How AI is decoding whale language

A single sperm whale click carries so much energy that it surpasses the force of a jet taking off, cutting through the darkness of the deep sea like an acoustic spotlight. Here you can listen to the rhythmic clicking in the original recording. For forty years, biologists recorded these sounds without suspecting how much structure they contained. Only now, with machine learning, does this structure become visible, casting new light on an old question: does something like language hide behind the clicking of the largest toothed whale on Earth?²

Why we have been listening to whales for 40 years

That we listen to whales at all is largely thanks to one person: the biologist Roger Payne. In 1970 he released the album “Songs of the Humpback Whale,” which for the first time brought the elaborate songs of humpback whales to a wide audience. The record sold more than a hundred thousand copies and sparked the global “Save the Whales” movement, which eventually contributed to the moratorium on commercial whaling. Payne made audible that whales are not silent giants of the sea, but highly social animals with a sound world of their own.

Half a century later, the circle closed. Payne became chief advisor to Project CETI and, in the final year of his life, advocated for a responsible use of technology to listen to whales and perhaps one day understand them. He died in 2023, but is listed as a co-author on the study that made headlines shortly afterward.¹ The tape recording has since become a dataset, and the trained ear a trained model.

Who are sperm whales, and what is a coda?

Sperm whales (Physeter macrocephalus) are the loudest animals on the planet. They produce clicks for echolocation, which they use to hunt in the darkness of the deep, and for social communication. Crucial to communication are short, rhythmic sequences of clicks that experts call codas. A coda usually consists of anywhere from a handful to several dozen clicks and lasts only seconds.

Codas are deeply social. Whales exchange them within their groups, often overlapping, so that several animals click at the same time. Different groups, known as clans, maintain their own coda dialects, similar to regional accents. It is precisely this diversity that makes the analysis so appealing, and so difficult.

The most important dataset comes from the Dominica Sperm Whale Project led by biologist Shane Gero, who has followed and recorded individually known whales around the Caribbean island of Dominica for over a decade.² Only this carefully annotated data treasure, built up over years, makes the work of AI possible in the first place. Without clean, context-rich data, even the best model remains blind, a principle that holds true in marine biology just as much as in any company.

What did the AI discover in 2024?

In May 2024, a team led by MIT researcher Pratyusha Sharma, together with Shane Gero, Roger Payne, David Gruber, Daniela Rus, Antonio Torralba, and Jacob Andreas, published a widely noted study in Nature Communications.² For it, machine learning analyzed nearly 9,000 codas from the Eastern Caribbean clan.

The result surprised even the experts. The clicks are not a simple “Morse code” of on and off, as had long been assumed. Instead, the researchers found a combinatorial system: from a few building blocks, a great many different signals can be formed, much as humans combine a small number of sounds into countless words.³ The authors therefore spoke, for the first time, of a phonetic alphabet of sperm whales.

Particularly striking was the context dependency. Certain features of a coda did not remain fixed but shifted systematically with the course of the exchange, depending on what had been clicked before. This very flexibility within a conversation is one of the hallmarks known from human communication, and it only became visible across thousands of codas once a model made it visible across the entire dataset. No human could have picked out these subtle regularities by ear from the noise of the deep sea.

What does an “alphabet” made of clicks sound like?

The whales’ alphabet is not made of letters, but of four features that shape the codas. The researchers borrowed terms from music for this:²

• Rhythm describes the timing pattern of the clicks within a coda. It is categorical and context-independent, essentially the fixed base form.
• Tempo refers to the overall duration and speed of a coda. Tempo, too, is a fixed, recognizable property.
• Rubato is the fine, continuous stretching or compressing that a whale applies during an exchange, much like a musician slightly stretches a melody.
• Ornamentation refers to additional clicks inserted into a familiar coda depending on the conversational situation, comparable to an embellishment.

