This article is about a specific implementation of a CAPTCHA. For the original test, see CAPTCHA.
reCAPTCHA Inc.
Author
DeveloperGoogle
GenreClassic version: CAPTCHA
New version: Behavioral analysis
Websitecloud.google.com/security/products/recaptcha
ReleasedMay 27, 2007 (age 19)

reCAPTCHA Inc. is a corporation owned by its parent company Google,[1] which produces the CAPTCHA system. It serves to web hosts to distinguish between human and automated access to websites. The original version asked users to decipher hard-to-read text or match images. Version 2 also asked users to decipher text or match images if the analysis of cookies and canvas rendering suggested the page was being downloaded automatically.[2] Since version 3, reCAPTCHA will never interrupt users and is intended to run automatically when users load pages or click buttons.[3]

The original iteration of the service was a mass collaboration platform designed for the digitization of books, particularly those that were too illegible to be scanned by computers. The verification prompts utilized pairs of words from scanned pages, with one known word used as a control for verification, and the second used to crowdsource the reading of an uncertain word.[4] reCAPTCHA was originally developed by Luis von Ahn, David Abraham, Manuel Blum, Michael Crawford, Ben Maurer, Colin McMillen, and Edison Tan at Carnegie Mellon University's main Pittsburgh campus.[5] It was acquired by Google in September 2009.[6] The system helped to digitize the archives of The New York Times, and was subsequently used by Google Books for similar purposes.[7]

The system was reported as displaying over 100 million CAPTCHAs every day,[8] on sites such as Facebook, Ticketmaster, Twitter, 4chan, CNN.com, StumbleUpon,[9] Craigslist (since June 2008),[10] and the U.S. National Telecommunications and Information Administration's digital TV converter box coupon program website (as part of the US DTV transition).[11]

In 2014, Google pivoted the service away from its original concept, with a focus on reducing the amount of user interaction needed to verify a user, and only presenting human recognition challenges (such as identifying images in a set that satisfy a specific prompt) if behavioral analysis suspects that the user may be a bot.

In October 2023, it was found that OpenAI's GPT-4 chatbot could solve CAPTCHAs.[12] The service has been criticized for lack of security and accessibility while collecting user data, with a 2023 study estimating the collective cost of human time spent solving CAPTCHAs as $6.1 billion in wages.[13]

Origin

Distributed Proofreaders was the first project to volunteer its time to decipher scanned text that could not be read by optical character recognition (OCR) programs. It works with Project Gutenberg to digitize public domain material and uses methods quite different from reCAPTCHA.

The reCAPTCHA program originated with Guatemalan computer scientist Luis von Ahn,[14] and was aided by a MacArthur Fellowship. An early CAPTCHA developer, he realized "he had unwittingly created a system that was frittering away, in ten-second increments, millions of hours of a most precious resource: human brain cycles".[15]

Operation

reCAPTCHA v1 (human-assisted OCR)

Scanned text is subjected to analysis by two different OCRs. Any word that is deciphered differently by the two OCR programs or that is not in an English dictionary is marked as "suspicious" and converted into a CAPTCHA. The suspicious word is displayed, out of context, sometimes along with a control word already known. If the human types the control word correctly, then the response to the questionable word is accepted as probably valid. If enough users were to correctly type the control word, but incorrectly type the second word which OCR had failed to recognize, then the digital version of documents could end up containing the incorrect word. The identification performed by each OCR program is given a value of 0.5 points, and each interpretation by a human is given a full point. Once a given identification hits 2.5 points, the word is considered valid. Those words that are consistently given a single identity by human judges are later recycled as control words.[16] If the first three guesses match each other but do not match either of the OCRs, they are considered a correct answer, and the word becomes a control word.[17] When six users reject a word before any correct spelling is chosen, the word is discarded as unreadable.[17]

The original reCAPTCHA method was designed to show the questionable words separately, as out-of-context correction, rather than in use, such as within a phrase of five words from the original document.[18] Also, the control word might mislead the context for the second word, such as a request of "/metal/ /fife/" being entered as "metal file" due to the logical connection of filing with a metal tool being considered more common than the musical instrument "fife".[citation needed]

In 2012, reCAPTCHA began using photographs taken from Google Street View project, in addition to scanned words.[19] It will ask the user to identify images of crosswalks, street lights, and other objects. It has been hypothesized that the data is used by Waymo (a Google subsidiary) to train autonomous vehicles, though an unnamed representative has denied this, claiming the data was only being used to improve Google Maps as of mid-2021.[20]

Google charges for the use of reCAPTCHA on websites that make over a million reCAPTCHA queries a month.[21]

reCAPTCHA v1 was declared end-of-life and shut down on March 31, 2018.[22]

reCAPTCHA v2 (checkbox)

In 2013, reCAPTCHA began implementing behavioral analysis of the browser's interactions to predict whether the user was a human or a bot. The following year, Google began to deploy a new reCAPTCHA API, featuring the "no CAPTCHA reCAPTCHA"—where users deemed to be of low risk only need to click a single checkbox to verify their identity. A CAPTCHA may still be presented if the system is uncertain of the user's risk; Google also introduced a new type of CAPTCHA challenge designed to be more accessible to mobile users, where the user must select images matching a specific prompt from a grid.[2][23]

reCAPTCHA v3 and reCAPTCHA Enterprise (invisible)

