The "Completely Automated Public Turing test to tell Computers and Humans Apart" (CAPTCHA) is the squiggly word that appears on Web sites to stop bots from sending spam and doing other vile deeds. In the Sept. 12 issue of SCIENCE magazine (Vol 321 p. 1465), computer scientists Luis von Ahn, Benjamin Maurer, Colin McMillen, David Abraham and Manuel Blum from the Computer Science Department of Carnegie Mellon University in Pittsburgh report on an innovative application of CAPTCHAs: potentially using the more than 100 million applications of human intelligence in decoding the symbols for useful work.

The first application involves supplementing machine intelligence applied to optical character recognition (OCR). Currently, there is an enormous worldwide effort to transcribe existing printed documents into digital form for increased availability and (one hopes) long-term storage (although the stability and usability of digital storage in the face of technological change is the subject of much concern). 

The reCAPTCHA system "is used by more than 40,000 Web sites… and demonstrates that old print material can be transcribed, word by word, by having people solve CAPTCHAs throughout the World Wide Web. Whereas standard CAPTCHAs display images of random characters rendered by a computer, re-CAPTCHA displays words taken from scanned texts."

Words which OCR programs have not been able to recognize are stored and then randomly supplied from a database as part of a two-word CAPTCHA; the second word is a regular computer-generated CAPTCHA. Both the graphical symbol from the database of uncertain words and a computer-selected ordinary-word CAPTCHA are suitably distorted to prevent machine recognition. If the user types in the correct spelling of the second CAPTCHA, then the user’s interpretation of the first CAPTCHA is recorded as a possible transcription.

"To account for human error in the digitization process, reCAPTCHA sends every suspicious word to multiple users, each time with a different random distortion" and combined with different control words. The process includes additional controls in cases of discrepancies. Careful analysis of the results suggests accuracy higher than 99% - acceptable by industry standards and better than standard OCR. Furthermore, the researchers found that it takes no longer (around 13 seconds) to decipher a two-word reCAPTCHA than to decipher a one-word CAPTCHA that uses gibberish.

M. E. Kabay, PhD, CISSP-ISSMP, is Program Director of the Master of Science in Information Assurance program at Norwich University.