n Berlin in the early 1900s, a retired schoolteacher named Wilhelm von Osten exhibited a horse he had taught, he said, arithmetic. Asked a sum, Hans tapped out the answer with a hoof. He also appeared to read, spell, and tell the time. Crowds came; so did newspapers; so, eventually, did science.

the commission

In September 1904 a commission of thirteen people, including the psychologist Carl Stumpf, a zoo director, veterinarians and a circus manager, examined Hans and found no sign of trickery: Hans answered even when von Osten was not the one asking. That was true and it was not the end of the story.

the experiment

Stumpf asked Oskar Pfungst, who worked in his Berlin institute, to investigate. Pfungst varied one thing at a time. When the questioner knew the answer and Hans could see the questioner, Hans was nearly always right. When the questioner did not know the answer, or Hans wore blinkers, he was nearly always wrong. Hans was watching people: as his taps approached the right number, the questioner tensed or leaned slightly, and when it arrived, relaxed or straightened. Hans stopped. Nobody, including von Osten, knew they were doing it.¹

The point is not that the horse was stupid. Hans did something difficult: he read tiny, involuntary cues in strangers reliably. He solved the problem he was actually rewarded for, which was not the problem everyone thought they had set.

the legacy

The Clever Hans effect is now the name for any performance that rests on unintended cues. It is why experiments are run blind, and why Rosenthal and others took experimenter expectancy seriously. In machine learning, Lapuschkin et al. (2019) used "Clever Hans" for classifiers that scored well by attending to watermarks and borders rather than the objects. See shortcut learning.

The tour has a room of machines that answer, and its kiosk is worth asking about horses. The plain tour lists it.

notes

1. Pfungst's report is on our archive page, with a link to the free text. If you would rather be the questioner, there is a small reconstruction where something taps and you decide when it has reached the answer.
   [playful] He is still about. On the research pages, type his name and he will tap out the hour for you.

Sources

1. Pfungst (1911). Clever Hans (the Horse of Mr. von Osten), trans. C. L. Rahn. New York: Henry Holt. German original 1907. Free edition at Project Gutenberg. www.gutenberg.org/ebooks/33936
2. Rosenthal (1979). The file drawer problem and tolerance for null results. Psychological Bulletin. doi:10.1037/0033-2909.86.3.638
3. Lapuschkin et al. (2019). Unmasking Clever Hans predictors and assessing what machines really learn. Nature Communications. doi:10.1038/s41467-019-08987-4

see also

	archive: Pfungst — the primary source
	shortcut learning — Clever Hans in machines
	essay: the polite machine — the most agreeable student ever
	glossary: Clever Hans effect