From Dog Soundboards to Smarter AI: What Animal Communication Reveals

Story by:

• Christine Clark- [email protected]

Media contact:

• Christine Clark- [email protected]

Share This:

• Share this story on Linkedin
• Share this story on Facebook
• Share this story on Threads
• Share this story on Twitter
• Share this story via email

Article Content

Why does studying animal communication matter - and what might it teach us about building better AI? What can humans learn by mapping the cognitive abilities of other species, including their own pets? Federico Rossano, associate professor of cognitive science at the University of California San Diego and director of the Comparative Cognition Lab, is pursuing those questions with groundbreaking work to understand animal intelligence - most famously through his research on how dogs use soundboard "word buttons" to communicate with humans.

What began as viral videos of dogs "speaking" through soundboards - pressing buttons to play recorded words and combining them in context - has become a rigorous scientific effort. Rossano was the first researcher to pair controlled experiments with one of the largest citizen-science datasets in animal communication, gathered from dogs and their owners in real homes around the world. Rossano's research doesn't claim that dogs can talk, but his published studies suggest that button-trained dogs aren't simply copying their owners - they're learning a shared communication system, responding to particular words as cues with consistent meaning and, for some dogs, stringing words together in recurring two-button pairings that appear purposeful rather than accidental.

The work has resonated far beyond academia, drawing global media attentionand recently featuring in a NOVA documentarythat has been viewed by roughly four million people on PBS and YouTube. Rossano recently sat down with university communications to discuss what secrets his research unlocks about intelligence in other species, why it matters, and what its bigger-picture implications could be - from how we interpret animal minds, to how we evaluate whether AI systems genuinely understand meaning or are simply producing convincing behavior.

From viral videos to rigorous science: what's the real impact of your work so far? How has your research changed - scientifically and culturally - how we investigate animal communication?

There is a general belief that science requires highly controlled lab testing and anything else is necessarily problematic. But that is a very 20th-century way of thinking of what science is all about. Historically, if we wanted to study animal minds or animal communication, the researcher would bring animals to the lab and control every aspect of their life - this also meant that many animals were kept separate from their conspecifics, living alone or with a single other member of their species. For social animals, you can only imagine what the effect of such social deprivation would be on their minds and behaviors. This scenario has been true for dogs as well: bring them to the lab to be tested in an empty room in an unfamiliar environment by a stranger and let's see what they can do so '...let's put our study participants in very stressful and uncomfortable situations and act like we are observing their everyday behavior… My hope with our project with dogs and cats using soundboards is to show that science can be done also outside the lab, that there are ways to control the training that do not require bringing the animal to the lab. And of course, experimenters can travel to test the animals in more familiar and comfortable environments. And we always compare performance with a stranger with performance with the pet owner because motivation to engage and attention would likely differ. In the modern world, problems come in a messy form, but we now have tools to analyze messy data and to do that at scale. And we can look at data that better reflects the day-to-day experiences of the animals, rather than some simplified scenario that is completely disconnected from anything they would ever experience.

I am also proud of the citizen science part of the study. Many scientists refrain from it again because the claim is that we cannot trust what people do at home or when the scientists are not controlling every step of the process. And in some cases that is true. But I have found that engaging with the community leads to exciting insights that would be completely lost if we refrain from engaging with the people who actually live with pets every day. And I am truly amazed by how motivated people are to help scientists and do scientific research. As a scientist, I think we need to get out of "the ivory tower" and try to engage and connect and explain what we do and why we care about it. Because if you are asking interesting questions, people will listen and try to help find some reliable answers.

Your work suggests dogs respond to the recorded words from the buttons - not just subtle owner prompting - and that they can use buttons in consistent, context-appropriate ways across situations. What are the implications for AI - especially for building systems that communicate with humans and for testing whether an AI system understands meaning rather than just performing well?

At a very basic level, current LLMs are doing the same that these dogs are doing at least in the beginning: learning to associate patterns with specific outcomes. So if dogs see that when they push "treat" they get a treat, that is pretty highly rewarding and if they can push sequences and they get rewarded enough times, or they see them produced in that exact sequence many times, they are learning how to structure things to get the desired outcome. LLMs are learning statistical regularities, which in humans occur because that is the appropriate, normative way to do it. The issue with current AI is that we do not quite know to what degree AI understands "meaning" because they do not have "world knowledge." Or better, when AI uses words in a correct sequence it does not quite mean that it understands what each word means. My two-year-old daughter told me "I need a credit card" but that was in response to spending the day with me going to different shops, and me buying toys. She clearly has no idea of what a credit card is… but realized that if she kept asking me for it, that once she would get it, she would get toys and spend the day with dad.

The challenge of current AI research on animal communication is that finding patterns is what AI is amazing at, but without "ground truth", i.e. without knowing what that communicative signal actually means, all we would know is that AI can identify several different signals and predict what follows them, without really knowing what those communicative signals are trying to actually communicate, content-wise.

In regards to other species, we can easily figure out if whales, for example, are singing and even identify how many songs they have in their repertoire… but figuring out what they are singing about and why they are singing that song at that particular moment and to that particular addressee is a very different question.

What are some of the strides scientists have made in understanding animal communications that leverages AI, which in turn, can also help build better AI and are you part of this effort?

In some areas of animal communication we have made humongous progress. I am thinking of examples of primate communication, starting from the early work of Jane Goodall to the dozens of scholars now studying gestures, vocalizations and facial expressions in great apes and other primates. And that is why my lab has started a couple of years ago a project aimed at using AI technology (especially computer vision and multi-modal models) to build a foundational model to study primate behavior. Not just communication, but behavior in general. Because by understanding what they are doing in terms of activities and who they are doing it with and when, we can learn a lot about context and improve our animal communication models. We call it "Primate-GPT" and it is part of a large international collaboration with several outstanding researchers worldwide. It will take a while but we truly believe that the payoff will be extremely valuable.