CetaNauts . org
This is a record of that work: what was demonstrated, what remained unresolved, and the controlled experiment that modern tools now make practical.
Who maintains this. cetanauts.org is kept by the CetaNauts project, which also produces a science-fiction series at cetanauts.com. The two are deliberately separated: nothing invented for the fiction appears anywhere in this record, and every claim here is sourced so you can check it against the original yourself. Where the evidence is thin or a question is unresolved, this site says so rather than filling the gap.
It is easy to assume that talking to another species is a thing people only imagine. It isn't. Human–dolphin interface research has a multi-decade publication record, and now includes work by Google, Georgia Tech and the Wild Dolphin Project.
Entries below cite published sources where those are available. Follow the links.
In 1947 Arthur McBride observed dolphins avoiding fine-mesh nets in dark, turbid water — an early observation suggesting acoustic obstacle detection. Unequivocal experimental demonstration of echolocation came later, in 1960.
John C. Lilly begins scientific research with bottlenose dolphins.
John Lilly publishes Man and Dolphin. The book became a bestseller and helped bring the question of human–dolphin communication to a broad audience.
Dwight Wayne Batteau and Peter Markey deliver the final report of a three-contract U.S. Naval Ordnance Test Station programme (N00123-67-C-1103), run at a lagoon on Coconut Island, Kāneʻohe Bay. Two bottlenose dolphins, Maui and Puka, responded to 35 response-demand messages built on a four-part frame — addressee, imperative marker, content word, terminator. Under a repeat command one animal reproduced four of the trained whistle contours. The report documents speaker independence across two trainers and two translators, and correct responses at 75 yards from the transmitting hydrophone; it also records that command sequencing failed, the second word releasing rather than queueing the first, and that the vocalizations were “not, at that time, under adequate control.” Its closing paragraph specifies the intended next work — extending the vocabulary toward vocal rather than gross body responses, and including the trainer as a responding element. Batteau is reported to have drowned in October 1967 (single secondary source; the report itself does not mention it), and the report was completed by Markey. Whether the specified next experiment was ever carried out has not been established.
Denise Herzing founds the Wild Dolphin Project, beginning long-term underwater observation of free-ranging Atlantic spotted dolphins (Stenella frontalis) on the Little Bahama Bank. The Wild Dolphin Project describes itself as the world’s longest-running underwater dolphin research project. Its multi-decade behavioral and acoustic recordings later contributed training data for DolphinGemma, developed with Google and Georgia Tech.
Laela Sayigh, Peter Tyack, Randall Wells and colleagues publish 'Signature whistles of free-ranging bottlenose dolphins (Tursiops truncatus): stability and mother-offspring comparisons' in *Behavioral Ecology and Sociobiology* 26:247-260 (1990). Working with the Sarasota Dolphin Research Program's multi-decade individually-known population, the study found signature-whistle stability over the sampled years, and sex-dependent patterns of mother–offspring whistle similarity.
Dr. Thad Starner of Georgia Tech's College of Computing begins collaborating with Dr. Denise Herzing of the Wild Dolphin Project, building CHAT (Cetacean Hearing and Telemetry) — a wearable underwater computer for two-way human-dolphin communication. CHAT combines underwater playback, hydrophone recording, and real-time signal recognition in a wearable field system. The method is deliberately NOT decoding: artificial whistles outside the natural repertoire are assigned to objects the dolphins already value.
Herzing, Delfour and Pack publish 'Responses of Human-habituated wild Atlantic spotted dolphins to play behaviors using a Two-Way Interface' in the International Journal of Comparative Psychology — a peer-reviewed study of a formal two-way human–dolphin interface, reporting sessions conducted from 1997 to 2000. The subjects were free-ranging, human-habituated Atlantic spotted dolphins whose engagement varied by individual, age, sex and social context.
Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund and Marco Tagliasacchi (Google Research) publish 'SoundStream: An End-to-End Neural Audio Codec' (arXiv:2107.03312). The architecture uses an end-to-end convolutional encoder-decoder with residual vector quantization to compress audio at bitrates 3-12 kbps with quality matching or exceeding traditional codecs, while running in real time on a smartphone CPU. SoundStream is foundational for subsequent audio LLMs, including Google's DolphinGemma — which uses SoundStream-style tokenization to convert dolphin acoustic data into a sequence-modelable token stream.
