The horse that remembers your face was angry, and the statistical error in the study before it. · published 24 September 2026
Horses that had seen an angry photograph of a specific person later looked at that same person — now neutral-faced and unaware of which photo the horse saw — more with the left eye, the one that feeds the threat-processing hemisphere, and the effect didn’t transfer to anyone else; the blinding is what makes it mean anything, and the earlier, flashier study on the same question was picked apart for proving an effect in one condition, failing to prove it in another, and calling that a difference.
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 first claim, 2016. Amy Smith, Leanne Proops, Karen McComb and colleagues, “Functionally relevant responses to human facial expressions of emotion in the domestic horse”, Biology Letters. Twenty-eight horses were shown photographs of human faces, angry or happy, with no training beforehand. The reported findings were a left-gaze bias to the angry photographs - animals tend to look at threatening things with the left eye, because the left eye feeds the right hemisphere - and a faster rise in heart rate to angry faces than to happy ones (median 14.7 seconds against 29). The experimenters could not see which photograph they were holding up, which matters on this subject more than most.
The challenge, the same year. Peter Schmoll published a re-analysis in the same journal: “Can horses read emotional cues from human faces?” His objection to the gaze result is a genuine and common statistical mistake. The authors had shown that the laterality index differed significantly from zero for angry faces and did not for happy ones - and treated that as showing the two differed from each other. It does not. A significant result in one group and a non-significant one in another is not a test of the difference between the groups; that comparison has to be made directly. He also questioned heart-rate latencies under a second, which are too fast to be a response to what was on the card. The authors published a reply defending the functional relevance of the findings. The exchange is on the record and was not settled by either side.
The stronger result, 2018. Proops, McComb and colleagues, “Animals remember previous facial expressions that specific humans have exhibited”, Current Biology. A horse is shown a photograph of a person’s face, angry or happy. Some hours later, that same person walks in - in the flesh, with a neutral expression, and not knowing which photograph the horse had been shown. Horses that had seen the angry photograph looked at the person more with the left eye: the threat-processing side. The effect was specific to that individual. A different person did not get it.
Why the blinding is the point. The person meeting the horse did not know which photograph it had seen, so they could not behave differently and be read. That is the Clever Hans control, built in from the start - a horse in Berlin appeared to do arithmetic in 1904 and was in fact reading the questioner’s involuntary posture, and the fix has been the same ever since. Here it is what makes the finding mean anything.
Limits
Twenty-eight horses in 2016, and a single laboratory group behind both papers. The 2018 measure is gaze direction, which is an inference about what the horse is attending to, not a readout of what it feels. “Perceived more negatively” is the authors’ interpretation of a looking bias, and a looking bias is a small behaviour. Whether the 2016 result survives Schmoll’s objection is still an open argument, and the 2018 study is not a replication of it - different design, different measure.
What this does not show
Not that horses read minds, feel human emotions, or know you are in a bad mood. Not that the 2016 authors were sloppy - the statistical error Schmoll names is one of the most common in published science and gets past reviewers constantly. And not that the dispute means the whole thing collapsed: the 2018 design is the stronger one and it stands on its own.