Sheep were trained to pick a familiar face out of a pair of photographs, and the result came with a published argument attached. · published 6 October 2026
Eight sheep learned to pick particular human faces out of pairs of photographs, managed it above chance when the faces were tilted to an unseen angle, and chose their own handler’s photograph in about 72 per cent of trials without being trained to - though a published response argued the sample was too small for the comparison with human ability.
These are the working notes, unedited. Everything the video
states as fact comes from here, and nothing else may be stated as fact.
The sections on limits and on what the finding does not show are
part of the brief, not an afterthought — see
the method.
The study. Franziska Knolle, Rita P. Goncalves and A. Jennifer Morton, “Sheep recognize familiar and unfamiliar human faces from two-dimensional images”, Royal Society Open Science, volume 4, issue 11, 2017.
What they did. Eight sheep were trained at a computer screen showing two images side by side. Choosing the correct one earned a food reward. The faces used in training were four celebrities, learned from photographic portraits - chosen simply because they were faces the sheep could have no prior contact with, not for any quality of fame.
The first result. Once trained, the sheep picked the learned face out of a pair well above chance. The researchers then tilted the portraits, showing the same person from a different angle the sheep had never seen. Performance fell, from about 79 per cent correct to about 67 per cent - but stayed well above chance. The size of that drop is similar to what happens when people are given the same kind of task, which is the detail that makes the result interesting rather than merely cute.
The second result. The sheep were then shown photographs of their own handler, someone they knew in person but had never seen as an image, against an unfamiliar face. They picked the handler in about 72 per cent of trials, with no training for that pairing at all.
The published objection. This did not go unchallenged. Towler and colleagues published a response in the same journal in 2019 arguing the conclusions were preliminary: too few animals, too few face stimuli, and no direct side-by-side comparison between sheep and human performance, which the claim of human-comparable ability leans on. A further objection is that the sheep were never tested on sheep faces - people are far better with human faces than with sheep faces, and a sheep would presumably be better with its own species, so a test using only human faces measures something narrower than “face recognition” in general. Keith Kendrick replied defending the work as consistent with thirty years of prior behavioural and neural evidence. The original authors did not concede an error.
Limits
Eight animals, four training faces, one handler. The tilted-image comparison with human performance is a comparison with results from other studies rather than humans tested alongside these sheep on this apparatus. Trained discrimination at a screen for a food reward is a narrow and well-defined skill, and calling it recognition imports more than the trials measured.
What this does not show
It is not evidence that a sheep knows who you are in the way you would mean it, nor that sheep think about faces the way people do. And the dispute above is not a scandal or a debunking: a short paper with eight animals drawing a comparison to human ability is exactly the sort of claim that should attract a written objection, and the objection appearing in the same journal is the process working, not failing.