reCAPTCHA: von Ahn turned 200M daily CAPTCHAs into free NYT archive digitization

In 2000, Yahoo Mail had a spam problem: bots filled out the free sign-up form day and night, generating millions of fake accounts. Carnegie Mellon grad student Luis von Ahn built a test that tells humans from computers and that a computer can grade. It went into Yahoo's registration flow within weeks of him showing Yahoo's chief scientist, and its long acronym — Completely Automated Public Turing Test To Tell Computers And Humans Apart — stuck as CAPTCHA. Others invented similar tests, but his version took off: 200 million solves a day, at email signups, ticket sales, MySpace.

Each solve wasted about 10 seconds, and von Ahn felt responsible for the annoyance — half a million-plus hours of human labor burned per day. Meanwhile the mid-2000s push to digitize old books was stuck: scanners made pictures of pages, but computers couldn't read them, so the scans weren't searchable. He reused the wasted seconds: take the words the computer couldn't recognize and make those the CAPTCHA, so people read them while proving they weren't robots. He stopped giving the test away and started a company, reCAPTCHA, to sell archive digitization one CAPTCHA at a time.

The New York Times became the first client after its CIO heard von Ahn give a talk: the extra word people typed next to the random letters came from its 130-year archive of old editions, one typed word at a time. Google — then out to digitize every book in the world — bought the company in 2009 to help digitize books. Later Google's CAPTCHAs showed street addresses from building facades, and every solve quietly improved Google Maps, unpaid. Von Ahn ran the same trick again: a language company where users translated texts while learning became Duolingo.

The test was already mandatory — so make the words OCR can't read BE the test: one set of keystrokes both blocks bots and transcribes the archive.

The cost already existed and was unavoidable; the marginal keystroke was free, so security stopped being pure friction.

Every side gained: sites got bot protection, archives got human reads of OCR failures, users paid only the 10 seconds they were already paying.

It scaled as a business: the NYT signed on as first client after its CIO heard von Ahn speak, and Google bought reCAPTCHA in 2009.

The move proved transferable: von Ahn reran the same do-two-things-with-one-action design in language learning and got Duolingo.

A mandatory cost paid by millions is hidden free capacity. Ask what payload unavoidable keystrokes could carry: pairing your task with the one people must do anyway turns friction into labor.

CAPTCHA-solving farms undercut the wall but not to zero: Kanich's 2008 study found services charging $1 per 1,000 human-solved tests — cheap, but still a price that stopped casual abuse. Tests hardened until humans struggled, and in 2014 Google rebuilt it as reCAPTCHA v2: an I-am-not-a-robot checkbox plus click-the-picture tests. Off-the-shelf deep learning broke v2 (published 2016), and engineering lead Aaron Malenfant warned the challenges would stop working within three to five years. Google's answer was v3: no test at all, just a per-site behavioral score of how human you look.