When AI Sees the Suffering Humans Ignore

A machine can now look at a bull’s face and tell when he is in pain. Better than humans, in some cases. That is the headline.

But it is not the whole story.

A study published in Scientific Reports compared trained veterinarians with AI models in detecting pain in bulls. The researchers used video footage from 17 bulls who had been subjected to castration surgery. The footage came from two key moments: before sedation, used as the “no pain” condition, and three hours after surgery, before post-operative pain relief, used as the “pain” condition. The humans used recognised pain scales. The machine used video frames. The results were striking.

The AI model achieved 97.06% accuracy and an F1 score of 96.97%. It had perfect precision and specificity in the study, meaning it did not wrongly classify any bulls as being in pain when they were not. It still missed a small number of genuine pain cases, but overall, the system performed as well as trained veterinarians in real-time assessment and better than human video assessment.

So now we have a machine that can detect pain in bulls.

Now what?

Because this is where the story becomes uncomfortable.

The study is being framed as a breakthrough for pain assessment. In one sense, it is. Pain in cows and bulls is difficult for humans to read. They are prey animals. They often hide signs of pain. Their suffering is subtle, because survival has taught them to make it subtle. Humans then take that subtlety and use it against them. If an animal does not scream in a way we recognise, we assume the pain is not serious. If he does not collapse, we assume he is coping. If his face does not look enough like ours, we pretend there is nothing to see. Then along comes AI and says, actually, there is something to see. There was always something to see.

Humans just missed it.

Or ignored it.

The researchers point out that pain in bovine species is frequently underestimated. Anyone who has looked honestly at farming should not be surprised. Pain is not an accident in animal agriculture. Pain is built into the system. Castration. Dehorning. Separation. Confinement. Lameness. Mastitis. Transport. Slaughter.

These are not rare exceptions. They are routine. They are normalised. They are priced in.

The industry does not need AI because pain is hard to imagine. It needs AI because the pain it causes is so widespread that it has become operationally inconvenient to notice. That is the brutal truth underneath the technology. A machine may detect the pain. But the farm still creates it.

The bulls in this study were not being monitored in some abstract ethical vacuum. Their pain followed castration surgery. Their bodies were altered because human systems treat animal bodies as units to be managed. Then the question becomes: what does pain detection actually mean inside a system that has no intention of stopping the violence?

Does the machine detect pain so the animal receives relief?

Or does the machine detect pain so the system can keep using him more efficiently?

Does it protect the bull?

Or protect the business model?

This is the question advocates need to ask every time AI is introduced into farming.

Not “can it work?”

Work for whom?

AI in animal agriculture is usually described as precision farming. The language is clean, technical, and deliberately dull. Cameras, sensors, algorithms, monitoring systems. It sounds like progress. It sounds modern. It sounds almost caring. But precision does not mean compassion. A system can become more precise without becoming less violent. You can precisely measure suffering while still causing it. You can detect pain earlier and still send the animal to slaughter. You can reduce false positives, optimise interventions, and improve productivity, all while leaving the basic injustice untouched. That is the danger.

Technology can make exploitation look smarter. Cleaner. More responsible. More difficult to criticise.

The machine sees the pain.

The industry sees data.

This study does include important caveats. The dataset was small, with only 17 bulls and 34 videos. The conditions were controlled. The model was tested on two time points, not a wide range of pain states. The classification was binary: pain or no pain. Real pain is not binary. It can be mild, moderate, severe, chronic, intermittent, hidden, confused with fear, or shaped by stress. The bulls had also been fasted before surgery. The researchers argue this did not significantly affect the baseline, based on previous validation work, but it remains part of the context. The veterinarians were unblinded during live assessment, meaning they knew which stage the bulls were in. That may have influenced their scoring. The AI may also have picked up on behavioural patterns linked to surgery, stress, or residual effects, rather than pain alone. So no, this does not mean a machine has fully solved animal pain. It means a machine may be able to detect some forms of pain, in some animals, in some conditions, with impressive accuracy. That is still important. But it should not become a moral escape hatch. The worst response would be to turn this into another argument for “better farming.”

Better monitoring.

Better pain scoring.

Better intervention.

Better management.

Better exploitation.

The problem is not that humans have been using animals without enough data. The problem is that humans have been using animals.

More information can help reduce suffering in the short term. Nobody serious should oppose pain relief. If AI can help identify animals in pain and get them treatment faster, that is obviously better than leaving them untreated. But we need to be honest about the limits.

Pain relief is not liberation.

Detection is not justice.

Monitoring is not consent.

A bull who has his pain recognised is still not being asked whether his body should be used.

A cow on a dairy farm whose lameness is detected earlier is still trapped in a system built around taking her milk, taking her calves, and taking her life when she is no longer profitable.

A machine that sees suffering does not change the fact that the suffering is human-made.

This is where advocacy has to be careful. There is a version of this story that the industry will love. It will say AI proves farming is becoming more responsible. It will say animals are being watched more carefully than ever before. It will say technology is helping farmers care. But care is a strange word for a system that breeds someone into existence for profit, controls their body, and kills them when convenient. The bar has been buried so low that noticing pain is treated as progress. Imagine doing that to humans. Imagine building a system around harming people, then celebrating a camera for detecting their suffering more accurately. No. The moral problem would not be poor detection. The moral problem would be the harm.

The same applies here.

Still, this technology matters because it exposes a lie. For years, animal industries have relied on distance. Physical distance. Emotional distance. Linguistic distance. They turn individuals into categories. Bulls become stock. Cows become production animals. Pain becomes a management issue.

AI, ironically, may make some of that harder to hide. If a machine can detect pain in a bull’s face, then the old excuses become weaker.

We did not know.

They do not feel it like we do.

They are fine.

They are used to it.

They do not show pain.

No.

They show pain. Humans just built entire industries around not seeing it. So what now?

Advocates should learn this technology, understand its limits, and challenge its use. Ask what happens after pain is detected. Ask whether treatment follows. Ask who controls the data. Ask whether the goal is reducing suffering or protecting profit. Ask whether AI is being used to help animals, or to make larger systems of confinement easier to manage with fewer humans present. Most of all, refuse to let technology soften the ethics.

The question is not whether AI can detect pain better than humans. The question is why humans keep creating so much pain for AI to detect. That is the part no algorithm can solve.

The machine may see the bull’s suffering. But only humans can decide to stop causing it.