Decoding whale language, and what it will mean for AI
Dear Reader,
There is a specific scientific project happening right now that I think will be remembered as one of the most important of our time, and that most people have not heard about. It is called Project CETI, the Cetacean Translation Initiative, and its goal is to decode the communications of sperm whales using machine learning. If it succeeds, we will have translated, for the first time in human history, the language of a non-human species.
I want to be careful about the word "language" here, because whether what whales do counts as language depends on definitions that are themselves contested. But whatever we call it, whales exchange highly structured acoustic signals that carry information, that vary across social groups, that are learned through cultural transmission, and that appear to be doing something functionally similar to what language does in humans. And in the last few years we have developed the tools to potentially crack the code.
If we succeed, the implications go far beyond marine biology. They will reshape our understanding of what communication requires, what consciousness looks like in radically different bodies, and how AI systems might one day mediate between different kinds of minds. This is one of the frontiers where I most want to be alive to see the results. Let me walk through what is going on.
Start with what we know about whale communication. Sperm whales are the specific focus of Project CETI, so let me focus on them.
Sperm whales produce sounds called codas. A coda is a short pattern of clicks, typically containing between three and forty clicks, separated by specific time intervals. Different codas have different click patterns, and different populations of whales use different repertoires of codas. Individual whales seem to have signature codas that identify them, similar to how songbirds have signature songs. Family units use codas that mark them as members of the same social group. Whole cultural regions of the ocean have their own coda dialects that other whale groups do not use.
This much is not disputed. What is disputed is what the codas mean.
The traditional view is that codas function primarily for social bonding and identification. They are like acoustic handshakes. They tell other whales who is present, who belongs to which group, and where they are in the water column. Under this view, the codas do carry information, but the information is limited to social positioning and does not have the complex referential structure of language.
The emerging view, driven partly by better recordings and partly by machine learning analysis, is that codas may carry more information than previously thought. Recent work at Project CETI and elsewhere has identified structural features in coda patterns that resemble what linguists call phonology, the way sounds combine to form larger units of meaning. There appear to be modifications to codas that function like inflections. There appear to be sequences of codas that may function like sentences. There appear to be behavioral correlations with specific coda types that suggest referential content.
Nothing about this is settled. But the direction of the evidence is toward whale communication being more complex than the traditional view suggested. Whether it rises to the level of language depends on what counts as language.
What Project CETI is trying to do is apply modern machine learning to the problem of understanding coda structure and meaning.
The approach is essentially this. Collect enormous amounts of high-quality recordings of sperm whale communications, tagged with information about the context in which they occurred. Feed the recordings into machine learning models designed to find structure in sequential data, similar to how large language models find structure in text. Look for patterns that correlate with specific behavioral contexts, specific social relationships, and specific environmental events. Build up, over time, a mapping between whale communications and the situations in which they occur.
This is a version of the approach that worked for translating between human languages before we had bilingual dictionaries. The Rosetta Stone approach was to find aligned texts in known and unknown languages and work out the mapping. Modern machine translation works by finding statistical patterns in massive datasets and inferring meaning from co-occurrence. Project CETI applies this to a species where we have no aligned texts and where the modality is not text at all but acoustic signals in a very different sensory environment.
The technical challenges are enormous. Sperm whales live in complex acoustic environments where multiple whales may be communicating at once. The recordings require underwater sensors deployed at scale across their habitats. The models need to handle a much sparser dataset than what is available for human languages, and the datasets need to be tagged with behavioral information that requires visual observation of whales who are often deep underwater. All of this is being worked on. None of it is fully solved.
But the ambition is real, the funding is significant, and the timelines being discussed are surprisingly short. Some researchers involved in the project have suggested that meaningful partial translations may be possible within the next decade. Others are more skeptical. What is clear is that this is the closest we have ever come to systematically decoding the communication of another species, and the tools we now have make the effort feasible in a way it never was before.
Let me tell you why this matters philosophically.
For as long as humans have wondered about the minds of other animals, we have wondered whether they have language. Language has been treated, since at least the ancient Greeks, as the specific feature that distinguishes humans from other animals. We have language. Animals have communication. And the difference between these two, on the traditional view, is what makes us the specific kind of being we are, capable of the specific kind of thought we have.
