AI Just Translated What Dolphin Are REALLY Saying — Marine Biologists Are Speechless

Ajouté : 2026-05-11

[00:00:00]On April 14th, 2025, on National Dolphin Day, Once a dolphin starts doing a vocalization like a whistle, it [music] can try to complete the end of it.

[00:00:13]Google CEO Sundar Pichai posts a single announcement on X that got everyone buzzing.

[00:00:19]No product launch. No new phone. No quarterly earnings.

[00:00:24]He announces that Google has built an AI that is learning to speak dolphin. And that it's already generating sounds so accurate that the lead researcher, a man who has spent his career trying to produce dolphin-like audio with conventional software and failed, danced around the room the first time he heard what the machine created.

[00:00:44]The model is called Dolphin Gemma.

[00:00:46]And what it's revealing about what dolphins have been saying to each other, right beneath the surface of the ocean, for millions of years, is forcing scientists to reconsider a question they thought they'd settled a long time ago.

[00:00:59]What if humans aren't the only species on this planet that has language?

[00:01:03]Now, before we get to what the AI found, you need to understand something about dolphins, because most people think they know these animals.

[00:01:11]Cute. Playful. Smart enough to jump through hoops at a marine park. But that image, that theme park version of a dolphin, is a lie.

[00:01:21]It's a children's cartoon version of an animal whose actual cognitive abilities are so advanced that some researchers have spent decades arguing they should be classified differently from every other non-human species on Earth.

[00:01:34]Dolphins recognize themselves in mirrors. That sounds simple. It's not.

[00:01:39]The mirror self-recognition test, developed in 1970 by psychologist Gallup, [music] is one of the gold standards for measuring self-awareness.

[00:01:48]You place a mark on an animal's body in a spot it can't normally see, then give it a mirror.

[00:01:53]If the animal uses the mirror to inspect the mark, it understands that the reflection is itself, not another animal.

[00:02:00]Itself.

[00:02:02]Only a handful of species on Earth have ever passed this test. Great apes, elephants, magpies, and dolphins. In fact, young bottlenose dolphins demonstrate mirror self-recognition as early as 7 months old.

[00:02:17]Human children [music] don't typically pass the same test until 18 to 24 months.

[00:02:22]But, it goes further than self-awareness.

[00:02:24]Dolphins have been shown to understand syntax.

[00:02:28]In laboratory experiments, they can distinguish between commands like, "Bring the ball to the hoop." and "Bring the hoop to the ball."

[00:02:36]They understand that word order changes meaning.

[00:02:39]They grasp symbolic language.

[00:02:41]>> [music] >> They show signs of metacognition, the awareness of their own uncertainty, >> [music] >> indicating in experimental settings when they don't know the answer to a problem.

[00:02:50]And their brains are enormous.

[00:02:52]Bottlenose dolphins possess brains nearly twice the size predicted for their body weight, with deeply folded neocortices and a highly developed limbic system, regions associated with sensory processing, social awareness, decision-making, and emotion.

[00:03:09]So, an animal that knows it exists, that understands grammar, that has a brain built for complex social reasoning.

[00:03:18]The obvious question is, "What is it saying?"

[00:03:22]For decades, scientists [music] have known that dolphins produce three basic types of sound: clicks, whistles, and burst pulses.

[00:03:31]Clicks are used for echolocation, mapping the underwater world through sound.

[00:03:36]Burst pulses, rapid-fire streams of clicks, are associated with social situations like fighting, playing, and courtship.

[00:03:45]And then there are the whistles.

[00:03:47]And the whistles are where this story breaks open.

[00:03:50]In the 1960s, researchers first proposed that each bottlenose dolphin develops a unique sound, a signature whistle, that functions essentially like a name.

[00:04:01]Not inherited genetically, not copied from a parent, invented.

[00:04:06]Each calf, within the first few months of life, listens to the whistles around it, and then creates a completely original sound pattern that will identify it for the rest of its life.

[00:04:17]And it doesn't just announce its own name.

[00:04:20]It calls other dolphins by theirs.

[00:04:22]Research on more than 250 wild bottlenose dolphins in Sarasota Bay, Florida, confirmed that when dolphins are separated, they copy the signature whistle of the individual they want to find, essentially shouting a specific dolphin's name across the ocean.

[00:04:37]When that dolphin hears its name, it responds. [music] Not to unfamiliar whistles, not to random sounds, only to its own name called by someone it knows.

[00:04:48]And recent research published in Science Advances when even further.

[00:04:52]Scientists demonstrated [music] that when a dolphin hears another dolphin's signature whistle, it doesn't just recognize the sound. It forms a mental image of that individual.

[00:05:02]The whistle is representational.

[00:05:05]It triggers a memory of who that dolphin is, what it looks like, what their relationship is.

[00:05:11]Just like hearing a friend's name makes you picture their face.

[00:05:14]And dolphins can remember these names for more than 20 years without contact.

[00:05:18]Two decades of silence, and a dolphin will still turn toward the speaker when it hears the whistle of someone it once knew.

[00:05:25]But here's the problem that has haunted marine biology for half a century.

[00:05:30]Signature whistles are just one piece of dolphin communication. Beyond the names, there are thousands of other vocalizations. Clicks arranged in patterns, burst pulses that vary in rhythm and intensity, whistles that aren't signatures but seem to carry meaning. And for decades, scientists simply [music] could not process this data fast enough.

[00:05:50]A few minutes of dolphin recording could take hours to catalog manually.

[00:05:55]The patterns were too dense, too fast, too alien for human ears to decode. And that's where the AI came in.

