The first thing I ever said to my dog was, “Do you want to come home with me?” He was six pounds, and 10 weeks old. He craned his head forward and sniffed my mouth.
In the four years since, I have continued to pepper him with questions that he cannot answer. I ask him what he’s up to, if he wants to go for a walk, if he’s feeling sleepy. When he is sick, I ask him what is wrong; when another dog growls at him, I pull him aside to ask if he’s okay. He does what he can to relay his thoughts back to me: He barks; he sighs; he scratches at the door.
But of course we have never talked to each other, not really. Some 15,000 years since humans first domesticated the wolf, scientists have learned that different barks mean different things—for instance, dogs use lower, longer barks for strangers—but our understanding of dog communication remains rather limited. (Researchers are careful to call it communication, not language, because no animal has been shown to possess the same complexity of verbal systems as humans.)
Although a bark at a squirrel is easy enough to decipher (I will eat you!), humans have more trouble knowing whether a whine is just a dog having random feelings on a Tuesday—or something far more serious. Dog owners often joke about how they’d give up years of their life just to have a chance to talk to their pet for a single hour or day. Meanwhile, hucksters posing as dog whisperers and pet psychics have happily taken their cash by claiming to be able to help them translate their dogs’ inner thoughts.
Now, amid a wave of broader interest in applications for artificial intelligence, some dog researchers are hoping that AI might provide answers. In theory, the technology is well suited for such a purpose. AI, at its core, is a pattern-recognition machine. ChatGPT is able to respond in language that seems human, because it has been trained on massive datasets of writing, which it then mimics in its responses. A similar premise applies to other generative-AI programs; large language models identify patterns in the data they’re fed, map relationships among them, and produce outputs accordingly.
Researchers are working with this same theory when it comes to dogs. They’re feeding audio or video of canines to a model, alongside text descriptions of what the dogs are doing. Then they’re seeing if the model can identify statistical patterns between the animals’ observed behavior and the noises they’re making. In effect, they’re attempting to “translate” barks.
Researchers have used similar approaches to study dog communication since at least 2006, but AI has recently gotten far better at processing huge amounts of data. Don’t expect to discuss the philosophy of Immanuel Kant with Fido over coffee anytime soon, however. It’s still early days, and researchers don’t know what kind of breakthroughs AI could deliver—if any at all. “It’s got huge potential—but the gap between the potential and the actuality hasn’t quite emerged yet,” Vanessa Woods, a dog-cognition expert at Duke University, told me.
Right now, researchers have a big problem: data. Modern chatbots are trained on large collections of text—trillions of words—that give them the illusion of language fluency. To create a model capable of translating, say, dog barks into English (if such a thing is even possible), researchers would need millions, if not billions, of neatly cataloged clips. These barks will need to be thoroughly labeled by age, breed, and situation—separating out a 10-year-old male labradoodle barking at a stranger from a six-week-old bichon frise puppy playing with its littermate.
No such catalog currently exists. This is one of the great ironies of the project: Dogs are all around us, constantly captured by phones and doorbell cameras and CCTV. You don’t need to watch Planet Earth to see the canine living in its natural habitat; the internet is filled with more clips of dogs than anyone could watch in a lifetime. And yet all of this media has never been cataloged in a serious way, at least not on the scale that would be necessary for us to better understand what their barks mean.
Perhaps the best catalog that exists is from researchers in Mexico, who have systematically recorded dogs in their homes in specific situations, getting them to bark by, say, knocking on a door or squeaking a favorite toy. A research team at the University of Michigan took some of the 20,000 recordings included in the dataset and fed it into a model trained to recognize human speech. They played barks for the model, and then had it predict what they were barking at, just based on sound. The model could predict which situation preceded the bark with about 60 percent accuracy. That’s nowhere near perfect, but still better than chance, especially considering that the model had more than a dozen bark contexts to pick from.
The same approach of using AI to decipher dog barks is happening with other animals. Perhaps the most promising work is with whale chatter, as my colleague Ross Andersen has written. Other researchers are tackling pigs, bats, chimpanzees, and dolphins. One foundation is offering up to $10 million in prize money to anyone who can “crack the code” and have a two-way conversation with an animal using generative AI.
Dogs probably won’t be the animals that help scientists win the prize. “I do not think they necessarily use words and sentences and paragraphs,” Rada Mihalcea, a co-author of the Michigan study, told me over Zoom. (Naturally, in the middle of our call, a stranger knocked on my door, causing my foster dog to bark.) As much as dog owners like myself might want something akin to Google Translate for dogs, Mihalcea’s starting with much more narrow ambitions. She hopes this line of research can “help us get an understanding into what is even there as a linguistic system—if there is such a system.”
Another research group, led by Kenny Zhu at the University of Texas at Arlington, is taking a different approach. His team is scraping massive amounts of dog videos from YouTube. But the data are extremely noisy—quite literally. The researchers have to isolate the barks from all the other sounds that happen in the background of the videos, which makes the process onerous. Zhu’s team does have preliminary findings: They had their algorithms process the sounds of six different breeds (huskies, Shiba Inus, pit bulls, German shepherds, Labradors, and Chihuahuas), and believe they’ve found 105 unique phonemes, or sound units, that span all the breeds.
Even if researchers are able to eventually get a perfect dataset, they’ll run into another problem: There’s no way to know for sure that whatever observations the AI makes is right. When training other AI models on human languages, a native speaker can verify that an output is correct, and help fine-tune the model. No dog will ever be able to verify the AI’s results. (Imagine a dog sitting in an academic research lab, nodding solemnly: Yes, that’s correct.“Ruff-ruff-ruff” means“Give me the chicken.”) The dream of AI as an intermediary between humans and dogs faces a fundamental bias: It is human researchers who are using human-made AI models and human ideas of language to better understand canines. No matter how good the technology gets, there will always be unknowns.
The focus on better understanding dogs’ verbal noises can obscure how much we already know about them. Dogs have evolved to better communicate with humans: Their barks have changed, and their eyes have grown more expressive. Feral dogs and wolves bark less than pets, suggesting that humans are a big reason why our pups make noise. “The whole thing about dog genius is that they can communicate with us without speaking,” Woods told me. “We can also read them really clearly, which is why we’re so in love with them.”
I know what she means. During a heat wave this summer, I decided to buy heat-resistant dog boots to protect my pup from the scorching pavement. You put them on by stretching them over your dog’s paws, and snapping them into place. The first time I put them on my dog, he stared at me. When I tried to walk him in them later that week, he thrashed in the grass and ran around chaotically. He did not want to wear the boots. And I did not need an AI to know that.