Most of the time when we talk about AI, we’re really talking about output: more code, content, or images generated on demand.
But one of the most disorienting ideas we heard at the SXSW Conference this week wasn’t about what AI can do for us. It was about what AI might finally let us hear.
In a keynote that felt more like fieldwork than a product demo, Aza Raskin, Co‑founder of the Earth Species Project, walked through how researchers are using AI to decode the communication systems of other species like dogs, whales, birds, primates, and more.
Not as a party trick, but as a serious attempt to answer a strange question:
If intelligence is all around us, what would it take to actually listen?
Instead of showing yet another tool that turns prompts into text or images, Raskin walked through something stranger and much more technical: using AI to help decode the communication systems of other species.
This isn’t the sci‑fi fantasy where you press a button and your dog suddenly speaks English. Nobody is shipping a “talk to your golden retriever” app. What these teams are doing is closer to reverse‑engineering an unknown protocol.
The work starts in the least glamorous place: with an unusually large and varied pool of raw data.
Researchers record thousands of hours of audio (barks, songs, clicks, rumbles) across multiple species, including dogs, whales, birds, primates, and others. Each sound is paired with sensor data that describes what was happening at the time: where the animal was, when the call occurred, what the environment was like, and how the animal was moving.
On its own, this combination of sounds and signals looks like chaos. But that’s intentional. Before a model can find structure, it has to see the full complexity of real behavior: overlapping calls, background noise, and subtle shifts a human ear would miss. That raw, high‑dimensional dataset becomes the foundation for everything that follows.
Once the raw recordings are in place, the next step is to give those sounds meaning by adding context.
For each sequence of calls, researchers annotate what was happening around it: which animals were nearby, whether there was food or a predator present, if a newborn had just arrived, and whether the group was moving, resting, playing, or under stress. They also capture what happened immediately afterward: did the animals approach, retreat, call back, or change direction?
By pairing each vocalization with this situational detail, the dataset shifts from “audio plus numbers” to “signals plus consequences.” The model is no longer learning from sound in isolation; it is learning from sound embedded in real scenarios.
And this is exactly the point we come back to again and again when we talk about AI: context is what makes a model useful. Without it, you only get pattern‑matching on noise. With it, you can start to connect specific patterns to consistent outcomes, and that’s where real insight (for animals or for software systems) begins.
While different research groups use different setups, a typical approach to this kind of problem looks something like this.
With sound and context in place, the models can finally begin to look for structure.
They search for patterns of calls that reliably show up before certain behaviors, sequences that tend to appear together, and recurring motifs that behave like reusable “words.” Under the hood, researchers use modern machine‑learning techniques to group similar calls into families and to map how those units unfold over time, then relate those patterns back to real‑world situations.
The result isn’t a neat dictionary, but it is a rough language map: clusters of calls that act like a vocabulary, informal rules for how they combine, and regularities in who “says” what, to whom, and when. In short, what once looked like noise starts to behave like a communication system you can actually reason about.
Once you see that process, it gets hard to keep your old mental models.
We’re used to thinking about language as text on a page or words in a UI. This work pushes a different view:
That’s a big shift for anyone building products or models: Your logs, traces, and usage analytics are not “just metrics.” They’re the behavioural language of your system and your users. You can either treat it as noise or treat it like whale song and look for structure.
Traditional analytics often treats data as a byproduct: something produced by a system you already understand. In the animal‑AI work, data is the environment you enter to discover what’s going on:
That mindset is powerful when you apply it back to software. Instead of defining a handful of KPIs and staring at them, you can ask: What are the latent “phrases” in our product usage? Instead of hand‑labeling a few events as “success” or “drop‑off,”you can ask: What pathways consistently precede churn or expansion?
At some point, you stop seeing animals as “things that make noise,” and start seeing them as agents in a live conversation. The same thing can (and should) happen with your stack:
Researchers working on interspecies communication treating animal communication as a living system to be decoded, not a static dataset to be summarized. Most engineering orgs aren’t there yet.
If AI can help us decode dog barks and whale songs, it raises an uncomfortable question: What excuse do we have left for not understanding our own customers and systems?
For AI‑curious engineers and leaders, this work should trigger at least three concrete questions.
If an animal‑language model can find structure in unlabelled noise, what could similar approaches reveal in onboarding flows, abandoned features, or “shadow workflows” teams invent around your tools?
Most teams still apply AI in narrow, output‑focused ways: generating copy, code, tests, or assets, sprinkling in autocomplete, or adding a chatbot layer on top of existing documentation.
Those are all output use cases. But what if you also asked:
If AI can decode non‑human communication, it can certainly help decode the “language” of your own product.
The most striking part of the SXSW conversation wasn’t the novelty. It was the stated goal:
“Open the aperture of our own empathy.”
For an engineering org, that could mean:
The real promise of AI is to stretch our empathy, not just our productivity. Empathy here isn’t sentimental. It’s operational:
Nature has been speaking this whole time. Our products, users, and engineering systems have been speaking this whole time, too.
AI might talk to animals before it replaces engineers. That might be the best outcome we could hope for. Because it leaves us with the question that matters:
If you treated your product data, customer behavior, and engineering workflows like a language to decode, what is your system already trying to tell you that you still can’t hear?