By Elizabeth Upchurch
Artificial intelligence is now outperforming humans at reading animal faces for signs of pain and stress. In a series of studies highlighted in Science, researchers have trained AI systems to analyze minute facial movements in animals—including horses—using established grimace scales and deep-learning models. In multiple trials, these systems identified pain and emotional distress with greater accuracy than experienced veterinarians and behavior experts. They sometimes detected discomfort humans missed, and other times confirmed that animals were pain-free when people believed otherwise.
These AI systems work by tracking subtle changes in the eyes, ears, and mouth. They look at micro-movements that often escape human notice. Unlike people, who rely heavily on context, habit, and expectation, the algorithms evaluate facial patterns consistently across thousands of images and videos. The findings suggest that equine pain and stress can be far more nuanced than traditional observation allows.
While most riders aren’t going to run out and get facial-recognition AI monitoring for their horses any time soon, the research challenges the assumption that we know when our horses are uncomfortable. What this research makes clear is that many signs of pain and stress are easier to miss than we realize, and that greater awareness alone can meaningfully change how we observe, manage, and advocate for our horses.
Why Horses Hide Discomfort So Well
Horses evolved as prey animals, and that evolutionary history still shapes how they express pain. Overt signs of weakness would have made a horse vulnerable in the wild, so discomfort often shows up subtly—if it shows up at all.
Scientists have developed species-specific “grimace scales” that focus on changes in facial muscles around the eyes, ears, and mouth. For horses, signs of discomfort can include ears rotating outward, tension that creates so-called “worry wrinkles” above the eyes, and tightening around the muzzle. These movements are often barely perceptible unless someone is specifically trained to look for them
Even then, accurately reading those expressions is challenging. Human experts must manually identify and code each facial muscle movement, a process that can take “an average of 100 seconds to identify the various facial muscles and code their positions in a single image,” according to the article. In real-world settings, those fleeting facial cues are easy to miss, which helps explain why discomfort is so often overlooked or misinterpreted.
For owners, this explains why discomfort is so often reframed as attitude or training resistance. A horse who feels tight, dull, or inconsistent may not look obviously lame but that doesn’t mean they’re fully comfortable.
What AI Is Seeing That Humans Often Miss
In these studies, AI detects pain by identifying details humans don’t consciously register. Machines can measure microscopic changes in facial landmarks like tiny shifts in distance between the eyes, the angle of the ears, or the tension in the muzzle, that are nearly impossible to track by eye.
In controlled studies, AI systems correctly identified pain in horses at rates that exceeded those of veterinarians and behavior specialists reviewing the same images and videos. In some cases, the technology also ruled out pain when experts believed it was present, suggesting that human perception can be influenced by expectation as much as observation.
How You Can Use This Research
The first learning is to focus on patterns, not isolated moments. A single pinned ear or tense expression may not mean much on its own, but repeated facial tension in similar situations can point to discomfort.
The second is to pay attention to faces at rest. Several studies noted that resting facial expressions can be especially revealing, sometimes more so than expressions under saddle. A horse standing quietly in a stall or on the cross-ties may show tension that disappears once movement masks it.
Finally, it’s worth remembering that “normal” is not the same as “comfortable.” Horses often adapt to chronic low-grade discomfort, and owners adapt alongside them, gradually accepting subtle tension as part of the horse’s baseline.
At its core, this research reinforces a principle most good horsepeople already believe: listening matters. But it also suggests that listening requires more humility than we sometimes allow ourselves. While researchers speculate that AI could one day support welfare standards in competition settings, today it reminds us that horses communicate constantly—often quietly. Improving welfare starts with learning to see what we’ve been trained to overlook.
Technologies like AI show us how much more there is to learn. This research invites horse owners to slow down, observe more closely, and take small changes seriously.
