Clever Hans: Right Answers for the Wrong Reasons

Ninety-Eight Percent, and Not One Sum

A horse answered 98% of number questions correctly without counting a thing. Catch him at it, then catch a machine doing the same.

You are standing in the crowd

The tapping in the courtyard

High apartment houses on all four sides, and a crowd pressed against the walls. In the middle stands Hans, a Russian Orlov trotter, beside a retired mathematics teacher in a wide-brimmed hat: Wilhelm von Osten, who has spent four years teaching him.

Somebody calls out a sum, and von Osten chalks it on a blackboard. Hans lifts his right forefoot, taps the flagstones (tap, tap, tap) and stops on the right number. He seems to tell the time, name coins, spell words and pick a tone out of a chord.

In September 1904 thirteen people convened by the psychologist Carl Stumpf, among them a zoo director, a veterinarian, a circus manager, military officers and schoolteachers, examined Hans. Their report of 12 September 1904 found no trickery. They were right, and the horse still could not add.

Everything that goes wrong here is still going wrong today, in hospitals, in cameras, in the models that sort your photographs. Only the horse has been replaced.

This lesson takes about 30 minutes. Every picture, model and chart on this page is computed live in your browser while you watch; nothing is a recording.

Act one · the bench

Ask Hans a question

Here is the courtyard and the horse. Pick a sum, ask it and watch him tap, then flip the two switches below one at a time and keep score. You are re-running the most famous experiment in animal psychology.

Predict first · Before you touch the switches: if the person asking does NOT know the answer, how often will Hans be right?

Trials run0
Hans was right0—
Condition—

The trace above is the height of the questioner's head during the last trial, sampled forty times a second. Ask a question and watch it.

Challenge. Pfungst wrote: “If I asked him to add 3 and 4, but did not move until 14 was reached, he would tap 14.” Make Hans answer 14 to 3 + 4.

What Pfungst actually measured

Oskar Pfungst, a student in Carl Stumpf's Berlin laboratory, ran the real tests. Between 13 October and 29 November 1904 he worked with the horse in the courtyard, under a canvas tent that kept the crowd away, and took the mystery apart with the same two switches you just flipped.

Pfungst's testConditionTrialsHans correct
Reading numbersQuestioner did NOT know the answer498%
Reading numbersQuestioner knew the answer4298%
AdditionQuestioner did NOT know the answer313 right
AdditionQuestioner knew the answer3129 right
Large blinders, 15 × 15 cmQuestioner NOT seen356%
Large blinders, 15 × 15 cmQuestioner seen5689%

Figures from Pfungst, Clever Hans (The Horse of Mr. von Osten), trans. C. L. Rahn, 1911, pp. 35–44. His summary (p. 40): with knowledge, “90 to 100% of the responses of the various series were correct”; without it, “10%, at most.” The blinders series had 102 tests; in 11 it was unclear whether Hans could see.

Notice which numbers belong to which switch: 89% and 6% are the blinders figures (could Hans see?), while 98% and 8% are the knowledge figures. Many retellings swap them.

Act two · swap places

Now you are the horse

Knowing the answer is not the same as feeling it, so stand where Hans stood. You won't see the sum, only a person, and you tap (button or space bar) until it feels time to stop.

Your taps0
The answer was—
Your record0 / 0

Press “Begin a trial”. Watch the person, not the numbers — there are no numbers.

If you got some right without the magnifier, you just did what Hans did, and it didn't feel like cheating. It felt like noticing. Hans was not a fraud: he was a superb reader of people, rewarded thousands of times for stopping at the right moment.

Pfungst found the cue. As the questioner finished, he bent his head and trunk slightly forward and Hans began to tap; at the right number he made “a slight upward jerk of the head” and Hans stopped. Pfungst made the same movements without knowing it, and people watching him “never perceived them, no matter how closely they observed me.”

He could not switch the cue off. Bending down to write a note was enough: “Hans would invariably begin to tap when I stooped to jot down some note I wished to make.”

Act three · the training floor

The same trick, in silicon

More than a century later, in 2019, a research team in Berlin pointed an explanation method at an image classifier and found the horse again. Its strongest evidence for “horse” in the PASCAL VOC 2007 photos was a small source tag printed in the lower-left corner of about one fifth of the horse pictures. Paste that tag onto a photo of a car and the model calls the car a horse.

So let's build one. Below, your browser will draw a few hundred tiny 32 × 32 pictures of horses and cars from scratch, stamp a little printer's tag into the lower-left corner of most of the horses, and train a real neural network on them — a two-layer network with hand-written backpropagation, about 16,000 weights, no libraries. Then it will test that network three ways.

Predict first · Predict: “in the wild” (the same animals, no tags anywhere) your network will score…

A sample of what the network is being shown. Horses on the top row, cars on the bottom. Look at the lower-left corner of each.

Accuracy on the pictures it studied—
Accuracy on a fresh test set from the same world—
Accuracy in the wild — same animals, no tags anywhere—
Cars wearing a tag that the model calls “horse”—

Press “Train the network”. It takes a second or two.

Challenge: get the wild score to 75% or more while at least 90% of the training horses still carry the tag.

The Ferrari moment

Here is one car. The model is fairly sure it is a car. Now stamp the tag on it — the tag has nothing to do with cars or horses, it is a printer's mark — and watch the verdict.

The model's verdict—

Train the network first, then press the button.

Act four · where is it looking?

Cover one square at a time

Pfungst could not ask Hans what he was looking at, so he used blinders. We can do something cruder: cover one small square of the picture, rerun the model and see how much the verdict moves. Do it for every square and you get an occlusion map, a map of where the evidence is.