[00:06:02]In April 2025, Google, in partnership with the Georgia Institute of Technology and the Wild Dolphin Project, announced Dolphin Gemma, the first large language model ever built for a non-human species. Trained on tens of thousands of hours of acoustic recordings of wild Atlantic spotted dolphins, collected over nearly 40 years by Dr. Denise Herzing and her team in the Bahamas, Dolphin Gemma doesn't translate dolphin sounds into English.

[00:06:29]It does something far more interesting.

[00:06:32]It learns the structure of dolphin communication, the way a language model learns the structure of human language.

[00:06:39]It identifies patterns.

[00:06:41]It predicts what comes next in a sequence.

[00:06:44]It detects relationships between sounds and behaviors that human researchers could never isolate on their own.

[00:06:51]And the results are staggering.

[00:06:53]As of early 2025, Dolphin Gemma was distinguishing over a dozen click variants that were previously indistinguishable to the human ear.

[00:07:02]It was predicting the next vocalization in a sequence with over 74% accuracy.

[00:07:08]And it was generating synthetic dolphin sounds so convincing that Dr. Thad Starner, Google DeepMind research scientist and Georgia Tech professor, said he couldn't produce burst pulses with regular software for years. But the model created them on its own from nothing but what it learned from the data.

[00:07:26]Denise Herzing, the woman who has spent over 40 years swimming with the same dolphin community in the Bahamas, who was named one of Time's 100 most influential people in AI in 2025, put the scale of this breakthrough in perspective.

[00:07:41]She said it would take human beings 150 years to manually comb through the data and pull out [music] the same patterns that Dolphin Gemma can find today.

[00:07:49]150 years. The AI does it in days.

[00:07:54]And the dream, the long-term goal, is bidirectional communication, not translation, conversation.

[00:08:02]The team has built a wearable underwater device called CHAT, Cetacean Hearing Augmentation Telemetry. A diver wearing the CHAT system can generate synthetic dolphin-like sounds, made up words, designed by AI to refer to specific objects that dolphins enjoy, like seagrass or sargassum.

[00:08:22]Two divers swim alongside a dolphin, use the synthetic sound to ask for an object, and pass it back and forth.

[00:08:31]If the dolphin mimics the sound and a researcher rewards it with the object, a shared vocabulary begins to form.

[00:08:39]Not human language imposed on dolphins, a new language built between species [music] from scratch.

[00:08:46]And dolphins aren't the only species under the AI microscope.

[00:08:50]In May 2024, researchers from MIT and Project CETI published a study in Nature Communications announcing they had identified what they called a sperm whale phonetic alphabet.

[00:09:04]Using machine learning to analyze over 9,000 recordings of Caribbean sperm whales, they discovered that the click patterns whales use to communicate, called codas, vary depending on conversational context. Some clicks act like suffixes attached to an unchanging rhythmic beat.

[00:09:22]The tempo shifts, the rhythm ornaments, the structure changes based on who the whale is talking to and what's happening around them.

[00:09:31]Daniela Rus, director of MIT's CSAIL, said the findings challenge the prevailing belief among many linguists that complex communication is unique to humans.

[00:09:41]And meanwhile, the Earth Species Project, a non-profit backed by LinkedIn co-founder Reid Hoffman and the Paul Allen Foundation, has built Nature LM audio, the first large-scale audio language model designed for all animal sounds.

[00:09:56]Not one species, all of them.

[00:09:59]It can identify species, distinguish sex and life stage from vocalizations, and analyze massive bioacoustic data sets in minutes rather than months.

[00:10:09]They're studying carrion crows, jumping spiders, [music] zebra finches, Hawaiian crows, and beluga whales.

[00:10:16]And a $500,000 prize is now being offered for the first AI-powered breakthrough in interspecies communication. The race isn't theoretical anymore. It's happening right now.

[00:10:29]In the water, on servers, on smartphones strapped to the chests of divers swimming alongside wild dolphins in the Bahamas.

[00:10:37]But here's where this story takes a turn that most people aren't expecting.

[00:10:41]Because the closer scientists get to understanding what dolphins are actually communicating, the more uncomfortable the implications become.

[00:10:49]If dolphins have names, if they call each other across the ocean and wait for a response, if they remember individuals they haven't seen in 20 years, if they understand grammar, if they show metacognition, if they have dialects that vary from pod to pod as differently as a New Yorker and a Texan, if AI confirms that their vocalizations contain structured, context-sensitive, [music] combinatorial patterns that look like language, then what are they?

[00:11:18]What does it mean that we've been keeping them in concrete tanks for entertainment? What does it mean that naval sonar exercises flood their world with noise so intense it can cause internal bleeding and permanent deafness?

[00:11:32]What does it mean that shipping lanes cut through their nurseries and feeding grounds, drowning out the calls they use to find their children?

[00:11:39]Thad Starner, the Google DeepMind scientist behind Dolphin Gemma, said something that stays with you long after you hear it.

[00:11:47]He said, "If dolphins have language, then they probably also have culture.

[00:11:52]And if they have culture, then we need to understand what priorities they have.

[00:11:56]What do they talk about? We're about to find out. And the answer, whatever it turns out to be, is going to change the way we think about every other species on this planet.

[00:12:07]For 40 years, Denise Herzing swam with the same dolphins, recorded their sounds, cataloged their lives, and waited for a tool powerful enough to crack the code. She described Dolphin Gemma as being like a new kind of microphone, one that hears patterns we could never recognize on our own.

[00:12:26]The microphone is on now. And for the first time in the history of life on Earth, we might actually be able to hear what another species has been trying to say.

[00:12:35]The question is whether we're ready to listen, and whether we'll like what we hear.