Researchers have reached a new milestone in interspecies communication with the debut of DolphinGemma, the first large language model (LLM) specifically fine-tuned on four decades of wild dolphin data. Denise Herzing, founder and research director of the Wild Dolphin Project, presented the findings on November 4, 2026, at the inaugural WIRED World Fair in Miami.
The model represents a collaborative effort between Google DeepMind and Georgia Tech, applying transformer architecture to the bioacoustic signatures of Atlantic spotted dolphins. Unlike previous efforts that focused on manual decoding or simple mimicry of artificial labels, DolphinGemma utilizes audio-in and audio-out capabilities to predict and generate vocalization sequences based on 40 years of field research.
From Mimicry to Predictive Modeling
The project marks a transition in how scientists approach non-human linguistics. For years, the Wild Dolphin Project utilized the Cetacean Hearing and Telemetry (CHAT) system, an underwater acoustic interface designed to see if dolphins would adopt artificial “whistles” for specific objects like scarves or balls. While successful in establishing basic label-based interaction, the new model moves beyond these controlled prompts.
According to Google DeepMind, DolphinGemma is a multimodal system. It correlates acoustic data with synchronized underwater video footage, allowing the AI to analyze how specific body language and social contexts coincide with vocalizations. This approach seeks to decipher intent rather than just identifying recurring sounds, effectively treating dolphin communication as a complex, contextual language rather than a series of isolated commands.
The hardware used to facilitate these interactions has also seen a significant overhaul. The original CHAT system, which once required researchers to wear bulky, chest-mounted submersible computers, has been miniaturized. Researchers now utilize two smaller wearable devices—one for the wrist and one for the chest—allowing for more fluid movement and natural interaction while in the water with the dolphins.
Open-Source Scientific Tool
Despite the advanced predictive capabilities of the model, Herzing clarified during the Miami presentation that the system is currently an “interface” for pattern recognition rather than a universal translator. The goal is to provide a framework for understanding the structure of dolphin sequences, which can then be tested in the wild to see if the animals respond to AI-generated whistles in predictable ways.
To accelerate research in the broader scientific community, DolphinGemma is intended to be an open-source tool. By releasing the model, the developers aim to assist other researchers in applying similar LLM architectures to various non-human species, potentially standardizing the way bioacoustic data is analyzed across the globe.
According to the Wild Dolphin Project, the ongoing research continues to prioritize the “CHAT” whistle protocol, where the AI acts as a bridge. This allows researchers to verify if a dolphin is initiating a request or simply responding to a human prompt, providing a clearer look into the cognitive autonomy of the species.
