Clever Hans: The Counting Horse That Fooled Berlin, and Still Fools AI

Around noon on most days in the summer of 1904, a crowd gathered in a paved courtyard hemmed in by tall apartment blocks in the north of Berlin. In the middle stood an elderly former mathematics teacher in a black slouch hat and, beside him, a Russian trotting horse. The teacher would ask a question, sometimes out loud and sometimes chalked on a board, and the horse would answer by tapping its right forefoot on the cobbles. Add two-fifths and one-half? Nine taps, then ten: nine-tenths. If the eighth of the month falls on a Tuesday, what date is the following Friday? Eleven taps. Nobody paid to watch. The owner never took a fee.

The horse was called Hans, and Berlin soon called him der Kluge Hans: Clever Hans. Within months he was on postcards and liquor labels, he had toys made in his image, and he had been studied by a commission of thirteen men that included a circus manager, a zoo director and one of the most respected psychologists in Germany. The commission could not find a trick. It took a young researcher named Oskar Pfungst, a canvas tent and several weeks of patient testing to discover what was really going on, and the answer turned out to be stranger and more useful than fraud.

Hans could not add. What he could do was read the people around him, down to movements smaller than a millimetre that the people themselves did not know they were making. More than a century later his name is still attached to that problem, in animal labs, in dog training and in artificial intelligence.

The schoolmaster and his pupil

Wilhelm von Osten was in his sixties, a retired teacher of mathematics at a gymnasium, the German academic secondary school. He believed that animals had been underrated, and that a horse given the same patient instruction as a schoolchild would show what it could really do. So he taught Hans the way he had taught boys: with a counting frame, with numbers and letters written on a board, with lessons repeated day after day, and with rewards of bread and carrots rather than the whip.

Hans tapped out numbers with his forefoot, nodded for yes and shook his head for no, and pointed with his nose at cards laid on the ground. He spelled words using a table in which every letter had a row number and a column number, so a word became a string of taps.

Carl Stumpf, the psychologist who would later lead the inquiry, set out the range of claims in his introduction to Pfungst’s book. Hans, his admirers said, could work with fractions and find the factors of a number. He could read, spell, tell the time and keep track of the calendar. He could recognise coins and tell musical tones apart, and was said to know thirteen melodies. He ignored questions put in French or Latin, which his friends took as proof that he was not simply guessing. Experienced educators, Stumpf wrote, placed him at about the level of a child of 13 or 14.

Stumpf called von Osten “fanatic in his conviction” that the horse could think, and that conviction is the hinge of the story. Von Osten was not running a con. He was right that Hans was answering. He was only wrong about how.

“Nothing but language lacking”

Berlin’s newspapers argued about Hans for months. One camp thought he could genuinely reason. The other was sure it was a trick and set about naming the signal: von Osten’s slouch hat, his long coat, his hand moving toward the carrot pocket, his eyes. More exotic theories had the horse picking up body heat, electric currents through the ground, or N-rays, a supposed radiation that briefly had a scientific following before it was shown not to exist.

What made Hans hard to dismiss was that he did not need his owner. Other people could ask him questions and get correct answers, sometimes when von Osten was not even in the courtyard. The explorer and naturalist Carl Georg Schillings, who first came to the yard in July 1904, became convinced and began testing Hans himself, with similar success. If it was a trick, it was one that worked for strangers.

The story reached the United States too. On 4 September 1904 The New York Times ran a long dispatch from Berlin under the headline “Berlin’s Wonderful Horse; He Can Do Almost Everything but Talk.” Stumpf, who was already following the case, had suggested in the Berlin press that an expert panel take a look.

The September commission

On 12 September 1904 the panel’s thirteen members signed a short report. They included the circus manager Paul Busch, the Berlin zoo’s director Ludwig Heck and his assistant Oskar Heinroth, a veterinary assistant, a retired major-general, a teacher and Stumpf himself. One series of tests was run with only Busch present, so no one in the audience could be signalling.

The commission’s conclusion is often summarised as “no trickery found,” which is true but incomplete. The members agreed unanimously that there were no intentional signs from the questioner, and they added that “unintentional signs of the kind which are at present familiar” were ruled out as well. They did not say Hans could think. They said this appeared to be a case “which appears in principle to differ from any hitherto discovered” and was “worthy of a serious and incisive investigation.” They had cleared Hans of the obvious explanations, not explained him.

