Google has initiated a notable partnership with the Wild Dolphin Project (WDP) and researchers from Georgia Tech, applying its artificial intelligence capabilities to the long-standing puzzle of dolphin communication. The collaboration introduces DolphinGemma, an AI model designed to analyze decades of acoustic data gathered from wild dolphins. The system uses Google Pixel phones for in-the-field analysis, aiming to identify patterns within dolphin vocalizations and potentially pave the way for basic forms of interaction between species.
Co-authored by WDP founder Dr. Denise Herzing and Google DeepMind/Georgia Tech professor Dr. Thad Starner, the announcement highlights the use of WDP’s unique dataset. Since 1985, WDP has conducted underwater research on a specific community of Atlantic spotted dolphins in the Bahamas, meticulously logging audio and video recordings alongside behavioral observations across generations.
This deep historical context allows researchers to link specific sounds to behaviors – for instance, the use of unique signature whistles for individual contact (like mothers and calves reuniting), burst-pulse “squawks” often heard during fights, or click “buzzes” associated with courtship or chasing sharks. Yet, fully comprehending the structure and potential meaning remains a considerable challenge. As Herzing noted in comments shared with Ars Technica, “We do not know if animals have words.”
Decoding Dolphin Sounds With AI
DolphinGemma attempts to find underlying structures within this complex acoustic data. The model is based on Google’s Gemma architecture—a family of relatively lightweight, often open AI models distinct from the company’s larger, proprietary Gemini series. It employs Google’s SoundStream technology, an efficient neural audio codec likely used to convert the high-frequency, complex dolphin sounds into a tokenized format suitable for AI processing.
With approximately 400 million parameters—a modest size facilitating on-device operation—DolphinGemma processes sound sequences to learn patterns and predict subsequent vocalizations, much like how language models generate text. This pattern analysis could help researchers spot previously unrecognized regularities in dolphin communication.
The project also integrates with and aims to enhance the existing CHAT (Cetacean Hearing Augmentation Telemetry) system, an underwater computer interface developed previously by WDP and Georgia Tech. CHAT experiments involve associating novel, synthetic whistles played by researchers with specific objects (like scarves or sargassum), testing if dolphins learn to mimic the sounds to make requests.
DolphinGemma, running locally on the Pixel 6 and planned Pixel 9 phones housed within the CHAT device, intends to accelerate the real-time recognition of potential dolphin mimics against background ocean noise, allowing researchers to respond faster to reinforce any observed learning.
AI for Science Context
DolphinGemma joins a portfolio of Google DeepMind projects applying AI to various scientific problems. AlphaFold, its AI system renowned for accurately predicting protein structures, is widely used by researchers and contributed to a Nobel Prize in Chemistry in 2024 for DeepMind CEO Demis Hassabis and lead researcher John Jumper.
More recently, on February 19, 2025, Google detailed its work on an AI co-scientist designed to generate novel research hypotheses by analyzing scientific literature. Earlier, in October 2024, Hassabis had spoken about AI’s potential role in planning and predicting experimental outcomes. DeepMind has also demonstrated advanced reasoning capabilities with models like AlphaGeometry2, which combines neural networks with symbolic logic to solve complex geometry problems at an elite level.
The Question of Openness
A key aspect of the DolphinGemma announcement is Google’s plan to release the model openly this summer for academic use, stating it might be adaptable for other cetacean species. This commitment to openness is interesting when viewed against recent trends.
Recent reports suggest that DeepMind is increasingly restricting the publication of research related to core competitive technologies like Gemini, citing competitive pressures. Examples given included the proprietary nature of Gemini Robotics models and the non-release of AlphaGeometry2.
The release history of AlphaFold 3 also provides relevant context. Initially launched in May 2024 with access limited to a web interface, significant pushback from the research community preceded the eventual open-sourcing of the code in November 2024. However, this release came under a non-commercial license, and access to the crucial model weights still requires an application.
DolphinGemma’s foundation on the Gemma architecture of AI models, which Google positions as more open (similar to the openly released TxGemma toolkit from March), might explain its different release trajectory compared to core Gemini or specialized models like AlphaGeometry2. Nonetheless, the specifics of the DolphinGemma release this summer will be closely watched.
Challenges and Future Work
While AI systems like DolphinGemma provide powerful tools for sifting through vast datasets and identifying potential patterns, the scientific process still relies heavily on human expertise for validation and interpretation. Any structural patterns or hypotheses generated by the AI must be rigorously checked against observations and potentially tested through further experiments.
The WDP’s long-term observational approach remains central, with AI serving as an analytical accelerator. The upcoming summer 2025 field season, deploying the upgraded CHAT system with DolphinGemma on Pixel 9 phones, will be the first real-world test of this new AI-assisted approach to listening in on the complex social lives of dolphins.
