Sperm whales produce structured patterns of clicks that carry information, context, and identity across miles of open ocean. Project CETI is using machine learning to decode them. What it is finding challenges everything we thought we knew about the boundaries of language.
In the deep ocean, a sperm whale speaks. The sound — a rapid sequence of clicks called a coda — travels thousands of metres through dark water to reach other whales. The receiving whale responds. The exchange continues. It has been going on, without human comprehension, for millions of years. We are, at last, beginning to listen.
Project CETI — the Cetacean Translation Initiative — is the most ambitious attempt in history to decode non-human communication. Using hydrophones, autonomous underwater drones, machine learning, and the same analytical frameworks developed for human language, researchers are asking a question that sounds more philosophical than scientific: do sperm whales have language? The 2024 data suggests that the answer may be yes — and that the implications reach far beyond marine biology.
Sperm whales are, in almost every physical dimension, extreme. The largest toothed predator on Earth, they can reach 18 metres in length and dive to depths exceeding 2,000 metres in pursuit of giant squid. Their brains are the largest of any animal ever to have lived — averaging 8 kilograms. And their acoustic system is the most powerful biological sound-production mechanism known to science.
The spermaceti organ — a massive structure occupying up to a third of the whale's enormous head — is filled with a waxy oil that focuses and amplifies sound. Sperm whale clicks can reach 230 decibels, the loudest sound produced by any animal, with the energy to stun or kill prey at close range. But these clicks serve another purpose that is only now being fully understood: communication.
Sperm whales produce two distinct categories of clicks. Usual clicks are used for echolocation — finding prey in total darkness. Codas are something different: slower, rhythmically structured patterns of 3–40 clicks that are exchanged exclusively between whales during social interactions. Codas are not random. They have structure, consistency, and — most strikingly — apparent social function.
Codas are not uniform across the species. Sperm whale populations around the world produce distinct coda dialects — shared repertoires that differ between what researchers call "vocal clans." The Caribbean clans produce different coda patterns than Pacific clans; clans in the Indian Ocean have their own distinctive repertoires. These dialects are learned and maintained socially, passed down through generations of females in matriarchal groups. The parallel with human language — structurally, socially, and geographically — is difficult to ignore.
Project CETI — the Cetacean Translation Initiative — was launched in 2021 by marine biologist David Gruber of City University New York and machine learning researcher Michael Bronstein of the University of Oxford (formerly of DeepMind). The project brings together bioacousticians, linguists, AI researchers, roboticists, and cryptographers with a single stated goal: to collect and decode the largest dataset of sperm whale communication ever assembled, and use it to determine whether whale codas constitute a language.
The method is modelled explicitly on the techniques used to build large language models like GPT. Those models learned language by training on billions of words of human text — finding statistical patterns, contextual relationships, and semantic structures without being told what language was. Project CETI is applying the same approach to cetacean communication: collect an enormous corpus of coda sequences, record the full social context in which each exchange occurs, and train machine learning models to find the structure within.
A landmark study published in Nature Communications in 2024 by Gero et al. and the Project CETI collaboration reported that sperm whale codas possess statistical features previously associated only with human language: combinatorial structure (a finite set of elements combined in varied sequences to create meaning), context-dependence (the same coda meaning different things in different conversational positions), and individual variation (identifiable "vocal signatures" within the shared dialect). The paper stopped short of calling it language — but described it as "the most complex communication system documented outside of humans."
The data collection infrastructure Project CETI has deployed in the waters off Dominica in the Eastern Caribbean is extraordinary. A network of underwater hydrophones records ambient sound continuously. Autonomous underwater vehicles follow whale groups at depth, collecting close-range acoustic data. Custom-designed acoustic tags — the size of a deck of cards — are gently attached by suction cup to individual whales, recording both what the whale says and what it hears. All data streams to machine learning pipelines running continuous pattern analysis.
By 2024, the project had collected more than 4 million individual sperm whale clicks across thousands of social exchanges. The models trained on this data are beginning to produce what the team calls "semantic maps" — representations of coda relationships that suggest some codas are used in response to specific social situations, some appear to function as identity markers, and some — perhaps most intriguingly — appear to be used differently by individuals of different social status within a clan.
Understanding what sperm whales are saying requires understanding what sperm whale society looks like — because language, in every species studied, is embedded in social context. Sperm whales live in a highly structured matriarchal society. Females and calves live in stable social units of 6–12 individuals, often comprising multiple generations of related females. Males leave these groups at maturity, living more solitary lives before returning to mate. The social unit is the fundamental communicative context: codas are exchanged almost exclusively within and between these groups.
Researcher Shane Gero, who has been studying a single community of sperm whales in Dominica since 2005, has documented something remarkable: individual whales have distinct coda styles recognisable across decades. Calves learn the dialect of their unit; they also develop individual variations within it. The cultural transmission of communication is not just at the group level — it appears to be at the level of individual relationships.
The Project CETI data has also revealed something unexpected about coda rhythm. The inter-click intervals within a coda — the timing between individual clicks — carry information independently of the pattern of clicks itself. Two codas with the same click count can have different temporal structures that appear to be used differently in social context. This is analogous to prosody in human speech: the meaning of a sentence is carried not only by the words but by the rhythm, stress, and timing with which they are spoken. Machine learning models were the first analytical tools sensitive enough to detect this layered structure in whale communication.
Cetacean and primate brains diverged roughly 94 million years ago, long before the evolution of complex cognition in either lineage. The large, structured brains of sperm whales — and the social, vocal communication that appears to rest on them — evolved entirely independently of the human lineage. If sperm whales have developed something approaching language, they did so via a completely different evolutionary route, from completely different neural architecture. This is the most profound implication of the CETI data: complex symbolic communication may not be a singular human accident. It may be what sufficiently complex social intelligence converges on, regardless of substrate.
Project CETI's ultimate goal — translation — remains distant, and the researchers are careful about the word. Translating whale communication would require not just identifying patterns but grounding those patterns in meaning: knowing what situation a given coda is used in, what response it elicits, and whether that relationship is consistent and generalisable. The team estimates this will require a corpus ten to a hundred times larger than what has been collected so far, and models significantly more sophisticated than current tools.
But the intermediate findings are already reframing the conversation about non-human intelligence. Parallel AI-driven projects are finding structured complexity in the vocalisations of humpback whales (whose songs show signs of cultural evolution resembling the spread of musical trends through human populations), dolphins (who appear to use individually distinct "signature whistles" functioning as names), and crows (whose problem-solving behaviours suggest causal reasoning far beyond what was previously attributed to birds).
What machine learning is revealing, across biology, is that we systematically underestimated the complexity of non-human minds — because we lacked tools sensitive enough to detect that complexity in the data. The same analytical power that allows AI to find patterns in the Herculaneum scrolls, or in ancient genomes, or in climate systems, is finding patterns in the communications of animals we shared this planet with for millions of years without understanding.
The ocean is not silent. It never was. We simply weren't listening carefully enough.
Buy me a coffee