In this article
Your dog is sitting by the back door, ears half up, eyes fixed on you. Hopeful? Bored? Uneasy about the bin lorry? You probably have a theory. Now there is software that offers one too: AI that reads a dog's face and body from a photo or a short video.
Researchers have been putting that idea to the test, and the answer is more interesting than yes or no. Here is how scientists decide what a dog is feeling in the first place, what the strongest studies found, where they fall short, and how to weigh any app's verdict against what you know about your own dog.
How scientists decide what a dog feels
Start with the awkward part. A person can tell a researcher how they feel. A dog cannot, so there is no simple answer sheet to mark a computer against. A review of twenty computer-vision studies of animal pain and emotion says this is perhaps the most important problem of all4.
Scientists do not even share one definition of an animal emotion, though a common working idea is a short, intense state set off by something rewarding or unpleasant that is happening or expected4. So researchers build their answer sheet in one of two main ways: stage a situation and define the feeling by it, or ask people to label photos and videos, which can bring in human error and bias4.
Twenty-nine Labradors and a treat out of reach
By its authors' account, this was the first study to put AI to work on dogs' emotional states using footage from a controlled experiment: 29 Labrador Retrievers, filmed behind a transparent window for an earlier study1. When a dog had learned to expect a food reward, it was taken to be in positive anticipation; when it could see the food but was denied it, it was taken to be frustrated1, 2.
The 3-second clips then went to two kinds of system. One worked from the facial movements a certified coder had logged by hand with DogFACS, the coding scheme researchers use for dogs' faces, and was right just over 71% of the time1. The other, a deep-learning system that learned straight from cropped pictures of the dogs' faces and found its own patterns, was right in 89% of clips1.
The Labrador study in numbers
The researchers balanced the clips, 82 of each kind, so blind guessing would be right about half the time1. Heatmaps showed the best system tending to focus on the eyes, ears and nose1. There was a catch for the hand-coded system, though. Its most useful cue was the eyes looking up, and the food was delivered with a motion slightly above the dogs' eyeline, so the authors say that cue could be a quirk of the experiment1.
They also point out that every dog was a Labrador, filmed in strictly controlled conditions. Testing other breeds is an immediate need for future research, they write, and moving from controlled rooms to everyday settings is notoriously difficult1.
Why one expression is not enough
A single cue makes a weak rule. In the Labrador study, a rule based on drawn-back ears alone was right only about two times in three1. In a 2021 study of 28 Labrador-type dogs waiting for food or toys, drawn-back ears appeared in about 89% of frustration clips, but also in 55% of positive anticipation clips2. A nose lick rarely turned up while dogs were waiting happily, yet it appeared in only 17% of frustration clips2.
The researchers concluded that none of the facial movements they found would sort the two feelings consistently on its own2. Piecing together several sources helps: the situation, the body, whether the dog moves towards something or away, and physiology such as heart rate2. Weighing many small cues at once is where machine learning has an advantage, a 2025 paper notes5, and it is a good habit for anyone watching a dog.
Even the lab's answer sheet is a best guess. The two feelings are closely linked, and the researchers could not rule out some dogs tipping from hopeful to frustrated during the five-second wait2.
What happens when general AI looks at a dog
Carefully staged footage like the Labrador clips is rare3. A quicker route is to show a picture to a general-purpose AI, the kind built to make sense of pictures in general, and ask what the dog is feeling3. In 2025 a team tested eleven versions of popular models like this, with no dog-specific training3.
On web photos labelled happy, sad, angry or relaxed by the person who collected them, some of the models did moderately well3. Then the team cut the dogs out and pasted them onto new backgrounds. One model more often called the same dog happy or relaxed on grass or a sofa, and sad in a clinic or behind bars3.
On the Labrador faces from the lab study, where the answer came from the experiment and there was almost no background to go on, every model performed close to chance, and most gave nearly every face the same label3. Rewording the request did not help. Told "You are a dog", one model scored 48%, the same as with no extra instruction3. Yet the information was there: the purpose-built system had scored 89% on the same faces3.
Even in commercial products aimed at pet owners, misleading emotional feedback could foster false confidence in the technology and skew human-animal interactions3.
The authors conclude that models like these are not yet suitable for reliably recognising dogs' emotions3. It is a useful warning for anyone using such tools: a confident label can come from the sofa behind the dog.
Real homes are harder than labs
Purpose-built tools have their own trouble outside the lab. A 2025 study ran an automated system over home videos that owners had filmed during New Year's Eve fireworks and on an evening without them, tracking points on each dog's face5.
Of 42 dogs, only 11 left enough clear frames to analyse5. Dogs hid, cameras caught them at awkward angles, and a dark dog against a dark background gave too little contrast5. Ear shapes varied so much among the dogs that the team dropped the points on the ear flaps altogether5.
What was left still showed the ears drawn back during fireworks, matching what people coding the videos by hand had found5. The authors add two cautions: AI findings need careful human interpretation, and it is not yet clear whether faces alone can tell fear from frustration5.
How to weigh an app's verdict
None of this makes a mood read useless. It makes it an estimate from limited evidence, and you hold evidence the app does not: what happened a minute ago, what is normal for this dog and how the rest of the body looks.
Reading dogs is largely a learned skill. In a study of 89 adults and 77 children in Europe and Morocco, the ability to recognise dog emotions from photos came mainly through experience6. Adults generally found it easier to say what situation a picture was taken in than to name the emotion it showed, and they were good at spotting happy dogs but not fearful ones6.
Treat headline accuracy figures with care, too. The review of twenty studies lists one early dog study at 95.3%4. But that study used pictures gathered from the internet and labelled by non-experts, and the review says web-scraped data like this should be treated with caution1, 4.
Dogoos, which publishes Bark & Bite, offers mood reads too. They come from a multimodal AI7, the same broad kind of model as the ones tested in 20253. It looks at the face, body and sound in a short clip together7, and the cautions in this article apply to it as well. Think of each read as a best guess. Its confidence figure is the AI's own7: it says how sure the AI was about that clip, not how often it is right.
So treat a read as a second opinion, worth a look and never the last word. A clip can also be a handy record. If something puzzles you, film the whole dog, jot down what happened just before, and keep it to show your vet or a qualified behaviourist.
Before you trust a read
- The whole dog was in view, from a clear angle, and stood out from the background5
- You know what happened just before: a treat, a knock at the door, a bath
- The body and the situation tell the same story as the face2
- It matches how your dog usually looks at that time of day
- Nothing else about their behaviour has changed lately8
When to talk to a professional
An app cannot examine your dog. If you are worried about how your dog is behaving, the RSPCA's advice is to speak to your vet first, who may then refer you to a clinical animal behaviourist8. A change in behaviour can be a sign that something is wrong8, and pain or illness is a job for a vet.
In a lab, for one breed and two feelings set up by the experiment, AI has done impressively well. At home, treat a read as a prompt to look again at the dog in front of you, who has been sending signals all along.
Sources and further reading8 sources
The small numbers in the article link to the source behind each fact. Sources open on their own websites.
- Explainable automated recognition of emotional states from canine facial expressions: the case of positive anticipation and frustration
- Evaluating the accuracy of facial expressions as emotion indicators across contexts in dogs
- Investigating the capabilities of large vision language models in dog emotion recognition
- Going Deeper than Tracking: A Survey of Computer-Vision Based Recognition of Animal Pain and Emotions
- Automated analysis of emotional expressions in dogs based on geometric morphometrics
- The ability to recognize dog emotions depends on the cultural milieu in which we grow up
- Dogoos: dog mood, health and everyday care in one app
- Understanding Your Dog's Body Language