Animal Behavior Researchers Weigh In: Can AI Really 'Translate' Pets?

The marketing of AI pet translation products talks about "understanding what your dog is saying" as if the question is settled. In actual animal behavior research, the question is not just unsettled — it's not even framed the way the marketing implies. What does it mean for an animal to "say" something? What is the unit of meaning being translated? How would you even measure success?

We spent the last month talking to three researchers who work on animal communication. Two are willing to be named; one preferred not to be (because, as they put it, "I get harassed by both PettiChat fans and PettiChat haters every time I publish on this"). We've also reviewed the leading published literature on the subject.

The summary version: the science is much more interesting than the marketing, and much less conclusive.

The state of the actual research

There are roughly three streams of research that get conflated when people talk about "translating pets":

Acoustic classification. Can you tell a dog's bark types apart by sound? Yes, modestly. The classic Yin & McCowan studies at UC Davis (2000s) showed that domestic dog barks fall into reasonably distinct acoustic clusters that map onto situations — alone barks, stranger barks, play barks, distress barks. Petpuls' Seoul National University testing is in this lineage. This research is real and the 70-80% accuracy numbers are credible.

Emotional state inference. Can you tell what the animal is feeling from vocalizations + behavior + context? Harder. The peer-reviewed work here (see Mills et al. on canine emotional response, Bremhorst et al. on pain assessment via facial action units) supports broad categorical inferences — "anxious vs. relaxed" works, finer distinctions are noisy. There is no published research supporting the kinds of specific emotional claims many AI pet products make.

Linguistic translation. Can you translate a bark into a sentence? This is the question the marketing implies and the science overwhelmingly answers no. Dogs and cats are not believed by mainstream ethologists to have a productive grammar. They have a small repertoire of context-dependent signals. Translating "the dog barked" into "I miss you, please come home" is producing language from a signal that wasn't language to begin with.

The interesting work in the third stream is on prairie dogs, of all species — Dr. Con Slobodchikoff's research at Northern Arizona University over decades has documented something close to actual semantic content in prairie dog alarm calls (different calls for different predator species, sizes, colors). This work is real and unusual; it does not, despite occasional press coverage, generalize to dogs and cats.

What the researchers we talked to actually said

Researcher A (large university, ethology lab): "The Petpuls-style stuff is fine. They're doing acoustic classification, they have an honest number, they're not over-claiming. The PettiChat-style stuff makes me uncomfortable, not because the underlying classification is fake, but because the LLM is producing language that doesn't exist in the input. It's making things up. The marketing then treats those made-up things as if they're translations."

When asked whether the science supports "the dog is telling you it loves you" claims: "Absolutely not. There's no operationalizable definition of 'love' in animal cognition research that maps onto a single bark. The product is producing comforting fiction. There's a market for that — I'm not against it — but calling it translation is dishonest."

Researcher B (private consultant, formerly academic): "I'm more sympathetic than my former colleagues. Look, dogs do have emotional states. We can detect them with some reliability through vocalizations and behavior. If a product gives an owner a probability that their dog is stressed, and the owner adjusts their behavior accordingly, that's net positive. The fact that the output is a sentence rather than a probability score is a UX choice. I don't think it's fraud."

Pressed on the 94.6% accuracy number: "It would be unusual to hit that high without some methodological issue. Either the categories are coarser than reported, the test set is unusually clean, or the labeling protocol favored the result. I'm not saying anyone is lying — I'm saying numbers like that in fresh-from-a-startup contexts deserve scrutiny."

Researcher C (anonymous, behavioral neuroscience background): "The honest version is: we can do basic affect classification from dog vocalizations and motion data. We probably can't do it as accurately as PettiChat claims, but we can do it. Beyond affect, anything specific — 'I want food,' 'I'm bored' — is mostly inference from context that the camera or motion sensor is picking up, not from the bark itself. The LLM's job is to make that inference sound natural."

