For about forty years, a researcher named Shane Gero has spent his summers off the coast of Dominica, listening to sperm whales. He knows individual animals by the notches on their flukes. He knows which whales are related, which ones babysit for each other, and which families sound different from the rest. What he doesn’t know is what any of them are saying.
That gap is the reason a growing number of scientists have turned to artificial intelligence. The idea sounds like science fiction: feed enough whale sound into a machine-learning model and see whether it can find the grammar. Yet over the past two years the work has produced real, peer-reviewed results. It has also produced a good deal of overexcited headlines, so it’s worth separating what has been shown from what is still a hope.
Why whales, and why now?
Whales are an obvious target for this kind of research. Sperm whales have the largest brains of any animal on Earth, live in tight matrilineal family groups, and communicate almost constantly with patterned bursts of clicks called codas. Different clans use different coda repertoires, and young whales appear to learn theirs from their elders, which is what researchers mean when they say these animals have culture.
Until recently, though, nobody could do much with that complexity. A human listener hears something like Morse code. A single hour of recording might contain thousands of clicks, and a research career might produce millions. Sorting through that by hand is not realistic, and it’s the sort of pattern-finding problem that modern machine learning handles well.
That’s the bet behind Project CETI, the Cetacean Translation Initiative. It is a nonprofit that applies machine learning and robotics to listen to and translate sperm whale communication, with its research centered on Dominica in the Eastern Caribbean. The first phase is plain data collection: building a huge, carefully labeled collection of whale sounds paired with what the whales were doing when they made them. Without that context, a translation model would have nothing to learn from. projectceti
What the research has found so far
The most talked-about result is the so-called phonetic alphabet. Using machine learning on recordings collected around Dominica, CETI researchers found that sperm whale codas aren’t just a fixed menu of rhythms. They vary in tempo, in how the rhythm stretches and compresses, and in small additions to the pattern, and these features combine in ways that look systematic. David Gruber, who founded the project, described this in a Bioneers interview as one of the team’s biggest breakthroughs, along with the finding that the silences between clicks seem to carry vowel-like features that may form another layer of the whales’ communication system. bioneers
The newest chapter came this spring. A study reported on by Mother Jones found that sperm whale codas have multiple interacting layers of structure, and are among the closest parallels to human phonology of any animal communication system that has been analyzed. That’s a careful scientific statement, not a claim that whales speak a language like ours. Mauricio Cantor, a behavioral ecologist who wasn’t involved in the work, said it shows that whale communication goes well beyond patterns of clicks. Mother Jones
It’s also worth knowing that the science doesn’t always point in one direction. In an earlier analysis covered by Science News, the team used a type of AI called a generative adversarial network and noticed that some codas sounded more like a clack than a click. Gero told the reporter the pattern might reflect emotional state, but not everyone was convinced. One outside researcher doubted that the pattern was something the whales produce or notice on purpose, and CETI’s own linguist pointed out that ripples caused by the recording itself can look much like the pattern they found. That’s how real science works: a finding gets proposed, then challenged, then tested again. sciencenews
A birth, recorded in sound
One of the more moving results has nothing to do with decoding words. In March 2026, CETI announced two papers describing the most comprehensive documentation of a sperm whale birth ever recorded, along with the first quantitative evidence of cooperative birth assistance among non-primates. The audio from that event showed the whales’ vocal style shifting at key moments, including vowel-like structures, according to the announcement on EurekAlert. eurekalert
That matters because it ties sounds to events. A model that only hears codas can say they’re structured. A model that hears codas alongside a birth, a dive, or a nursing session has a chance of linking sound to meaning. It’s slow, painstaking work, and it’s exactly what the translation effort needs.
The hardware problem nobody talks about
Machine learning gets most of the attention, but a large share of the effort is just getting the recordings. Sperm whales dive for long stretches and spend only a short time at the surface, so researchers can wait for hours with nothing to show for it.