If whales turn out to have something that meets any reasonable definition of language, this framework collapses. The specific human superpower is no longer unique. It is one instance of a broader capacity that also shows up in cetaceans, and possibly in corvids, dolphins, and other cognitively sophisticated species. Whether human language is more complex than whale language, or structured differently, would be an empirical question rather than a definitional one. And the philosophical consequences would ripple through everything from ethics to epistemology to our sense of our own place in nature.
I want to be honest that this is one possible outcome. It is also possible that as we decode whale communication, we find that it is genuinely different from human language, not less complex but structured on different principles that do not map cleanly onto our categories. This would also be revolutionary, because it would show us a form of communication that had evolved independently and that we could learn to understand without pretending it was just a variant of what we already know. Either outcome would be philosophically significant. What is unlikely is that decoding whale communication would leave our current concepts intact.
There is a specific reason whales are particularly interesting for questions about AI.
Cetaceans have large brains. Sperm whales specifically have the largest brains of any species on Earth, weighing around 8 kilograms, several times heavier than a human brain. This is not just a function of body size. Even accounting for scale, cetacean brains are highly developed, with specific neural structures that in some ways exceed what mammals typically have. They have large frontal cortices. They have highly folded cortical surfaces indicating extensive processing capacity. They have specific structures like spindle neurons, which had been thought to be exclusive to great apes and were implicated in social cognition and self-awareness. These structures show up in whales and dolphins too.
What this means is that whales have the biological machinery for cognitive complexity comparable to or exceeding what we have. What they may lack is the specific evolutionary and cultural context that in humans led to the development of technology, agriculture, and civilization. They have been highly intelligent for tens of millions of years, presumably doing something with all that neural capacity, but not building anything we recognize as culture in the human sense.
Or perhaps they have been building something we have not recognized. If cetacean communication carries the kind of complex information that language carries, then cetacean culture may be enormous, with dialects, traditions, histories, and knowledge structures we have simply never had access to. Decoding whale communication would not just tell us about whales. It might tell us about a form of culture that has been running in parallel to ours for a very long time, without us being able to see it.
This is one of the more disorienting possibilities in the whole field. The idea that there could be complex non-human cultures on Earth that we simply cannot perceive because we do not share the modality is at once obvious in retrospect and shocking to fully take in.
Let me tell you what I find hardest to sit with.
If we succeed in decoding whale communication, the first thing we may learn is what whales have been saying to each other for as long as we have been recording them and not understanding. Some of that recording may have happened during whaling, when whale populations were being reduced by ninety percent or more. If whales have anything that functions like a shared cultural memory, they may know very well what happened. They may have said things about it to each other. Things that we, if we ever decode them, will have to hear.
I do not think this is likely in the specific way I am describing. Sperm whale populations mostly did not survive intact through the whaling era, and whatever cultural transmission was possible would have been disrupted. But something about the shape of the possibility haunts me. We may be about to discover that we have been living alongside minds that have been aware of us all along, in ways we did not know how to detect, and that have been forming their own understanding of what we are and what we have done.
Whether or not this is literally the case, the general structure applies. Cetaceans have watched us for a long time. They have complex acoustic communication that we do not understand. If we come to understand it, we will not just be gaining a tool. We will be entering a conversation that has been ongoing without us. What we find there is not entirely up to us.
What does this have to do with AI.
The connection is not superficial. Project CETI is using AI, specifically large language models and related techniques, to try to understand a non-human mind. The tools being developed to potentially decode whale communication are the same class of tools that could in principle be applied to any communication system, biological or artificial.
This has two implications I want to draw out.
The first is that we are learning how to build bridges between different kinds of minds. If AI can help us translate whale communication, it can in principle help us translate between any two systems that have communication and where we can collect enough data. This is a genuinely new capability. The ability to systematically decode the communication of arbitrary systems has never existed before. And the implications go far beyond marine biology.
The second is that we are learning something about AI itself. The techniques that work for decoding whale communication tell us something about what language and communication are, at a mathematical and structural level. This is being fed back into how we build AI systems. If we can find general principles that unify whale codas, human languages, and possibly AI-produced text, we will have discovered something deep about the nature of communication that is not tied to any specific substrate.
