Human Assisted Computing: Putting CAPTCHAs to Work

If you've ever commented on someone's blog, or registered for a web application, it's quite possible that you've had to complete a CAPTCHA, a "Completely Automated Public Turing test to tell Computers and Humans Apart".

Used with permission

CAPTCHA's take their name from a "game" proposed by Alan Turing, one of the pioneers of both computer science and what came to be called artificial intelligence. Robotics and the Meaning of Life describes the Turing Test as follows:

The Turing Test involves a man and a woman answering questions by passing written messages to a questioner. The questioner can ask any question in an attempt to tell the respondents apart. The man and woman can give any answer in order to fool the questioner. You can imagine how hard it would be to tell the man and woman apart. Turing then supposed that either the man or the woman is replaced by a computer. With the same restrictions as before the questioner now has to decide which one is the computer. If the questioner can't tell the difference the computer is deemed to have passed the test.

(If you're interested, you can read Turing's original paper here: Computing machinery and intelligence.)

CAPTCHAs actually implement a "reverse" Turing test, in the sense that it is a computer that is being used to distinguish between a person and a computer, rather than the other way round.

What surprised me was that the question-response structure of CAPTCHAs is now being put to use as the basis of several "games" in which people are used to perform classification tasks, whilst "solving" CAPTCHA puzzles that authorise human activity on a website.

In particular, the reCAPTCHA test uses words taken from scanned books that computers have been unable to read properly using optical character recognition (the process that lets you turn scanned documents into word documents). So whenever you enter a reCAPTCHA word, you're actually helping a computer out somewhere...

A quick scout around actually turns up several other examples of "human assisted computing". Originally released as the ESP GAME, the Google Image Labeler shows the same image to two independent players of the game (who may be continents apart) and challenges them to come up with a descriptive label for the image that they both agree on. This label can then be used to help locate the image in the context of a web search.

Whoever thought we'd be doing unpaid piece-work for computers?!