#1227 2007 · reCAPTCHA (Luis von Ahn, Ben Maurer, Carnegie Mellon University) · Web security / digitization
reCAPTCHA made every bot-check also digitize a word no computer could read
问题
Old scanned books had text too degraded for software to read, while millions typed CAPTCHAs for free daily
背景
By the mid-2000s, CAPTCHA — distorted text that separated humans from bots at login and signup screens — was solving security but wasting enormous amounts of human attention: Luis von Ahn, the Carnegie Mellon computer scientist who had helped popularize CAPTCHA, calculated that roughly 200 million CAPTCHAs were being solved every day worldwide, at about ten seconds each, adding up to over 500,000 human-hours daily spent typing squiggly letters that vanished the instant they were verified. Meanwhile, the Internet Archive and newspaper publishers held hundreds of thousands of scanned pages whose text optical character recognition software could not reliably read: old fonts, ink bleed, and creased paper defeated the machines even as the images themselves were perfectly legible to a person.
Von Ahn's earlier project, the ESP Game, had already shown that gameplay could extract useful labels from players as a side effect of entertainment. CAPTCHA presented the same opportunity in reverse: it was labor people were already performing, at scale, for free, but the output — a solved squiggle — was thrown away the moment it verified a login.
换别人会怎么做
The available options were to pay human transcribers at commercial rates to key in the unreadable words, which does not scale to hundreds of thousands of pages, or accept the OCR software's error rate and publish flawed digital text.
他们看到了什么
A security check and a labeling task ask the same thing: identify a distorted word. Splitting the CAPTCHA in two let the same ten seconds verify a human and transcribe an archive, unnoticed.
那一手
Von Ahn and Ben Maurer built reCAPTCHA to show users two distorted words instead of one: one word was a known CAPTCHA control the system could grade, the other was an actual unrecognized word from a scanned book or newspaper page. Typing the known word proved the user was human; typing the unknown word — checked for agreement across many different users before being accepted — supplied the correct transcription back to the digitization project, so the same ten seconds that used to vanish now did double duty.
为什么管用
It works because the marginal cost of adding a second word to an existing CAPTCHA is close to zero for the user — the friction of typing distorted text was already accepted as the price of using the site — while the value of a correctly transcribed archival word, multiplied across hundreds of millions of daily solves, is enormous. Cross-checking each unknown word against multiple independent users' answers turns an unreliable single guess into a statistically confident transcription without needing to trust any one solver.
值了多少
Digitized 13M+ NYT articles back to 1851 and ~2M books a year for Google Books, solved by over a billion people before Google retired it.
什么时候会失灵
The pattern only works where the byproduct task is genuinely as easy as the task users already tolerate — if it's harder or arouses suspicion (as reCAPTCHA increasingly did once people wondered exactly what they were transcribing), completion and goodwill both drop. It also has a shelf life: once machine vision got good enough to solve distorted text itself, the security rationale evaporated, and reCAPTCHA had to be replaced by non-text behavioral checks.
后来呢
reCAPTCHA became the standard proof-of-humanity widget across much of the web for close to a decade and established human-computation-as-byproduct as a recognized design pattern, directly informing von Ahn's next venture, Duolingo, which used the same idea to have language learners translate real sentences as they practiced.
资料来源
- [1]reCAPTCHA: Using Captchas To Digitize BooksTechCrunch, 2007techcrunch.com
- [2]reCAPTCHA (a.k.a. Those Infernal Squiggly Words) Almost Done Digitizing the New York Times ArchiveNewsweek, 2011newsweek.com