That phrase, “of the kind which are at present familiar,” left room for an unintentional sign nobody yet knew about. Stumpf took that seriously and gave the job of finding it to a student at his Psychological Institute at the University of Berlin.

Pfungst’s tent

Oskar Pfungst began work on 13 October 1904 and finished on 29 November. With Stumpf, Erich von Hornbostel, who kept the records, and Friedrich Schumann, he set up a large canvas tent in von Osten’s courtyard so the horse could be tested without a crowd, and he went back several hours at a time, about four times a week.

His key idea sounds obvious now. Pfungst separated two kinds of tests: those in which the questioner knew the answer, and those in which nobody present did. For number cards, he would show Hans a card that he himself had not looked at. For sums, two people would each whisper a number into the horse’s ear, so that neither knew the total. He also used blinders so that Hans could not see the questioner at all.

The results were lopsided. In the number-card tests, Hans got 98 percent correct with a knowing questioner and 8 percent without. With printed words he went from perfect to zero. In the whispered addition test he got 29 of 31 right when the questioner knew the total, and 3 of 31 when no one did, which Pfungst put down to chance.

Even counting fell apart. Pfungst set up the counting frame von Osten had used in training, pushed some of its balls to one side and turned his back on it. In eight such tests Hans failed every time nobody watching knew the number, and succeeded every time someone did. Across the board, Pfungst found 90 to 100 percent success when the questioner knew the answer and at most 10 percent when he did not.

That was not a verdict against von Osten. It meant the answer was somewhere in the person asking, whoever that was.

Blinders, whispers and a bite

The blinder tests narrowed it down further. Pfungst fitted Hans with large blinders and stood where the horse would have to turn its head to see him. Hans made, in Pfungst’s words, “the most strenuous efforts to get a view of the questioner,” and each test was logged by whether he had managed it.

In tests where Hans could see the questioner, he answered correctly 89 percent of the time. In tests where he could not, his success fell to 6 percent. He was watching for the answer.

The sessions were not gentle. Hans fought being tied, and Stumpf noted that when the horse was loose and excited he could make the courtyard “an unsafe place”; both Schillings and Pfungst “suffered from more than one bite.”

The millimetre that moved a horse

So Pfungst started watching the questioners. As soon as someone asked a question, they leaned forward a little to watch the horse’s foot. That small forward bend was the signal to start tapping. When the count reached the correct number, the questioner relaxed, and the relaxing came with a slight upward movement of the head. That was the signal to stop.

The movements were tiny. Pfungst found that von Osten’s were among the smallest he ever saw, although the wide brim of his slouch hat made them easier to see, magnifying the motion in a ratio of about 3:2. Bare-headed questioners still gave the signs, and Pfungst found he could see them from a metre and a half away. Even with the head covered, he could follow them in the eyebrows.

To prove it, Pfungst used the cues on purpose. Simply stooping, as if to write a note, was enough to make Hans start tapping. Of thirty such tests, twenty-nine worked. When he asked Hans to tap 13 and then stayed bent until 20, Hans tapped 20. When he asked for three plus four and did not straighten until fourteen, Hans tapped 14. Pfungst could produce wrong answers at will, because Hans was not answering the question at all.

Pfungst could produce wrong answers at will, because Hans was not answering the question at all.

The unsettling part was that Pfungst was doing it too. He noticed his own small lift of the head only after he had spotted von Osten’s, and his movements were so slight they escaped notice even when he was concentrating on them. When he tried to make the jerks on purpose, he could not keep them uniformly fine; a bigger one would slip out now and then, and he found he could not regulate the impulse beforehand.

Playing the horse

In November 1904 Pfungst took the question back to the Psychological Institute and turned it around. This time he played the horse. A volunteer stood in front of him, thought of a number, and concentrated on it while Pfungst tapped with his right hand. He had 25 subjects of all ages, including children of five and six. All but two of them gave the same upward head movement when he reached their number, and none of them realised they were doing it.

He measured the movements with recording equipment. In four subjects whose movements were traced on paper, the average upward jerk of the head was about 1 millimetre, with individual readings from a tenth of a millimetre up to about 2.3 millimetres.

Pfungst then tried reading minds the way Hans had. He asked 12 people to concentrate on concepts such as “up” or “down,” “left” or “right,” “yes” or “no,” and guessed from their expressive movements. In 350 tests he was right 73 percent of the time, and in favourable cases he reached 90 to 100 percent. He was doing what Hans did, with less practice.

