In 2000 a talk by Yahoo's chief scientist reached Carnegie Mellon: the portal was drowning in accounts opened by programs and wanted a gate no program could pass. Luis von Ahn, Manuel Blum, Nicholas Hopper and John Langford built one and named it CAPTCHA — a Completely Automated Public Turing test to tell Computers and Humans Apart. Their paper states the goal plainly, a puzzle people solve without effort and no program can solve, and adds the clause that makes the subject interesting: any program that does pass has, by that fact, solved an open problem in artificial intelligence.
L. von Ahn, M. Blum, N. Hopper, J. Langford, "CAPTCHA: Using Hard AI Problems for Security", Eurocrypt 2003, Lecture Notes in Computer Science 2656, pp. 294—311.So the test was built to be lost, and its designers said as much. Every CAPTCHA is a standing prize for whoever cracks it, and the prize was collected on schedule: distorted text, then pictures, then the checkbox. What follows is what became of the task over twenty-five years, and what now stands in its place.
| When | What happened | What it changed |
|---|---|---|
| 1996 | Moni Naor circulates an unpublished note proposing that a reverse Turing test be used to filter out programs.M. Naor, "Verification of a human in the loop, or Identification via the Turing Test", manuscript, Weizmann Institute, 1996; never formally published, but cited as the first statement of the problem. | The idea is named before anyone needs it. |
| 1997 | AltaVista puts distorted text on its add-URL form, against bulk automated submissions.M. Lillibridge, M. Abadi, K. Bharat, A. Broder, US patent 6,195,698, filed 1998, granted 2001. | The first deployment, before the name existed. |
| 2000—2003 | A Carnegie Mellon group states the problem and names it: CAPTCHA, a Completely Automated Public Turing test to tell Computers and Humans Apart.L. von Ahn, M. Blum, N. Hopper, J. Langford, Eurocrypt 2003, Lecture Notes in Computer Science 2656, pp. 294—311. The trigger was a talk by Yahoo's chief scientist at CMU in 2000. | With it, the requirement: a program that passes has solved an open AI problem. |
| 2003 | Mori and Malik read Gimpy and EZ-Gimpy, the CAPTCHAs Yahoo was then using.G. Mori, J. Malik, IEEE CVPR 2003: 92% on EZ-Gimpy, 33% on full Gimpy. | The first scheme lasts about two years. |
| 2007—2008 | reCAPTCHA attaches work to the check: on the way through, the user deciphers a scanned word OCR could not read.L. von Ahn, B. Maurer, C. McMillen, D. Abraham, M. Blum, Science 321, 12 September 2008, pp. 1465—1468. | For the first time the test produces something besides access. |
| 2014 | A Google network trained on Street View house numbers takes the hardest text variant. In December the text gives way to an "I'm not a robot" checkbox and picture grids.I. Goodfellow et al., arXiv:1312.6082, 2014 — 99.8% on the hardest category; Google Online Security Blog, April and December 2014. | The scheme is broken not by a spammer but by its operator. |
| 2016 | Measurement shows what the checkbox rests on: a nine-day-old Google cookie is enough to be handed the easy version.S. Sivakorn, I. Polakis, A. Keromytis, IEEE European Symposium on Security and Privacy, 2016. | The task is not what decides; what is known about the visitor is. |
| 2018 | reCAPTCHA v3: no task at all. A background score from 0.0 to 1.0, and the site decides what to do with it.Google Online Security Blog, 29 October 2018; reCAPTCHA v3 developer documentation. | The check stops being visible to the visitor. |
| 2022—2024 | Private Access Tokens: a site accepts a token stating the request came from a genuine, unmodified device. Privacy Pass becomes an IETF standard.Apple, WWDC22 session 10077, "Replace CAPTCHAs with Private Access Tokens"; Cloudflare and Fastly issuer launches, 2022; RFC 9576, 9577, 9578, IETF, June 2024. | The vouching is done by the hardware vendor. |
The first working CAPTCHAs asked for a word printed in warped type over clutter: Gimpy and EZ-Gimpy, both in use at Yahoo. They lasted about two years. Greg Mori and Jitendra Malik treated the problem as shape matching rather than reading, and got EZ-Gimpy right 92% of the time and full Gimpy 33% — plenty for an attack, which needs to succeed cheaply rather than always. Chellapilla and Simard at Microsoft drew the general lesson: what is hard for a machine is cutting the string into characters, not recognising a character, and cutting can be learned. The answer for the next decade was more distortion, which people paid for first. By 2014 a Google network trained on Street View house numbers was taking the hardest text reCAPTCHA at 99.8% — above what measured groups of people manage on the same material. Google published the result itself, with its own product manager's conclusion: distorted text alone was no longer enough.
