believes the capacity for language is a specific feature of the human brain, that language arose from a single genetic mutation possibly as recently as 70,000 years ago. He has utterly rejected primatologists who claimed they had managed to train apes on crude linguistic tasks; for him, all such communication is strictly imitation. “It’s about as likely that an ape will prove to have a language ability as there is an island somewhere with a species of flightless birds waiting for humans to teach them to fly,” he once wrote. Chomsky has rejected GPT, and other large language models with the same absoluteness. But recursion, the ability to place one linguistic component inside another, is Chomsky’s defining property of universal language reserved strictly for human beings. ChatGPT can do it easily.

Since the launch of ChatGPT, the experience of language itself has arguably changed. Overnight, activities once considered the sole province of the human mind were not. A machine can tell you where to get coffee in downtown Tokyo. A machine can compose a novel. A machine can imitate dead people. Everyone has had their “aha” moment with AI, the shock of seeing a machine do something that feels particularly human or transcendent of the human. By now, some of the shock has worn off. But the new reality we inhabit is still an immense leap from the old one: We live in a world where we are not the only linguistic intelligence. Computers can speak somewhat sensibly. 

But probably we have never lived in a world where humans were the only linguistic intelligence. One of the great hopes of CETI is that communication with whales will lead to a deeper respect for non-human life. A natural world that could “talk” with us would be different from a natural world that couldn’t.

Something human beings figured out early on about sperm whale brains is that they’re delicious. “The brains are accounted a fine dish,” Herman Melville noted in Moby Dick. “The casket of the skull is broken into with an axe, and the two plump, whitish lobes being withdrawn (precisely resembling two large puddings), they are then mixed with flour, and cooked into a most delectable mess.” If other species have linguistic intelligence, it would be inconvenient at the very least. Would we still eat steak if we could understand what a cow means when it moos?

Science has a way of humbling humanity. The exploration of space has revealed our unimaginable smallness. Life, in this entirely irrelevant corner of unimaginable emptiness, is a brutal struggle for mere survival, to no final purpose, at least according to Darwin’s theory of evolution. The lesson of artificial intelligence could be that human intelligence is just not that special. There is machine intelligence. There is human intelligence. There is whale intelligence. There are other meanings than human meanings. And there is no evidence whatsoever that a language-processing brain is the best way to understand the world. Indeed, there’s a great deal of evidence that it’s terminally unsound. Look at the lives of the language-processing brains around you.

The goal of CETI’s research is to perform playback experiments, to intervene with the whales. But any intervention would require answers to some basic questions: “Do we know what to playback? Do stimuli replicate biological signals? Can we recognize a response?” Beguš asks.

There is also the question of what we would say to the whales. What exactly do we have to communicate? What do we have to offer the whales except our apologies about what we’ve done to the ocean? Why are we trying to talk to them? Would they even want to listen?

We may never be able to cross the border between whales and humans. There is a whale-world and there is a people-world, and even if whale-world has whale language and people-world has human language, translation between those languages may be no more possible than a whale walking into a grocery store or a human diving to 1,000 metres.