Stumpf signed off the second report on 9 December 1904. Pfungst’s full account came out in German in 1907 as Das Pferd des Herrn von Osten, and the English translation, Clever Hans (The Horse of Mr. Von Osten), followed in 1911. It is free online through Project Gutenberg.

What became of Hans

Von Osten took the result badly. According to accounts of the case he was at first angry with the horse, then soon recovered his faith in him, refused to allow any further experiments and went on showing Hans to visitors. He died in 1909, still convinced that his horse could think.

Hans passed to Karl Krall, an Elberfeld merchant and devoted believer, who set up a laboratory in a stable and trained more horses, among them Muhamed, said to extract cube roots, and Zarif. Krall published Denkende Tiere (Thinking Animals) in 1912. Visiting scholars came away divided, and the Elberfeld horses never won over mainstream science.

The ending is the one part of the story nobody can tell with confidence. Hans and the other Elberfeld horses were taken for military service during the First World War, and after that his fate is not recorded. Some accounts say he was killed in action in 1916 or eaten by hungry soldiers. Neither has ever been confirmed.

The Clever Hans effect

What survived was the lesson. Psychologists began using “Clever Hans” for any case in which an animal, or a person, picks up unintended cues from the people testing it. The fix Pfungst used, making sure that the person running the test does not know the right answer, is the root of the blind and double-blind designs that are now routine in medicine and psychology.

The effect has never gone away. In a 2011 study in Animal Cognition, researchers at the University of California, Davis, ran 18 detection-dog teams through searches in which there were no drugs or explosives at all, so every alert was wrong. Handlers had been told falsely that red paper markers showed where the scent was. The teams called 225 alerts, and they were concentrated on the marked spots. What the handlers believed shaped where their dogs seemed to find things.

Animal-cognition researchers now design around it. When a German team tested a border collie named Rico in 2004, a dog whose owners said he knew the names of more than 200 toys, they put the toys in a separate room so Rico could not see his owner while choosing. He brought back 37 of 40 correctly, results the team published in Science. A 2024 study of dogs trained on recordable soundboard buttons, run from the University of California, San Diego, had owners wear sunglasses, hats and masks and had buttons pressed by strangers as well as owners. The dogs still responded appropriately to words such as “outside” and “play,” which suggested they were reacting to the words themselves rather than to their owners’ faces. If your own dog seems to understand everything you say, try asking with your face out of view, the way Rico’s testers did. Bright dogs need busy bodies too; here is how much daily exercise a dog actually needs.

The newest Clever Hans is not an animal at all. In 2019 a team from Fraunhofer Heinrich Hertz Institute and TU Berlin published a paper in Nature Communications titled “Unmasking Clever Hans predictors and assessing what machines really learn.” They examined an image classifier that scored well at recognising horses in a standard benchmark set of photos. When they looked at which pixels drove its decisions, it was ignoring the horses. About a fifth of the horse photos carried a small source tag in a lower corner, and the model had learned to look for that. Pasting the tag onto a photo of a car could make the model call it a horse.

It is a neat echo that the AI version also involves a horse, but the structure is what matters. In both cases a system gave right answers on the tests people set, for reasons nobody suspected, and the only way to find out was to change the conditions and watch it fail. The authors warned that standard accuracy scores can miss exactly this kind of shortcut.

Why the story still matters

The story is usually told as a debunking, with the horse as the fool. It reads better the other way around. Hans was not stupid. He had learned, without any instruction, to read human posture well enough to pick out a head movement of a tenth of a millimetre. Pfungst, a trained observer, needed weeks to find what Hans had found on his own, and even then he could not fully control his own.

Nor were his believers. They brought in experts, tried strangers and changed questioners. What they could not test was the one thing they could not see: themselves. Every other kind of public deception aglena has covered, from the Mechanical Turk, which hid a chess player inside a cabinet, to the séance rooms Houdini spent years exposing, depended on someone hiding something. Clever Hans hid nothing. The signal was right there in the courtyard, made by the very people trying to catch it.

That is why the name keeps coming back. A dog that seems to understand, a brain scan that seems to light up, as in the famous dead salmon in an fMRI machine, or an AI that seems to see horses can all be telling us more about our tests than about themselves. Pfungst’s fix still works: take away what the subject could be reading, keep the person asking in the dark, and see whether the answers survive.

Further reading