July 8 2023 (published April 2026): Project CETI captures the first complete drone + hydrophone record of a sperm whale birth in Unit A, a well-studied Dominica matriline. Ten females (grandmothers, mothers, sisters) + one subadult male named Alan attended the birth of a ~1,000-pound calf to a female named Rounder. Six hours of collaborative lifting at the surface before the calf's fluke unfolded enough to swim. The behavior — phylogenetically traceable back ~35 million years — provides quantitative evidence of what Gruber calls 'one of the most collaborative and empathetic entries into the world that we've ever seen, maybe beyond a human.' A pod of pilot whales (documented antagonists of sperm whales) was visible in the upper-right of the drone footage during the event.
Project CETI discovers sperm whale communication has combinatorial structure like human language.
Google announces DolphinGemma, a ~400M parameter model trained on 40 years of Wild Dolphin Project recordings of Atlantic spotted dolphins in the Bahamas. It uses an audio tokenizer and sequence model to predict likely subsequent sounds. Google said it planned to share the model openly; we have not yet located an official public model release or repository. CHAT Jr uses Pixel phones for real-time field analysis.
Semenzin, Mustun, Dessi, Orhan, Emanuelli, Lakretz, de Polavieja and Sumbre publish Dolph2Vec (arXiv 2606.12503), applying wav2vec-style self-supervised learning to a new corpus of bottlenose vocalizations. The dataset is the notable part: recorded in a natural marine enclosure in the Red Sea where a resident pod of Tursiops truncatus ponticus lives alongside human caregivers and visitors — semi-captive, but ecologically valid. Representation learning only; no vocabulary training and no two-way interface.
Most interface research does not begin by claiming to decode a natural dolphin language.
The approach that has carried this field since the 1980s is close to the opposite. Researchers invent whistles — sounds deliberately built to sit outside anything the animals already say, so a match can never be coincidence — and attach them to things the dolphins genuinely want. Then they wait to see whether a dolphin chooses to use one.
Denise Herzing, who has run this experiment longer than anyone alive, puts it plainly: CHAT is an interface, not a translator.
It is a patient, modest, honest experiment. It has been running in open water since 2011, and the hardware behind it was built at Georgia Tech.
Wild and managed-care studies have each demonstrated useful parts of an interface. The public record we have located does not show one controlled study combining dolphin initiation, a human response, acoustic signals, semantic grounding, tests of generalization, and strong cue controls.
That design has a real strength. Participation is unambiguous, because nothing compels it. A wild dolphin has food, family, and an entire ocean. Whatever interest it shows is genuine.
And it owes a human nothing.
A wild dolphin has an ocean full of problems worth solving. A human's invented word for a scarf is not one of them.
Here is what makes that worth asking about: the studies that did record dolphins using artificial signals to request things — Diana Reiss's underwater keyboards in the 1980s, Louis Herman's language work — were done with animals in human care. Animals that had a relationship with the people asking.
The question this organisation exists to ask is whether motivation, rather than method, is the thing standing in the way.
The timeline above is the short version. Four longer records go underneath it — a timeline of what happened, a library of the stories people have told, a census of every facility on earth that holds dolphins, and an account of each research programme followed to the point where it stopped.
104 events, 1551 to the present — the science, the whaling, the captures, the protections and the deaths. An event earns a place because a whale or a dolphin is what it is about. Forty-two of them forbid generated imagery, and rather than fill those gaps with something invented, each says what its picture will be and where it comes from.
A hundred stories, from 430 BCE to now, each rated for how well the evidence supports it — and rated claim by claim, because a story can be documented in its essentials and still carry a famous detail that is not. Nine of them are rated legendary: told as fact, and contradicted by the evidence. That is a finding, not a rejection.
Our census names 404 institutions. Fifty-three priority rows have received an independent verification pass; most remaining classifications are provisional. We identified one current two-way interface collaboration and did not establish a defensible exact global count.
Seventy years, and how each effort ended →
In 1967 two dolphins responded to thirty-five synthesized commands and reproduced four trained whistle contours on a repeat cue. The Navy report proposed extending the vocal channel and making the human a responding participant. We have not located a public record showing that full experiment was completed.
CetaNauts.org is a small independent research effort based in Georgia. We are not affiliated with any of the programs described above. We think their work is the most interesting science almost nobody is watching.
We have not run an animal trial. We do not have a result. Anything we learn will be published whether or not it supports what we expected.
Any work with animals requires appropriate institutional, legal and welfare review, and would be done only with qualified partners who hold the necessary authorisations. We will say so plainly on this site at every stage.