A camera watched 870 cattle. It caught nearly every case of pinkeye before the vet did.

A fixed camera and an artificial-intelligence system watched 870 cattle at four locations and flagged 169 of 170 cases of pinkeye, formally bovine infectious keratoconjunctivitis, a contagious eye infection, before a veterinarian caught them, according to a peer-reviewed study from the U.S. Department of Agriculture's Agricultural Research Service. The trial's authors report 99.4 percent sensitivity and 97.6 percent specificity, meaning the system missed one case in 170 and rarely raised a false alarm. RFD News first reported the numbers this week.

The mechanism is a camera trained to read the muzzle and eye of each animal. Each cow's face acts as a biometric, a way to link the image back to a health record without scanning an ear tag, and a second pass classifies whether the eye shows early signs of IBK. For a producer, the difference between catching pinkeye on day one and catching it on day five is the difference between isolating one animal and treating a pen. Pinkeye spreads through face-to-face contact and flies, and the damage adds up: ulcerated corneas, weight loss, labor to handle sick cattle, and a sale barn that discounts animals with scarred eyes.

Related work at the ARS Livestock Issues Research Unit in Lubbock is testing a different approach. An infrared scan of the eye takes roughly 20 to 30 seconds per animal, compared with about two minutes for a rectal thermometer, and is paired with cameras that watch behavior for signs of fever or heat stress. The Lubbock group, led by animal scientist Nicole Sanchez, is building toward continuous screening for cattle in feedlots, dairies, and similar enclosed facilities. The IBK study sits alongside that effort, but on a different disease and a different input: visible light, not infrared.

The trial's authors frame their result as a research milestone, not a product launch. They name what still stands between a published paper and a working ranch tool: independent field validation across varied breeds and conditions, a shared definition of what counts as a confirmed case, and a price that pencils out against the cost of a treated animal.

A 99.4 percent sensitivity figure from a controlled trial does not guarantee the same hit rate under feedlot dust and summer glare, and the trial's definition of a "true positive" is the study's own, not a regulator's. A camera that flagged every watering-hole sneeze as pinkeye would still be right 99 percent of the time on a healthy herd and still be wrong about the alert that mattered.

The next step the researchers point to is a smartphone version: a producer holding a phone over a chute, the model running on-device, the result tied to the animal's record. That is the same architecture anyone who has used face unlock already understands. It is also, for now, a roadmap, not a rollout.