CETI’s engineers have tackled this from several angles. One team built a framework called AVATARS that predicts when and where a whale will surface, a method The Robot Report described as using sensor data and models of dive behavior to make better use of the small window at the surface. Harvard engineers, working with CETI, built an open-source bio-logger. It attaches to sperm whales and records high-fidelity, multi-channel audio plus behavioral and environmental data, all tailored for machine-learning analysis. harvard
The latest tool is an underwater glider. In April 2026, CETI said it was working with the French firm Alseamar on gliders that follow whales by listening for them. The glider uses four hydrophones, and software on board works out where a sound is coming from and steers the vehicle toward it. Roee Diamant, CETI’s underwater acoustics lead, told The Robot Report that the team can currently hear whales from about 12 kilometers away, depending on the type of vocalization. Compared with boats and buoys, which pick up a lot of unwanted noise, a glider can quietly stay with a group for far longer. Marine Technology News noted a further benefit: the same system could show how whales respond to shipping, offshore construction and fishing noise. Project CETI Tracks Sperm Whale Conversations in Real Time +2
Dolphins join the conversation
Whales aren’t the only animals getting this treatment. In April 2025, Google announced DolphinGemma, working with Georgia Tech and the Wild Dolphin Project, a nonprofit that has been studying Atlantic spotted dolphins in the Bahamas for about four decades. Scientific American reported that the team used the Wild Dolphin Project’s acoustic archive to train what they described as the first large language model for dolphin vocalizations. Scientific American
The approach borrows from how text-prediction systems work. AIwire explained that DolphinGemma is an audio-in, audio-out model that converts dolphin vocalizations into a structured format. Rather than predicting the next word in a sentence, it tries to predict the next sound in a dolphin’s sequence. It’s also small enough to run on a Pixel 9 phone in the field. hpcwire
The goal goes beyond listening. DolphinGemma is meant to work alongside a system called CHAT, which plays specific synthetic whistles to dolphins. As Entrepreneur summarized it, if a dolphin mimics an AI-generated sound, researchers respond by handing over a treat. The hope is that naturally curious dolphins will learn to request objects like seagrass or scarves with particular whistles. Denise Herzing, who founded the Wild Dolphin Project, has said that picking out the same patterns by hand would take people roughly 150 years. Entrepreneur
That’s a modest, concrete experiment, and it shows what “communicating with animals” is likely to look like first: a small shared vocabulary for a handful of objects, not a conversation.
What the skeptics say
It would be easy to write this story as a straight march toward a universal translator. The researchers themselves don’t. Arik Kershenbaum, a zoologist at Cambridge, compared the DolphinGemma effort in Scientific American to the Star Trek film where the crew can imitate whale song but has no idea what it means. Generating realistic animal sounds is not the same as understanding them.
Several honest limits stand out.
Sample size. The Dominica sperm whales are one population. Dolphins from the Bahamas are one population of one species. Whether the findings travel to other groups is an open question.
Meaning is hard to prove. A pattern in the data isn’t automatically a unit of meaning. To show that a certain coda means “dive” or “come here,” researchers need repeated, controlled evidence of how whales react, which is difficult to get from animals that live in the open ocean.
Human interpretation. We tend to look for human-like categories such as letters, vowels and words, and that can lead us to see structure that is really our own projection. CETI has said its approach is to listen in context, but the risk remains.
Recording artifacts. As the clack debate showed, instruments can create patterns that aren’t in the animal’s voice.
The realistic near-term goal is also much smaller than “translating whales.” According to Mother Jones, Project CETI aims to understand about 20 distinct vocalized expressions, tied to actions such as diving and sleeping, within the next five years. Twenty expressions is not a dictionary. It would, however, be a genuine first. Mother Jones
The question of whether we should
There is also a harder question underneath the technical ones: what happens if this works? Bioneers published a companion piece by legal scholar César Rodríguez-Garavito asking what responsibilities come with the ability to decode another species’ communication. The questions are practical. Would hearing whales objecting to shipping noise change how we regulate it? Should recordings of distress calls be shared freely? Could playing synthetic calls to wild animals disturb them or interfere with their social lives?
Gruber has been open about the conservation angle, saying that understanding another form of intelligence changes how we think about communication and life beyond our own species. There’s a long record of science changing law and public attitudes this way. The discovery that humpback whales sing, popularized in the 1970s, helped turn whaling from an ordinary industry into a global controversy. It isn’t hard to imagine that evidence of whale families grieving, cooperating and teaching their young could shift opinion again.
That’s also why the field needs strict standards. Playback experiments that change animal behavior should be reviewed like any other research involving wildlife, and claims should be held to the same evidence bar as any other science. A mistranslation that gets picked up by headlines could do real harm to public trust, to funding, and potentially to the animals.
Where this goes next
Three trends are worth watching.
First, the data is growing quickly. CETI’s drones, tags and gliders are producing the kind of large, context-rich dataset that earlier researchers could only wish for. The project plans to widen its datasets beyond Dominica, and the open-source bio-logger lets other labs gather comparable recordings.
Second, the tools are spreading. Google said it intended to release DolphinGemma as an open model so researchers could adapt it to other species. Newsweek quoted a Google spokesperson describing the potential to speed up research within and possibly between species. That cross-species angle could matter a great deal, because patterns that show up in dolphins, whales and other animals are more likely to reflect something real about how vocal communication works.
Third, expectations are slowly becoming more realistic. The most credible researchers talk about listening before speaking, about context before translation, and about small steps. That’s a healthier attitude than the old fantasy of asking a whale what it thinks.
The bottom line
Can scientists communicate with whales using AI today? Not in any way that deserves the word “conversation.” What they can do is record whales more completely than ever, find structure in their sounds that human ears missed, and begin to link some of those sounds to behavior. Sperm whale codas appear to be far more intricate than anyone assumed. Dolphin researchers are testing whether a small shared vocabulary is possible.
That’s less dramatic than a headline about talking to whales, and it may be more important. Every new layer of complexity found in these animals’ voices is a reminder that we share the planet with other minds, and that we’ve been mostly deaf to them. AI isn’t a magic translator, but it may turn out to be a very good hearing aid. If the next few years go well, the first thing we learn may not be a clever phrase, but simply how much the whales have been saying all along.