The interplay is decisive: rhythm and tempo form the stable framework, while rubato and ornamentation carry the context-dependent variation. This combinatorics produces a repertoire far larger than the sum of the individual coda types. The whales thus appear to have a system with which they can span a large space of possible signals.

What was added in 2025?

The story did not end there. In late 2025, a team led by Berkeley linguist Gašper Beguš, together with Project CETI, followed up with a study in the Proceedings of the Royal Society B on the phonology of sperm whale codas.⁴

The researchers found that whales apparently also deliberately vary the frequency of their clicks. Two recurring patterns appear across individuals, which the authors, drawing on human phonetics, call vowels, an “a-coda” and an “i-coda.” Even diphthong-like transitions, in which the sound spectrum shifts within a single coda, were described. This adds another layer to the picture: alongside the timing pattern of the clicks, their tonal coloring also carries information. Here, too, the sober caveat applies: this concerns structural similarity to human spoken language, not proven meaning.

How does the AI behind this actually work?

It starts not with the model, but with the measurement. Project CETI deploys bio-loggers, small sensors temporarily attached to the whales that record high-resolution audio together with contextual data such as diving depth, movement, and behavior. Underwater microphones complete the picture. These datasets are designed from the outset to be machine-analyzable.³

Behind the analysis stands an entire technical ecosystem. Project CETI brings together AI specialists, marine biologists, roboticists, and acousticians who use autonomous buoys, underwater robots, and drones to capture audio, video, and whale behavior around the clock. Only this volume and density of synchronized data makes it possible to statistically confirm rare patterns at all.

On this basis, machine learning models search thousands of codas for recurring patterns, group similar signals, and reveal how click patterns change with conversational context. The AI does not provide the finished answer, but the map: it shows researchers where a closer look is worthwhile. The actual interpretation only emerges from the interplay of computing power and biological expertise.³

It is often said that “the same technology behind ChatGPT” is used here. That is a catchy but imprecise simplification. Project CETI uses methods from the broad field of AI and machine language processing, which also includes large language models. The structural findings from 2024 and 2025, however, come mainly from statistical pattern and sequence analysis, not from a “ChatGPT for whales.” A more accurate description: these are AI methods from the same research family, applied to an entirely new terrain.

Does this mean we will soon be able to talk to whales?

Here, sobriety is warranted. As impressive as the structure is, the meaning of the codas remains unknown. The AI recognizes patterns, not semantics. As long as no one knows what a particular click pattern conveys to a whale, it cannot be said whether this counts as language in the human sense.²

There is also substantive criticism. Some marine biologists consider certain of the patterns found to be recording artifacts or byproducts of attention rather than linguistic signals. And much about whale behavior does not resemble human speech: sperm whales often click simultaneously and align their rhythms with one another, which looks more like a choir or a duet than an alternating conversation.

Ethical questions also arise. An analysis by Mark Ryan and Leonie Bossert names anthropomorphism, the premature transfer of human concepts, as one of several challenges in trying to use AI to “speak whale.”⁵ The authors see great value in decoding the sounds for species conservation, but warn against actively trying to communicate with whales, since this could harm the animals. There also remains the ironic limit of all language models: even a system that could fluently click along would not automatically understand the content, any more than we would.

Why this matters beyond whales

Why should you, as a reader, care what sperm whales call to one another? Because it shows, in miniature, what AI today is genuinely good at, and where its limits lie. It finds structures in vast, unwieldy amounts of data that remain hidden to the human eye and ear. But it does not replace the expertise needed to interpret those structures. It is precisely in this division of labor, AI as pattern finder, humans as meaning maker, that the value of the technology lies, whether in the ocean or in a company.

And the benefit for whales is real. Whoever understands how whales are doing, what stresses them, and how they respond to noise, ship traffic, or climate change, can protect their habitat more effectively. AI thus becomes a kind of digital stethoscope for a realm that, until now, was mostly just noise to the human ear. What kind of understanding will eventually emerge from this remains open. That we can listen to the deep sea better today than ever before is, however, certain.¹