Wispr Flow turns your voice into clean, ready-to-send writing — speak naturally, it strips the filler and fixes the punctuation. I've used it daily since February 2026 to build this newsletter. Read the full review →
In April 2025, Google unveiled DolphinGemma, an AI model built with Georgia Tech and the Wild Dolphin Project to study the sounds wild dolphins make. Newsweek ran “Google Launches AI That Talks to Dolphins.” Other coverage suggested a conversation with another species was months away.
Here is where it actually stands. Google DeepMind’s own model page says DolphinGemma “is currently in development” and “will be openly available” — future tense, both times. No download. No peer-reviewed result. And the model was never built to translate anything.
That sounds like a debunking. It isn’t. The honest story is stranger and more useful, and it turns on a distinction that reaches far past dolphins: predicting what comes next is not the same as understanding what it means.
What the model actually does
DolphinGemma is deliberately tiny — around 400 million parameters — because it has to run on a Pixel phone in the water beside the researchers. It is audio-in, audio-out: it takes dolphin sound, compresses it into tokens, and predicts what is likely to follow. Google’s own wording is that it “processes sequences of natural dolphin sounds to identify patterns, structure and ultimately predict the likely subsequent sounds in a sequence” — much like a language model predicting the next word.
That is a real scientific instrument. The hardest problem in studying an unknown communication system is that you cannot tell where one unit stops and the next begins; dolphin sound is a continuous stream of whistles, burst-pulse squawks and click trains. A model that gets good at predicting the next sound has necessarily learned where the boundaries fall and which sequences recur.
But a model can reach near-perfect prediction on a signal while having no idea what any part of it refers to. Prediction reveals structure. It does not reveal meaning.
No follow-up questions required
Every sales leader knows the feeling. You walk into a pipeline review with a number you believe in, and twenty minutes later, you're defending every line item to a CEO who just wants to know what's actually going to close.
HubSpot Sales Hub ends that conversation. Every deal, every rep's activity, and every buyer signal are all in one place and updated automatically. So your forecast is built on what's actually happening. And when you present that number, you can stand behind it.
The astonishing part predates the AI
Dolphins call each other by name — and that finding has nothing to do with this model.
Each bottlenose dolphin develops a unique “signature whistle” early in life. In 2013, Stephanie King and Vincent Janik at the University of St Andrews showed in PNAS that dolphins respond to computer-generated copies of their own signature whistle, and not to other dolphins’ whistles played back the same way. Learned, individually distinctive vocal labels, used to address one another, in a mammal that is not us. No AI involved in establishing any of it.
The limits are real too. Arik Kershenbaum, a Cambridge zoologist who studies animal communication, told Scientific American: “It’s not immediately clear that dolphins have words.” Having names is not the same as having a vocabulary, and a vocabulary is not the same as a language.
CHAT is a different machine
Most of the confusion comes from merging two separate systems.
DolphinGemma listens to natural dolphin sound and looks for structure. CHAT — Cetacean Hearing Augmentation Telemetry, from Thad Starner’s team at Georgia Tech — does close to the opposite. It does not decode anything. It invents a handful of synthetic whistles, deliberately unlike anything dolphins already say, each tied to an object the animals enjoy: a scarf, a piece of sargassum seaweed. If a dolphin mimics the whistle, a researcher hands the object over.
The Wild Dolphin Project is admirably unsentimental about how far that has got. Of the moment a dolphin mimicked the sargassum whistle, their own write-up says “this is not to say that a dolphin knew what it was saying,” with “nothing to indicate that this ‘word’ was used in context.”
Where it stands now
Google said it would share DolphinGemma as an open model “this summer” — meaning 2025. It did not. There is no Kaggle page for it, no Google-published weights on Hugging Face, and no entry in Google’s own list of Gemma variants. Several secondary sites state confidently that it is downloadable from all three. We checked each one; it is not.
The field has not stood still. In June 2026 an independent team published Dolph2Vec, trained on more than five years of recordings, which they report outperforms general-purpose audio models at classifying and detecting signature whistles. What they claim is structure and interpretable acoustic units — not meaning, and not translation.
So: an unreleased model, a forty-year field study now in its 42nd season, one carefully-caveated mimic, and steady real progress on finding structure in a signal nobody can yet read. That is a good story. It is simply not the one about talking to dolphins.
The site version has a one-minute demo that makes this concrete — ten rounds of predicting the next symbol in a sequence you have never seen, followed by one question you cannot answer — plus a copy-paste prompt for running the same test on a pattern in your own numbers.
The lesson transfers. Whenever you are told an AI system “understands” something — your customers, your market, a disease, another species — the useful question is whether it has demonstrated understanding or demonstrated prediction. Those are very different claims, they get reported in identical language, and only one of them tells you why.