Google is learning to speak dolphin – but what are they really saying?

Google has trained an AI to “talk” to dolphins. Science is fascinating. The implications? Maybe a little dystopian.

Disclaimer: This isn’t a Black Mirror plot.

This summer, Google and the Wild Dolphin Project will begin testing Dolphin Gemma, an AI model designed to communicate directly with one of the smartest creatures on our planet.

As dolphins are known to communicate subliminally, the question is: will AI be out of its depth?

The system is built from Google's existing Gemma systems, which are a family of light open-source AI models that can be deployed on various devices.

In this case, Dolphin Gemma is trained on decades-worth of dolphin audio and predictor data using an “audio-in, audio-out” method – mimicking dolphin responses, much like an LLM like ChatGPT, for example – predicts text.

The idea of an AI model “speaking dolphin” may sound fantastical, but technically, it's doing what ChatGPT does – just underwater.

The illusion of understanding

Dolphins show signs of complex cognition – they recognize themselves in mirrors, pass knowledge socially, and use signature whistles akin to names.

However, the software might not be able to understand semantically. Instead, it would most likely be sophisticated mimicry.

Applying human understanding to a realm that’s beyond our scope risks anthropomorphizing a communication medium – and that no real dialogue will actually be reached.

Still, on the off-chance that dolphins did respond meaningfully, we might discover some shared points of reference.

AI's attempt to decode dolphin language mirrors new research on sperm whale communication, questioning whether true understanding is possible.

A mirror with flaws

There’s also a danger that AI could misinterpret a dolphin signal. Just like it hallucinates and is quite error-prone when it comes to recognizing human facial expressions and intentions. How do we know if the dolphin sound isn’t an abstract cry?

One wonders what the end goal is for this kind of enterprise. Of course, there’s merit to using tech to decode non-verbal communication, if you’re wielding it in a crime investigation.

But it seems like we’re on the road to decodification for everything. Unlocking the mystery of the ocean, perhaps we can turn to decoding infant cries or an individual's internal monologues.

Dolphin Gemma is a mirror we’re holding up to another species – but it might also reflect our own desire to control and decode everything.

Whispers in everyone’s hands

The portability of this kind of technology – which is enabled locally on Pixel devices – will be great for researchers, as they’ll be able to record sounds and respond in real time.

The tech will no longer be confined to labs. This means that anyone with the hardware and software would be able to interact with dolphins – and potentially other species, if the curiosity doesn’t stop there.

If someone manipulates dolphin sounds in a way to “train” them, then AI-mediated communication may become trending across the board – and that sounds like a Pandora’s box.

And once the box is open, who decides what’s worth saying – and in whose voice?