The CAPTCHA That Helped Read Books
Early reCAPTCHA puzzles did more than decide whether a visitor was human. They also helped turn scanned pages into searchable text. When a library or archive scans a book, optical character recognition (OCR) software tries to identify each printed letter and word. But stains, faded ink, unusual typefaces, or crooked pages can leave the software unsure.
Those uncertain words became part of a CAPTCHA: the little reading test a website used to screen out automated programs. A puzzle typically paired a word the system already knew with one it could not confidently read. The known word checked whether the person was solving the puzzle reliably; the answer to the unknown one supplied a human reading of the scanned page.
From Many Guesses to Usable Text
A single answer could be mistaken, so the system could compare responses from multiple people. When people independently entered the same reading, that agreement provided evidence for what the word said. The result could then be used to improve the text record of the page.
That mattered because a scan is only an image: computers cannot search its words until they have been converted into text. Human readers, encountering the hard-to-read fragments in the course of routine website visits, helped fill the gaps that automated recognition left behind. The small task of passing a security check could therefore contribute to making old printed material searchable.