The picture

Where the evidence is

The two, laid on top of each other

Train the network in the act above first.

The measured curve

How reliable does a shortcut have to be?

How reliable must the tag be before the model falls for it? Your browser trains a fresh network at each tag rate from 0% to 100% and measures three numbers. The green line is the score papers report; the orange line is the one that matters outside the lab.

Game · think like the machine

Horse or Car? Read the Model

Hans won by reading his questioner. Now you read the network you trained. Each round shows a picture: tap what the model will say, not what the picture is. Ten rounds, eight seconds each.

Round 0 / 10 Score 0 Streak 0 Best streak 0

What will your model say?

Act five · the ward

The score goes up. The medicine does not.

In 2018 a team led by John Zech tested pneumonia detectors across three hospital systems. Pneumonia showed up in 34.2% of the chest X-rays from Mount Sinai, 1.2% from the National Institutes of Health (NIH) and 1.0% from Indiana.

The hospitals leave fingerprints: a metal marker that technicians place in the corner of the image, and portable bedside scanners that make visibly different pictures. A network trained to name the hospital got 99.95% of NIH films right, 99.98% of Mount Sinai's and 95.59% of Indiana's. Inside Mount Sinai it told inpatient films from emergency ones with 100% accuracy.

Watch what that does to a score. Below is a simulated hospital system with the real prevalences, where each patient has a genuine lung signal and a token that says which hospital they are in. Train one model with the token visible and one with it hidden, and compare.

Weight on the lungs—
Weight on which hospital—
The hospital matters this much more—

This is a simulation built to reproduce the mechanism, not the study's data. The real published figures: a model trained on NIH scored 0.750 on NIH's own held-out films and 0.695 at Mount Sinai; a model trained on both together scored 0.931 on the pooled test set — and that 0.931 is the confound talking, not the lungs.

A single leaderboard number cannot tell you whether your model is reading the lungs or reading the hospital. Nothing about a good score rules out a shortcut — a shortcut is defined as the thing that gets a good score.

The receipts

Six times a machine was Clever Hans

Every one of these is documented, published and numbered. Not one of them involved a broken model. Each model was working perfectly — on the problem it had actually been given, which was not the problem anyone meant to give it.

Act six · a story about a story

The tank that never was

Every machine-learning course tells the same story. The army trained a network to spot tanks in a forest; it was perfect in the lab and useless in the field, because the tank photos were taken on a cloudy day and the empty-forest photos on a sunny one. The network had learned the weather.

It is a wonderful story. It is almost certainly not true. Click along the chain below and see how far back it goes.

The trail ends at an email. In 2013 Edward Fredkin wrote that, at a talk in Los Angeles in the early 1960s, he had asked whether the network had “merely trained itself to recognize the difference between a bright picture and a dim picture.” Nobody tested it, and over fifty years the question hardened into a “fact” with dates and photo counts.

A real anecdote from that era is better. Marvin Minsky recalled a friend in Italy whose perceptron learned to tell Bach chorales from chorales by conservatory students. Minsky looked at the scores: one set had a single whole note in the lower corner of each page, and that alone could split them.

A legend survived half a century without a shred of evidence because everyone who heard it recognised something true. The phenomenon never needed the tank. It has hospitals, dermatology clinics, camera traps and a horse.

Where this page is thinner than the truth

Six honest wobbles

One hundred and twenty years

From a courtyard to a chest X-ray

Click any stop.

The examination

Eight questions, and no cues from the examiner

Wrong answers are allowed and are answered honestly. There is no head-jerk hidden anywhere on this page.

Where the facts come from

Sources

Every figure on this page was checked against these pages in October 2026. Titles are given in their original language.

  1. Pfungst, O. (1911). Clever Hans (The Horse of Mr. von Osten), trans. C. L. Rahn. Henry Holt. Project Gutenberg edition · www.gutenberg.org
  2. “Clever Hans.” Wikipedia · en.wikipedia.org
  3. „Kluger Hans.“ Wikipedia (Deutsch) · de.wikipedia.org
  4. “Muhamed (horse)” — Krall’s Elberfeld horses. Wikipedia · en.wikipedia.org
  5. Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W. & Müller, K.-R. (2019). Unmasking Clever Hans predictors and assessing what machines really learn. Nature Communications 10:1096 · arxiv.org
  6. Zech, J. R. et al. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs. PLOS Medicine 15(11):e1002683 · journals.plos.org
  7. Ribeiro, M. T., Singh, S. & Guestrin, C. (2016). “Why should I trust you?” Explaining the predictions of any classifier. KDD 2016 (§6.4, Table 2) · arxiv.org
  8. Beery, S., Van Horn, G. & Perona, P. (2018). Recognition in Terra Incognita. ECCV 2018 · arxiv.org
  9. Winkler, J. K. et al. (2019). Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA Dermatology 155(10) · jamanetwork.com
  10. Geirhos, R. et al. (2019). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. ICLR 2019 · arxiv.org
  11. Geirhos, R. et al. (2020). Shortcut learning in deep neural networks. Nature Machine Intelligence 2:665–673 (§6.1, “Morgan’s Canon for machine learning”) · arxiv.org
  12. DeGrave, A. J., Janizek, J. D. & Lee, S.-I. (2021). AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence 3:610–619 (preprint) · www.medrxiv.org
  13. Mordvintsev, A., Olah, C. & Tyka, M. (18 June 2015). Inceptionism: going deeper into neural networks. Google Research Blog · research.google
  14. Branwen, G. “The Neural Net Tank Urban Legend” (Dreyfus 1992, Pereira 1994, Fraser 1998, Fredkin 2013, Kanal & Randall 1964, Minsky 2011) · gwern.net