If you have ever stared at your dog while it stared back at you, and felt an irrepressible urge to know what it was actually thinking, a Chinese startup believes it has built the device you’ve been waiting for your whole life.
The product is called PettiChat. It’s a small AI-powered collar that the company says can translate a dog’s barks or a cat’s meows into full human sentences, in real time, with up to 94.6 percent accuracy. It costs about $118. It launched on Kickstarter in April 2026. Within weeks, more than 10,000 pet owners had preordered one.
The company behind it, Meng Xiaoyi — Mandarin for “cute little translator” — is only four months old. It has already secured roughly $1 million in angel funding and announced ambitions to ship globally before the end of the year.
There is just one slightly awkward problem with all of this. There is no published evidence that the device actually works.
What PettiChat says it can do
The pitch is genuinely striking. According to Dexerto’s reporting, the collar uses a microphone and motion sensors to pick up vocalisations and body language from a pet, runs the data through an AI model built on Alibaba Cloud’s Qwen large-language-model platform, and outputs a translation on the owner’s smartphone in about 1.2 seconds.
The hardware is light — 27 grams — and reasonably engineered. The collar is bite-resistant, IP65 water-resistant, and rated for around 1,000 translations or 100 hours of GPS tracking per charge.
The company says the system was trained on more than one million voiceprint and behavioural data samples gathered from real cats and dogs over two years of testing across more than 1,000 animals. It claims to recognise more than twenty different emotional states — including fear, excitement, hunger, frustration, and affection — and to produce contextual sentence-level translations rather than just labelling broad mood categories.
The viral marketing has leaned hard into specificity. Demo videos show beagles supposedly saying “Leave me ALONE! I don’t like you” and golden retrievers articulating particular frustrations. The implication, never quite stated outright, is that the device isn’t just guessing at moods — it’s translating animal thought into language.
That implication is where the trouble starts.
What the science actually says
Animal communication is a real, active field of scientific research. Multimodal animal-behaviour studies — combining audio, video, and motion-sensor data — are a genuine ongoing area, with research groups including teams at Google DeepMind contributing meaningfully.
But the actual peer-reviewed state of the art is well below what PettiChat is claiming.
According to a detailed assessment of the device, independent studies place the achievable accuracy of analysing the acoustic signal alone at around 57.3 percent. The theoretical ceiling for multimodal systems combining audio, video and posture data is closer to 89 percent. PettiChat’s claimed 94.6 percent accuracy sits noticeably above what scientists have so far been able to demonstrate in controlled conditions — and Meng Xiaoyi has not published the data behind the figure.
There is also a deeper question about what “translation” even means in this context.
Cats and dogs almost certainly do not produce language in the way humans do. They produce vocal signals that convey emotional states, immediate needs, and social cues. An AI model can probably learn to classify those signals reasonably well — the bark that means “stranger at the door” is, statistically, distinguishable from the bark that means “I want to play.” But turning a classified emotional state into a specific human sentence is, at best, a creative reconstruction. The dog isn’t actually saying “Leave me ALONE! I don’t like you.” The model is generating that sentence to represent a detected emotional state.
This isn’t necessarily fraud. It is, however, a marketing leap. The collar may be doing something useful — classifying pet behaviour reasonably well, perhaps better than humans alone could — but the gap between “this dog appears agitated” and a typed sentence in apparent first-person quotation marks is a gap the company has chosen to step across without showing its working.
Why this is selling anyway
Pet owners are an unusually motivated market. The global pet care industry exceeds $300 billion annually. People who already buy GPS collars, premium foods, and customised insurance plans for their pets are not going to be deterred by $118 and an unverified scientific claim.
There is also something psychologically powerful about the idea of the device, regardless of whether it works as advertised. Owning a pet is, in part, an exercise in projection — you read mood into a tail wag, intention into a meow, personality into a head tilt. A device that confirms (or appears to confirm) what you already believe your pet is feeling will feel, to many owners, like it works, because it gives them what they were already looking for.
This is the same psychological territory that has powered horoscopes, personality quizzes, and a long history of products that succeed by telling people what they already half-believe.
PettiChat’s accuracy claim may or may not survive contact with independent testing. Its commercial momentum, in the meantime, doesn’t depend on that. It depends on the much more reliable engine of pet owners wanting, very badly, to believe.
What’s actually impressive about it
It’s worth saying that there is genuine technology underneath all this, even if the headline claim is overstated.
Building a compact, comfortable, bite-resistant collar with reliable sensors and a 1.2-second translation pipeline is real engineering. Training a behavioural model on a million-sample dataset of pet vocalisations and body language is a real machine-learning effort. The team reportedly shares academic lineage with the founders of DeepSeek and Unitree Robotics — two of the more serious AI and robotics groups in China — and they cite Google DeepMind’s animal-behaviour research as inspiration.
The honest version of what they appear to have built is probably this: a real, working pet behaviour classifier, dressed up in marketing language that takes it well beyond what classifiers can actually do. The classification part is interesting. The “translation into human sentences” part is, at best, an interpretive layer on top.
What happens next
This is where the story actually gets interesting.
If PettiChat publishes its data and submits the device to independent testing, we will quickly find out which parts of the claim hold up. The company may be sitting on a real advance in animal-behaviour classification. They may also be selling an extremely well-engineered confidence trick.
If they do not publish, we will know something different — that the company is content to sell preorders and ride viral momentum rather than face external scrutiny. That is itself a data point, and not a flattering one.
For now, the situation sits in an unusual middle ground. The device is real. The funding is real. The preorders are real. The science is, at best, unconfirmed.
Ten thousand pet owners have already decided that’s good enough. Whether the actual technology is good enough is, for the moment, a question only the dogs and cats can answer — and they aren’t quite saying.
