One day, the longtime New Yorker cartoon editor Bob Mankoff fed an illustration — a clown car at the repair shop, with three clowns and a businessman in place of the engine — to several large language models and asked them to explain the joke. One after the other, the models misattributed the humour to the clowns, but finally, one got it: the joke was the businessman, out of place in an otherwise coherent, fantastical scene. “I found something serious under the hood,” went that week’s winning entry to the New Yorker Cartoon Caption Contest.
It was Mankoff who created the magazine’s caption contest in 1998, challenging readers to write the funniest caption for a wacky illustration. Drawing thousands of submissions each week, the contest has become a cultural phenomenon and an irresistible dataset for computer scientists. Now, researchers see it as key to developing artificial general intelligence. “It’s like the final frontier,” said computer scientist Dafna Shahaf. AI might be able to reason through a maths problem or the LSAT, but understanding humour, in all its thorny subjectivity and intimation, is a harder task.
And for more than a decade, Mankoff has studied the caption contest alongside computer scientists, drawing on his expertise to parse which facets of humour are legible to machines and which are squarely human. “There’s this thing in psychology called the Zeigarnik effect, which is that people who didn’t finish something, for the rest of their lives, do want to finish it,” Mankoff, who is 81, told me. In the 1970s, he dropped out of his doctoral programme in experimental psychology. Today, his name is on more than a half dozen computational humour studies. “In a way, I ended up getting that PhD.”