Animal Welfare Advocates Are Betting AI Can Decode What Animals Actually Feel

The shoes come off at the door at Mox, a coworking space in San Francisco's Mission District that looks more like a Moroccan riad than a tech incubator. Persian rugs, mosaic lamps, potted palms. In early February, amid that deliberately unhurried atmosphere, a quietly consequential meeting took place: animal welfare advocates and AI researchers sat together on cushions and couches, trying to figure out whether machine learning could do something humans have never managed to do reliably — understand what animals are experiencing from the inside.

The gathering was part of a broader, accelerating effort within the Bay Area's animal welfare community to recruit artificial intelligence as a tool for one of the oldest and most philosophically thorny problems in biology: measuring animal consciousness and suffering. Wildlife advocates, farm animal researchers, and AI engineers are increasingly convinced that the same pattern-recognition capabilities now reshaping medicine, climate science, and drug discovery could be turned toward decoding the behavioral and physiological signals that animals use to communicate distress, fear, or wellbeing.

The appeal is obvious. Animal welfare science has long struggled with a fundamental asymmetry: the beings whose suffering researchers are trying to measure cannot describe their experience in language. Veterinary pain scales exist, but they are largely observational, dependent on trained human judgment, and difficult to standardize across species. A pig in a factory farm and a chimpanzee in a sanctuary may both be suffering, but the behavioral signatures look nothing alike, and the humans watching them bring their own cognitive biases to every assessment.

AI systems trained on large datasets of animal vocalizations, movement patterns, and physiological signals could, in theory, identify distress markers that human observers consistently miss. Research groups have already demonstrated that machine learning models can distinguish between the calls of animals in pain versus those at rest, and that computer vision systems can detect subtle postural changes in livestock that precede illness by hours or days. A 2022 study published in Scientific Reports showed that deep learning could classify pig vocalizations associated with positive and negative emotional states with meaningful accuracy, suggesting that the emotional lives of farm animals may be more legible to algorithms than to the farmers who raise them.

The implications for industrial agriculture alone are staggering. Roughly 70 billion land animals are raised for food each year globally, the vast majority in conditions that welfare scientists consider chronically stressful. If AI monitoring systems could be deployed at scale inside those facilities, the feedback loop between animal experience and human management decisions could tighten dramatically. Farmers would no longer need to wait for visible signs of illness or distress; the system would flag problems earlier, potentially reducing both suffering and economic loss from sick animals. That alignment of welfare and profit motive is precisely the kind of incentive structure that tends to drive adoption.

But the second-order consequences deserve careful attention. If AI welfare monitoring becomes standard in agriculture, the most immediate beneficiaries may not be the animals themselves but the corporations that can use the technology to preempt regulatory scrutiny. A farm that can point to a continuous AI welfare audit may face less pressure from inspectors, advocacy groups, and consumers, even if the underlying conditions remain far below what welfare scientists would consider acceptable. The technology could, paradoxically, provide a layer of legitimacy to systems that still cause enormous suffering, simply by making that suffering more legible and therefore more manageable within existing frameworks rather than prompting a fundamental rethink.

There is also the question of what AI can actually measure versus what it can infer. Behavioral and physiological signals are proxies for subjective experience, not direct windows into it. The hard problem of consciousness does not dissolve because a neural network has processed ten million hours of chicken vocalizations. Researchers at institutions like the Cambridge Declaration on Consciousness have long argued that the scientific evidence for sentience in non-human animals is robust, but translating that philosophical position into actionable welfare standards remains contested terrain.

What the San Francisco gathering reflects, more than anything, is a growing impatience within the animal welfare movement with the pace of change through traditional channels. Legislation moves slowly. Consumer behavior shifts incrementally. But AI is moving fast, and advocates are increasingly determined to be in the room when its applications are being designed, rather than arriving afterward to argue about the consequences.

The more interesting question may not be whether AI can measure animal suffering, but whether the humans who deploy it will be willing to act on what it finds.