G. Mori, J. Malik, IEEE CVPR 2003; K. Chellapilla, P. Simard, Advances in Neural Information Processing Systems 17, 2005; I. Goodfellow et al., arXiv:1312.6082, 2014; Google Online Security Blog, April 2014. Human accuracy on reCAPTCHA samples is about 75%: E. Bursztein et al., IEEE Symposium on Security and Privacy, 2010.In 2007 von Ahn's group noticed the size of what was being discarded: by their estimate people were typing more than a hundred million CAPTCHAs a day, a few seconds each, adding up to hundreds of thousands of human hours daily. reCAPTCHA attached a job to that. Each challenge showed two words. One came from a scanned page where two OCR programs had failed or disagreed; the other was a control whose answer was known. Get the control right and you are human, and your reading of the other word counts as a vote on what the page says — agreement between users settles it. In its first year the system passed more than 1.2 billion CAPTCHAs and deciphered over 440 million words, which the authors converted into 17,600 books. Hiring it done would have taken more than fifteen hundred people typing forty hours a week. Word accuracy came out at 99.1% against 83.5% for standard OCR on the same material, with the New York Times archive among the targets. In September 2009 Google bought reCAPTCHA and pointed it at Books and the news archive. The labour was real, unpaid and never negotiated: each contribution was a few seconds, which is exactly why there seemed to be nothing to ask about.
L. von Ahn, B. Maurer, C. McMillen, D. Abraham, M. Blum, "reCAPTCHA: Human-Based Character Recognition via Web Security Measures", Science 321, 12 September 2008, pp. 1465—1468; Carnegie Mellon University announcement of Google's acquisition of reCAPTCHA, September 2009.December 2014 replaced the typing with an "I'm not a robot" checkbox, and gave a grid of photographs to anyone the checkbox did not vouch for. The grid moved the task to where machines were still weak: naming what is in a picture. That window was already closing. Convolutional networks had won the ImageNet contest in 2012 and kept improving, and by 2016 Suphannee Sivakorn, Iasonas Polakis and Angelos Keromytis were solving the image challenge automatically 70.78% of the time, at 19 seconds each, using off-the-shelf image-tagging services and Google's own reverse image search. The same paper lists what the picture sets were made of — house numbers and street scenes among them, whose answers make Google Maps more precise. The digitising trick had not been retired, only re-aimed: from books to maps and to labelled images, the material machine vision is fed on. The check you pass to prove you are not a machine is, in part, what the machine is taught with.
Google Online Security Blog, December 2014; A. Krizhevsky, I. Sutskever, G. Hinton, Advances in Neural Information Processing Systems 25, 2012; S. Sivakorn, I. Polakis, A. Keromytis, "I Am Robot: (Deep) Learning to Break Semantic Image CAPTCHAs", IEEE European Symposium on Security and Privacy, 2016.reCAPTCHA v3, announced in October 2018, has no task in it. It runs in the background and returns a number between 0.0 and 1.0; Google's documentation calls this adaptive risk analysis and leaves the site to decide what to do with the score. Nothing is asked of the visitor, and the visitor demonstrates nothing. The number is made of the request itself: how the pointer moved, what the browser reports about itself, what the tracking cookie has seen. How far that already outweighed the task was measured on the previous version. A Google cookie nine days old was enough to be given the easy checkbox, with no browsing history behind it beyond the request that created it; the authors write that the risk analysis can be neutralised by attaching such a cookie. In the other direction, a fresh Google account — including one verified by phone — waited about sixty days for the same treatment, and a browser reporting an out-of-date version got the harder fallback challenge before the checkbox was even clicked. The authors' conclusion for an attacker was that it is better to carry no account at all.
