Google made an AI model to talk to dolphins

In a bold step toward bridging the communication gap between humans and one of the planet’s most intelligent animals, Google has helped develop an AI model that could potentially allow people to "talk" with dolphins. This pioneering system, named DolphinGemma, is expected to debut in the coming months, alongside a companion system called CHAT (Cetacean Hearing Augmentation Telemetry). Together, they’ll be tested to see whether AI can successfully translate, mimic, and even engage with dolphins’ famously intricate vocal language.

For over 40 years, researchers at the Wild Dolphin Project (WDP) have been meticulously documenting the lives and sounds of a community of Atlantic spotted dolphins in the Bahamas. Their database, filled with decades of underwater audio and video, connects dolphin vocalizations like whistles, clicks, and burst pulses to behaviors ranging from courtship to playful disputes — and even evidence of individual dolphins having unique names for each other.

Though scientists have long theorized about the possibility of human-dolphin conversations, technology to decode and reproduce dolphins’ complex underwater sounds has always lagged behind — until now. The recent explosion of large language models (LLMs), like ChatGPT, inspired WDP to explore whether similar AI could parse and predict dolphin speech patterns. Partnering with Google and the Georgia Institute of Technology, they trained an LLM on a massive, labeled dataset of dolphin sounds.

The result is DolphinGemma, built using the same AI technology that powers Google’s Gemini systems. It’s designed around 400 million parameters and works much like predictive text AI for human language, but for dolphin communication. It receives dolphin audio inputs, interprets them, and predicts what sounds should follow — offering potential to mimic and respond in real time.

Paired with the CHAT system installed on modified Google Pixel smartphones, this technology isn’t meant to immediately crack dolphin language. Instead, CHAT aims to establish a simplified, shared set of synthetic whistles that dolphins can learn to associate with specific objects they like, such as seagrass, sargassum, or even researchers’ scarves. The goal is to gradually build a two-way vocabulary — one simple whistle-word at a time — and eventually allow dolphins to make requests or express preferences during interactions.

While true interspecies conversations remain on the horizon, this inventive use of AI represents a huge step forward in marine research, conservation, and our understanding of non-human intelligence. As DolphinGemma prepares for its first tests in the wild, the prospect of humans finally chatting with dolphins no longer seems like pure science fiction — but an achievable future.