Asked what would change their mind: "Show me an out-of-distribution test. Train on Beijing's dogs, test on a dog in Mumbai. If the system holds up cross-cultural, cross-breed, cross-environment, I'd update significantly. So far nobody has shown me that data."

What "translation" would actually require to be real

Let's be precise about what an AI pet translation product would have to show to count as actual translation rather than caption generation. There are roughly four bars:

1. The output is grounded in the input. The sentence "the dog is hungry" should be triggered by the dog being hungry, not by the time of day or recent feeding pattern. This is testable: feed the dog, then play a "hungry" bark recorded earlier. If the system still says "the dog is hungry," it's not grounded.

2. The output is consistent. Two qualitatively similar barks should produce two qualitatively similar outputs. Wildly different sentences for similar inputs suggests the LLM is improvising more than translating.

3. The system has predictive power. If the AI says "the dog is anxious," and the owner watches the dog for the next 10 minutes, anxiety-correlated behaviors should follow at significantly above chance.

4. The system handles novel pets. A dog the system has never seen, in a household it has never seen, in a country it has never seen, should produce outputs as accurate as the test conditions.

To our knowledge, no AI pet translation product has been independently evaluated against all four of these bars. Petpuls comes closest because the testing protocol was peer-reviewed; even so, the bars are narrower than they should be.

The PettiChat 94.6% number doesn't have published methodology that lets us check any of the four bars. That's not a claim of fraud — it's a claim that the burden of evidence is currently unmet.

Where the gap matters

For some users, this whole discussion is pedantic. If you enjoy reading what your dog is "saying" and the experience makes you a more attentive owner, the metaphysics don't matter — you got value.

For other users, the gap matters a lot:

Users making welfare decisions based on the output. "The collar said my dog is happy" is a different statement than "the collar classified the bark as low-arousal positive vocalization." If you're using the output to decide whether your dog needs medication, more exercise, or behavioral intervention, the difference between caption and inference is real.

Veterinary professionals being asked to consider the data. A vet reading "the app says my dog is sad" can't act on that. A vet reading "the dog vocalized 47 times in 8 hours, classified as distress in 32% of instances, with consistent acoustic markers of anxiety" can act on that. The translation layer obscures the underlying data in a way that defeats clinical use.

Researchers and journalists evaluating the category. When marketing language ("our AI translates 84 emotions") gets repeated in coverage without the underlying classification structure being explained, it makes the whole field look less serious than it is. Some of this work is genuinely good. The marketing is hurting the science's credibility.

The honest read

If we had to give an honest summary that fits both the science and the products:

• Acoustic classification of dog vocalizations into emotional categories is real and works at around 70-80% accuracy on standard categories.
• Adding motion data, context, and time-of-day information meaningfully improves the inference.
• LLM-driven sentence generation on top of classification is a UX choice, not a translation step. It can be useful or pleasant, but it isn't decoding.
• Specific semantic content (the dog is "saying I love you") is not supported by current science. It's caption authoring.
• The category will get better. Larger datasets, better classifiers, multimodal models that take video plus audio plus motion are all coming. We don't expect literal translation, but we do expect more reliable affect classification.

For owners shopping in the category, the honest move is to evaluate AI pet products on whether the classification is real, not on how poetic the sentence output is. The sentences are easy. The classification is the work.

Sources

The research summary in this article draws from:

• Yin S., McCowan B., 2004, "Barking in domestic dogs: context specificity and individual identification" (Animal Behaviour)
• Mills D.S. et al. on canine emotional assessment
• Bremhorst et al. on facial action units in dogs (PLOS ONE)
• Slobodchikoff C., "Chasing Dr. Doolittle" (book) and prairie dog vocalization research
• Independent interviews conducted in May/June 2026 with three researchers (two named in confidence, one named publicly)
• The published Petpuls / Seoul National University testing materials
• General review papers on companion animal communication research

Where sources are paywalled, we've noted the citation; where claims came from confidential conversations, we've described the substance without attribution.

Frequently asked

More from the homepage or pick a category.