Will We Talk to Animals? AI Is Finding Structure in Their Sounds — Meaning Is Another Story

THE THESIS. AI has not handed us an animal translator. It has handed us something prior and more important: a microscope for sound. For the first time we can see structure where we used to hear only noise, and that has revived a field stalled for decades. The results of the past two years are real: a 'phonetic alphabet' in sperm whales, dolphin whistles that work like words, a $10 million prize for the first genuine two-way communication. But let's be plain: no one has translated a single animal 'sentence'. Detecting structure, predicting responses and translating meaning are three different rungs; the hype welds them into one. Our bet: within the next decade we will have partial functional dictionaries — which signal appears in which context — for a handful of intensively studied species. That is not Dolittle-style conversation. And it will still transform conservation, law, and our sense of who else 'speaks' on this planet.

TRANSLATING WITHOUT A ROSETTA STONE. Why is this dream back now? Because of one specific technique. In 2017, research teams showed you can translate between two human languages without a dictionary: each word becomes a point on a mathematical map (an embedding — a list of numbers capturing which other words it appears with), and it turns out the maps of two languages have similar shapes; you just align them. That idea — translation without a Rosetta stone — is what Project CETI and the Earth Species Project want to apply to animal signals, and there is even a published mathematical framework studying when it could work (Goldwasser and colleagues). Self-supervised models — the kind that learn from unlabeled data, like LLMs — are a perfect fit: with animals we have oceans of audio and almost no labels. The results are already here. NatureLM-audio, from the Earth Species Project, couples an audio encoder with a Llama 3.1 language model and can classify species and call types, and even count individuals in a recording (presented at ICLR 2025). Google trained DolphinGemma — about 400 million parameters, small enough to run on a Pixel phone, per Google — on 40 years of Wild Dolphin Project audio and announced its release as an open model in 2025: it works like a dolphin LLM, hearing a sequence of sounds and predicting the next one. The money is following: the Earth Species Project raised $17 million in 2024 ($10M from Reid Hoffman, $7M from the Waverley Street Foundation).

WHAT HAS ACTUALLY BEEN PROVEN. Let's separate the verified from the headline. One: structure. CETI analyzed more than 8,700 codas — the click sequences sperm whales communicate with — and found they combine rhythm, tempo, 'rubato' and 'ornamentation' into at least 156 distinct combinations, a combinatorial system published in Nature Communications (May 2024). In November 2025 (in the journal Open Mind), linguist Gašper Beguš and CETI added patterns reminiscent of vowels and diphthongs. Two: prediction. CETI's famous '~80% accuracy' is not understanding whales: it is a drone system (Science Robotics, October 2024) that predicts where and when a sperm whale will surface — 69% with acoustics alone, over 80% when tag data is added. Impressive logistics; zero semantics. Three: the first signal-to-meaning associations. The Coller Dolittle Prize ($100,000 a year, with $10 million for genuine two-way communication) went in 2025 to the Woods Hole team for showing that dolphins' 'non-signature' whistles work like words with shared meaning — one for alarm, another for unexpected situations — and in 2026 to Julie Elie (Berkeley) for sorting the zebra finch repertoire into categories of 'core words' that the birds themselves can tell apart. That is genuine rung-climbing.

THE LADDER THE HYPE SKIPS. Keep the map handy. Rung one: detect structure — non-random patterns. Achieved in sperm whales, dolphins and birds. Rung two: link signal to context and predict the response ('this call appears during alarm; if I play it, the animal reacts this way'). That is where the best teams are today: the Coller Dolittle winners and the Wild Dolphin Project's CHAT system, which tests synthetic whistles tied to specific objects. Rung three: translate meaning and answer back. No one is there. And the jump is enormous for a deep reason: we do not share a world. A sperm whale lives in a universe of sonar, pressure and darkness; even if we aligned its 'map' of signals with ours, there may not be enough common points. A model that predicts the next click does not understand clicks, just as predicting the next word is not, by itself, understanding the world. LLMs already taught us that lesson. Let's apply it.

LANGUAGE OR RHYTHM? WHAT THE SKEPTICS SAY. The scientific community is not buying the headline, and rightly so. Luke Rendell (University of St Andrews) argues that some of the new AI-detected 'sounds' inside codas may be recording artifacts: echoes of the click itself. Stephanie King (University of Bristol) doubts the patterns are intentional; they could reflect physiological arousal, not a message. And Denise Herzing (Florida Atlantic University) warns against words like 'vowel', recalling how inflated dolphin claims in the 1960s discredited the field for decades. The underlying debate is serious: a combinatorial system does not imply syntax, and syntax does not imply semantics. That sperm whales combine rhythms does not prove they combine meanings. Our position: the skeptics are not this field's brake — they are its quality control. To CETI's credit, it publishes data and methods openly and usually distinguishes 'complex communication system' from 'language'. The headlines, not always.

THE DILEMMA OF SPEAKING TO A WHALE. Here the short term gets serious. If we learn to generate credible signals, someone will use them. Ethicists are already cataloguing the risks (in the journal AI and Ethics, 2025) and the debate has reached the mainstream press (Bloomberg, September 2026): playback — broadcasting recorded or synthetic sounds to wild animals — could alter songs, migrations or entire cultures within a population; tour operators could lure wildlife with AI-generated calls; and fishing or poaching would gain the perfect decoy. This is not theoretical: human hunters have mimicked animal calls for centuries; AI merely industrializes the deception, just as it is industrializing fraud among humans — a thesis we have argued here before. The field itself has two souls: David Gruber (CETI) defends listening without speaking ('I really just want to hear and know more about them,' he says), while Brenda McCowan (Whale-SETI) counters that without playback there are no testable hypotheses. Both are right — which is exactly why this cannot rest on each lab's personal ethics.

OUR READING. Optimism, but in the right order. In the short term, the honest statement is that we will not talk to animals this decade, and that the most real risk is not misunderstanding them — it is manipulating them well. Governance must come before capability: playback protocols, limits on commercial use of synthetic signals, and something like an agreed 'contact protocol' across projects; NYU Law's MOTH program is already exploring which rights understanding another species would trigger. In the long term, this is one of our favorite uses of AI: pure knowledge. A partial functional dictionary of sperm whale would be enough to redesign shipping lanes, fishing gear and protected areas with surgical precision; planet-scale listening will be to ecosystems what monitoring is to a computer network. And there is a bigger prize. Every time science showed that 'they' were more like 'us', the moral circle widened. If AI confirms that we share the planet with other complex minds, the question will stop being whether we can talk to them and become what we owe them. On that day, the microscope will have been worth it — even if the conversation never comes.