CAPTCHA Began in 2000 When Yahoo's Udi Manber Told Carnegie Mellon That Bots Posing as Teenagers Were Working Its Chat Rooms

Manber’s complaint was practical. Besides the fake teenagers, “spam companies were creating havoc by writing programs that swiftly registered for hundreds of free Yahoo e-mail accounts,” the Times reported. “What we needed,” Manber said, “was a simple way of telling a human user from a computer program.”

Blum, a Carnegie Mellon computer science professor who was on that call, flipped the problem around. Computers were still bad at things a 5-year-old does without thinking, like reading messy text. So why not build a test out of exactly those failures? With his Ph.D. student Luis von Ahn and others, he built puzzles that “computers can generate and grade the tests even though they cannot pass them.” They named them Captchas, short for “Completely Automated Public Turing Test to Tell Computers and Humans Apart.”

The project’s own Gimpy page says it “was originally built for (and in collaboration with) Yahoo! to keep bots out of their chat rooms,” to keep scripts from collecting e-mail addresses and to stop programs from posting classified ads. Its words came from “an 850 word dictionary based on Ogden’s Basic English.” The full version showed seven distorted, overlapping words and asked you to type three. Yahoo’s registration page got the easy one: a single warped word on a busy background.

The team’s paper, “CAPTCHA: Using Hard AI Problems for Security” by von Ahn, Blum, Nicholas Hopper and John Langford, spells out the bargain. If the test holds, you can tell people from programs. If someone breaks it, “a useful AI problem is solved.” The authors call that “a win-win situation.” Blum put it more simply to the Times: “It’s like there are two lollipops and no matter what you get one of them.” The paper’s acknowledgments thank Manber “for his suggestions.”

The paper’s examples show how ordinary the bot problem already was. In November 1999, a tech news site’s online poll asked readers to name the best computer science graduate school. Carnegie Mellon students wrote a program to vote thousands of times, MIT students wrote their own the next day, and MIT won 21,156 to 21,032. Every other school got under 1,000.

The first lollipop didn’t last long. By October 2002, Berkeley’s Jitendra Malik and his student Greg Mori had a program that cracked Yahoo’s easy Gimpy “over 80 percent of the time,” in a few seconds a puzzle. Manber had already left Yahoo for Amazon.

I like how honest the original pitch was. Blum didn’t promise Yahoo a wall. He promised a race, and he told the Times he’d be happy with either result. Every grid of buses and crosswalks you’ve clicked since is another lap of that race. The thing worth remembering is that the people who invented the test knew it had a shelf life on the first day, and planned for it.