Google Online Security Blog, 29 October 2018, and reCAPTCHA v3 developer documentation; S. Sivakorn, I. Polakis, A. Keromytis, IEEE European Symposium on Security and Privacy, 2016 — the nine-day cookie, the sixty-day threshold for a new account and the response to an outdated User-Agent.The last step removes even the appearance of a question. Private Access Tokens, shown by Apple at WWDC in June 2022, let a site accept a token attesting that a request came from a genuine, unmodified device — without saying which one. The device asks its manufacturer to vouch for it, an independent issuer — Cloudflare and Fastly were first — mints a token, and the site sees a token and no identity. The underlying protocol, Privacy Pass, became an IETF standard in June 2024: RFC 9576, 9577 and 9578. As engineering it is an improvement: no puzzle, no tracking cookie, less data about you leaving the machine. As a test it is something else entirely. Nothing about a person is checked. What is checked is that a piece of hardware is in good standing with the company that made it.
Apple, WWDC22 session 10077, "Replace CAPTCHAs with Private Access Tokens", June 2022; Cloudflare and Fastly issuer launches, 2022; RFC 9576, "The Privacy Pass Architecture", RFC 9577, RFC 9578, IETF, June 2024.It is tempting to say the CAPTCHA stopped working. It did not. Sites are still defended, and on the whole better than before. What stopped is the thing it was originally for. A test of humanity asks you to do something only a person can do. A test of reputation asks whether your device, your browser and your address resemble the ones that have behaved well. The first kind was retired because the machines caught up. The second works precisely because it never asks the question the machines have already answered.
The price of that is not evenly shared, and it was measurable before the shift was complete. The 2008 Science paper reports reCAPTCHA success rates of 97.1—97.4% from addresses in English-speaking countries against 92.6—96.9% elsewhere, a gap its authors call statistically significant. Since then the criteria have moved into territory where trying harder does not help. An old phone, a browser reporting an out-of-date version, a new account, a shared or unusual address, a network the scoring model has seen little of — each buys a harder challenge or a lower score, and none of them makes anybody less of a person. The W3C working group has made the parallel point about disabled users since 2005: the interactive task itself excludes people.
von Ahn et al., Science 321, 2008 — the by-country success rates, reported at P < 0.01. W3C, "Inaccessibility of CAPTCHA: Alternatives to Visual Turing Tests on the Web", Accessible Platform Architectures Working Group Note, first published 2005, updated 2019 and 2021. The 500 human-years figure is Cloudflare's own estimate, from a 32-second average: T. Meunier, Cloudflare blog, 13 May 2021.There is no conspiracy in any of this, which is the awkward part. Every step was a reasonable answer to the step before it, and the bill being avoided is real — Cloudflare, multiplying a 32-second average by its traffic, estimated the world was spending something like 500 human years a day on these puzzles. A reputation check is cheaper, quieter and harder to automate around. It simply is not the test that was advertised. The machine at the door no longer asks whether you are a person; it asks whether you look like the usual traffic.
Reverse CAPTCHA puts the distorted text back in front of you, because that is the one place where the original bargain still roughly holds: type OCR gives up on, a person reads at a glance. Worth a few minutes, to see what the machines of 2003 were up against — and to find out that people are not flawless at it either.
Spot the Machine runs the same experiment on writing, and it is the one to expect to lose. The direction is what matters: one boundary held for decades, the other moved in a few years.
There is an older rhyme here worth stating. In 1770 Wolfgang von Kempelen wheeled out a chess-playing automaton for Maria Theresa, and Europe spent decades arguing about how its machinery could think. A person sat inside; the Turk burned in Philadelphia in 1854 with the argument still going. What it demonstrated was the audience: people took a human for a machine, because a machine was the impressive answer. The present arrangement runs the other way. Machines pass for people well enough that nobody bothers asking any more, and the door opens on the standing of the hardware instead. The question has not changed in two and a half centuries. Only the direction of the mistake has.
The Turk: built by Wolfgang von Kempelen in 1770 for the court of Maria Theresa, destroyed by fire in Philadelphia in July 1854. T. Standage, "The Turk", 2002.