This ‘Talking Dog’ Learned 100 Words On A Soundboard — A Biologist Explains What That Proves

The dog behind this headline is Bunny, a sheepadoodle in Tacoma, Washington, whose soundboard videos have drawn millions of followers online. Her owner, Alexis Devine, has tracked Bunny’s vocabulary climbing steadily since 2020 — from a single “outside” button to just over 100 words today.

What makes Bunny more than an internet curiosity is that she didn’t stay one: she became a research subject. Cognitive scientist Federico Rossano at UC San Diego enrolled her in a study called TheyCanTalk, which eventually grew to include well over 1,000 dogs and other pets, all wearing the same basic setup — a floor mat of buttons, each one playing a recorded word.

To understand why Rossano’s team built their study the way they did, it helps to go back more than a century, to a horse in Berlin named Hans. His owner, a retired math teacher named Wilhelm von Osten, believed people badly underestimated animal intelligence, and spent years teaching Hans to tap out answers with his hoof — arithmetic, spelling, even the day of the week for a given date. By 1904 crowds were gathering to watch, and Hans could apparently solve problems even when von Osten wasn’t the one asking. That convinced a panel of 13 experts, assembled by the German board of education, that no deliberate trick was involved.

Then a young psychologist named Oskar Pfungst took over the real investigation, and he tested one variable at a time. Could Hans still answer if the questioner stood farther back? If the person asking didn’t know the answer themselves? If Hans couldn’t see them at all?

The pattern that emerged was stark: when the questioner knew the answer and stood in Hans’s view, the horse was right the overwhelming majority of the time. When the questioner didn’t know the answer, Hans’s accuracy collapsed to almost nothing. Pfungst eventually traced it to something almost invisible — a small, involuntary shift in posture or tension that questioners made without noticing, right as Hans’s tapping reached the correct count. The horse had never been doing arithmetic. He’d been reading people, with a precision that, in its own way, was just as remarkable.

That episode gave science a permanent name for the trap: the Clever Hans effect — any case where an animal appears to understand language or concepts, but is actually just picking up unintentional cues from a human who already knows the answer.

This is exactly the shadow every soundboard dog operates under, and it’s why Rossano’s team designed their experiment around ruling it out. In a 2024 study published in PLOS ONE, they tested 59 dogs on real words — “outside,” “play,” “food” — alongside a nonsense word invented purely for the study, one the dogs had never heard and couldn’t have any pre-existing association with. If dogs responded to the real words but not the nonsense one, and did so even when the tester didn’t know which word was which, that’s a dog reading the word, not a modern Hans reading his owner.

A separate line of research looked at what happens when a dog presses two buttons back to back. That study found the pairings weren’t random and weren’t simply echoing what a person had pressed moments earlier — a real signal, however narrow, that something more than chance is happening, published in Scientific Reports. Rossano has described some of Bunny’s own multi-button sequences as having a structure that resembles how a toddler strings early words together, not full grammar, but not noise either.

Even researchers excited by these results stay guarded. A 2025 study published in Scientific Reports found that dogs’ ability to recognize a recorded word drops noticeably as the audio degrades, and soundboard buttons, with their tinny playback, are often the weaker signal compared to a human voice.

That raises a real possibility: some of what looks like comprehension may still lean on context — the time of day, a hand near the door — more than the word itself. Rossano’s team has been candid that the open question is still whether a dog like Bunny is generating genuinely new combinations, or has simply learned which sequences get rewarded (a subtler version of the same problem Pfungst uncovered in 1907).

The doubt doesn’t just end there. Learning that a paw-press reliably produces an outcome — a walk, a treat — is well within what operant conditioning, the basic process by which any animal links an action to a consequence, already explains. Language requires more: rules for combining units into new meanings, not just a longer list of memorized requests. Most researchers land in the middle: dogs like Bunny are almost certainly doing more than blind association, but there’s no solid evidence yet of full human-style grammar.

What does hold up, cautiously, is that dogs seem to build something like categories — food-words, place-words, people-words — solid enough that stringing two together isn't just static. Bunny herself remains one of the best-documented cases of this, precisely because she's been tested, not just admired.

None of this proves dogs have grammar. But it proves the underlying premise wasn’t wishful thinking : animals really can be given a structured channel for expression, and researchers really can test whether the other end of that channel is meaningful. That premise is already being pointed at far higher stakes than “walk” or “treat.”

The most immediate payoff is welfare, not conversation. A soundboard-equipped dog that can distinguish and reliably press “hurt,” “scared” or a specific body part gives a veterinarian information a wagging tail or a whimper never could, and some behaviorists are already exploring buttons as a diagnostic tool for exactly that reason. The same logic scales to working and service animals, where a richer, testable channel for a dog to flag distress or a problem could make partnerships with humans measurably safer.

Zoom out further and the same toolkit is already running on wild animals, at far greater scale. A 2024 study published in Nature Communications, part of the MIT-led Project CETI, used machine-learning models trained on tens of thousands of sperm whale clicks to find structure resembling a phonetic code — a project that treats a whale pod the way Pfungst once treated Hans, testing rigorously rather than assuming.

A 2024 study published in Nature Ecology & Evolution reported that wild elephants use distinct calls that function something like names for each other. None of this is decoding in the “translate it live” sense yet, but the direction of travel is real: better microphones, more data and far more powerful pattern-recognition than existed even five years ago.

Put together, the honest, cautious version of this story is also the exciting one: after a century of being warned off by a horse in Berlin, biologists finally have the tools to ask “does this animal understand me?” and get a rigorous answer instead of a guess — and the early answers, from a sheepadoodle’s button mat to a sperm whale’s clicks a mile down, keep coming back more interesting than expected.

Ever wonder just how deep your own bond with your dog really goes? Find out with this science-backed test: Pet Owner Connectedness Scale