Both of these possibilities are speculative in the specifics. They are also, in the abstract, exactly what we would expect if we were on the verge of a general theory of communication that transcends the biological cases we have historically studied. Whether we get there in the next decade or the next century, we are moving in that direction. Project CETI is one of the specific projects that will tell us how fast we are moving.
I want to close with something I have been thinking about for a while.
The nineteenth and twentieth centuries were dominated by the assumption that language was a specifically human capacity. This assumption structured everything from linguistic theory to philosophy of mind to how we treated non-human beings. When we started seriously asking whether other species had communication systems complex enough to count as language, the answer for a long time was no, because we were looking for exactly the features that made human language distinctive. When we did not find them, we concluded language was unique to us.
What we are now doing, with sperm whales and with AI, is asking the question differently. Instead of asking whether other communication systems have human-like language, we are asking whether human language is a specific instance of a more general capacity that might be present in multiple forms across different substrates. This is a different question, and it may have different answers. It also fits a pattern that has been consistent across many domains. Whenever we have asked whether some capacity is specifically human, and we have looked carefully enough, we have usually found that it is more widely distributed than we thought. The specific human version may still be distinctive. But the underlying capacity almost always turns out to be more general.
I expect this pattern to hold for language, if we can figure out what to look for. And if it holds, then the question of whether AI can have language becomes an empirical question about specific systems, rather than a metaphysical question that can be settled by definition. Whale research is one of the paths to answering it. AI research is another. They are converging faster than most people realize.
Next month I want to write about Nagel's original question, what is it like to be a bat, and what has become of that question fifty years after he asked it. It turns out we have learned a great deal about bats in the intervening decades, and the specific way Nagel's question has aged tells us something important about the whole project of understanding non-human minds. Stay with me.
— Transmission Sent —
Niklas Hanitsch
Reference materials
- Project CETI — Cetacean Translation Initiative (https://www.projectceti.org/)
- David Gruber et al. — Whale Vocalizations and Machine Learning: A Framework for Decoding (2023)
- Hal Whitehead — Sperm Whales: Social Evolution in the Ocean (2003)
- Carl Safina — Beyond Words: What Animals Think and Feel (2015)
- Michael Bronstein et al. — Toward Understanding the Communication in Sperm Whales (2021)
- Shane Gero — Dominica Sperm Whale Project research publications
- https://www.pnas.org/doi/10.1073/pnas.2405266121
- https://en.wikipedia.org/wiki/Sperm_whale_coda
Continue reading
- The octopus knows something we do not
- Crows, ravens, and the intelligence we underestimated
- What is it like to be a bat? Nagel's question in 2026
- Can AI Lie? The Surprising Philosophy of Machine Deception
- Post-human futures: Three scenarios I can actually believe
Frequently asked questions
Do whales actually have language? Depends on your definition of language. Whales produce highly structured acoustic signals that carry information, that vary across social groups, and that are learned through cultural transmission. Whether these signals meet the technical criteria for language, which require features like recursion, displacement, and productivity, is disputed. What is not disputed is that whale communication is far more complex than the classical view acknowledged.
What is Project CETI? Project CETI, the Cetacean Translation Initiative, is a scientific effort to decode the communications of sperm whales using machine learning. It combines large-scale acoustic recording, behavioral observation, and modern AI techniques to look for meaningful structure in whale codas and to potentially translate them.
When might we understand what whales are saying? Timelines vary. Some researchers involved with Project CETI have suggested that meaningful partial translations may be possible within the next decade. Others are more cautious. The technical challenges are substantial, but the tools available for this kind of work are advancing rapidly.
What are sperm whale codas? Codas are short patterns of clicks produced by sperm whales, typically containing three to forty clicks with specific rhythmic patterns. Different codas are used by different social groups, individuals have signature codas, and there appear to be modifications that function like inflections. The full structure and meaning of codas is still being investigated.
How does whale research relate to AI? The same machine learning techniques that make Project CETI possible are also central to how modern AI understands and produces language. Success in decoding whale communication would tell us something general about how communication works, which would inform both animal linguistics and AI development. The two fields are increasingly connected.