A Horse That Seemed to Do Math Was Reading His Questioner
Researcher Bob Sturm argued music classifiers were all such horses after equalization filters and delays changed their labels.
Computerphile · Why AI is like a (Clever Hans) Horse - ComputerphileIn Germany more than a century ago, a horse was put on display in public. Somebody asked aloud, “What’s 3 + 3, Hans?” They began counting upward, and Hans neighed and stomped his feet. When the count reached six, he stopped.
Hans was born in Germany in the late 1890s and was described as extraordinary. He seemed to answer mathematics problems and other kinds of questions. Since horses cannot speak, the stomps and neighs counted as replies. He was said to handle other types of problems besides arithmetic.
A German psychologist, a scientist who studies behavior, argued that something else was happening. Hans was not counting, he said, but watching the questioner. The claim matters beyond one animal because the same shortcut, a correct answer reached for the wrong reason, can appear in artificial intelligence.
On this account, the questioner would ask the question and start the count. As the correct total approached, the questioner gave a wordless cue without meaning to, and Hans had learned to treat it as the signal to stop. He was reading expectation, not adding numbers.
A modern computer scientist put it this way: “So, uh Hans would go ooh, almost I suppose like a Pavlovian response, but a bit more sophisticated because there were more types of trigger.” Pavlovian here means learned by association.
The point is that an AI model can do the same thing. About 12 years ago, a researcher named Bob Sturm, who wrote about music classifiers, argued that those models were all such horses. He took classifiers and music, applied equalization filters and delays, and the classifications changed. Rock music with delay added was suddenly labeled reggae, and he did not know why. That turn raises a why question that calls for causal study, asking what in the music caused one label rather than another.
Whether Hans was truly doing mathematics remains disputed, and it cannot now be settled by new tests. He did not live past wartime, and his remains were likely consumed.
Questions people ask
What is meant by a Pavlovian response?
It refers to a learned association in which a cue comes to trigger an automatic response through repetition. It is named for the physiologist Ivan Pavlov, whose dogs learned to salivate at signals associated with food. In behavior science it is now a basic example of classical conditioning.
Who is Bob Sturm?
He is a researcher who works on music informatics, the use of computing to analyze and classify music. He is known for probing tests that ask whether genre classifiers understand music or use shortcuts. His filter and delay experiments became a widely cited caution about accuracy that hides flawed reasoning.
What does the Clever Hans effect mean for testing today?
It is a warning that a subject can appear to solve a task while responding to unintended hints. Animal researchers guard against it by controlling what cues the animal can see and who knows the answer. AI researchers use the same idea when they test whether a model depends on the right features or on background artifacts.