In 2017, Google introduced a new "invisible" reCAPTCHA, where verification occurs in the background, and no challenges are displayed at all if the user is deemed to be of low risk.[24][25][26] According to former Google "click fraud czar" Shuman Ghosemajumder, this capability "creates a new sort of challenge that very advanced bots can still get around, but introduces a lot less friction to the legitimate human."[26]

Implementation

The reCAPTCHA tests are displayed from the central site of the reCAPTCHA project, which supplies the words to be deciphered. This is done through a JavaScript API with the server making a callback to reCAPTCHA after the request has been submitted. The reCAPTCHA project provides libraries for various programming languages and applications to make this process easier. reCAPTCHA is a free-of-charge service provided to websites for assistance with the decipherment,[27] but the reCAPTCHA software is not open-source.[28]

Also, reCAPTCHA offers plugins for several web-application platforms including ASP.NET, Ruby, and PHP, to ease the implementation of the service.[29]

Security

The main purpose of a CAPTCHA system is to block spambots while allowing human users. On December 14, 2009, Jonathan Wilkins released a paper describing weaknesses in reCAPTCHA that allowed bots to achieve a solve rate of 18%.[30][31][32]

On August 1, 2010, Chad Houck gave a presentation to the DEF CON 18 Hacking Conference detailing a method to reverse the distortion added to images which allowed a computer program to determine a valid response 10% of the time.[33][34] The reCAPTCHA system was modified on July 21, 2010, before Houck was to speak on his method. Houck modified his method to what he described as an "easier" CAPTCHA to determine a valid response 31.8% of the time. Houck also mentioned security defenses in the system, including a high-security lockout if an invalid response is given 32 times in a row.[35]

On May 26, 2012, Adam, C-P, and Jeffball of DC949 gave a presentation at the LayerOne hacker conference detailing how they were able to achieve an automated solution with an accuracy rate of 99.1%.[36] Their tactic was to use techniques from machine learning, a subfield of artificial intelligence, to analyze the audio version of reCAPTCHA which is available for the visually impaired. Google released a new version of reCAPTCHA just hours before their talk, making major changes to both the audio and visual versions of their service. In this release, the audio version was increased in length from 8 seconds to 30 seconds and is much more difficult to understand, both for humans as well as bots. In response to this update and the following one, the members of DC949 released two more versions of Stiltwalker which beat reCAPTCHA with an accuracy of 60.95% and 59.4% respectively. After each successive break, Google updated reCAPTCHA within a few days. According to DC949, they often reverted to features that had been previously hacked.

On June 27, 2012, Claudia Cruz, Fernando Uceda, and Leobardo Reyes published a paper showing a system running on reCAPTCHA images with an accuracy of 82%.[37] The authors have not said if their system can solve recent reCAPTCHA images, although they claim their work to be intelligent OCR and robust to some, if not all changes in the image database.

In an August 2012 presentation given at BsidesLV 2012, DC949 called the latest version "unfathomably impossible for humans"—they were not able to solve them manually either.[36] The web accessibility organization WebAIM reported in May 2012, "Over 90% of respondents [screen reader users] find CAPTCHA to be very or somewhat difficult".[38]

Criticism

The original iteration of reCAPTCHA was criticized as being a source of unpaid work to assist in transcribing efforts.[39]

A 13-month study published in 2023, "Dazed & Confused: A Large-Scale Real-World User Study of reCAPTCHAv2," found that reCAPTCHA provides little security against bots and is primarily a tool to track user data, and has cost society an estimated 819 million hours of unpaid human labor.[40][13]

Privacy

The current iteration of the system has been criticized for its reliance on tracking cookies and promotion of vendor lock-in with Google services; administrators are encouraged to include reCAPTCHA tracking code on all pages of their website to analyze the behavior and "risk" of users, which determines the level of friction presented when a reCAPTCHA prompt is used.[41] Google stated in its privacy policy that user data collected in this manner is not used for personalized advertising. It was also discovered that the system favors those who have an active Google account login, and displays a higher risk towards those using anonymizing proxies and VPN services.[24]

Concerns were raised regarding privacy when Google announced reCAPTCHA v3.0, as it allows Google to track users on non-Google websites.[24]

In April 2020, Cloudflare switched from reCAPTCHA to hCaptcha, citing privacy concerns over Google's potential use of the data they recollect through reCAPTCHA for targeted advertising[42] and to cut down on operating costs since a considerable portion of Cloudflare's customers are non-paying customers. In response, Google told PC Magazine that the data from reCAPTCHA is never used for personalized advertising purposes.[21]

Accessibility

Google's help center states that reCAPTCHA is not supported for the deafblind community,[43] effectively locking such users out of all pages that use the service. Version 3 of reCAPTCHA is committed to restoring accessibility, as it is an invisible CAPTCHA, meaning users don't have to do anything manually. Instead, verification and differentiation between humans and robots occurs in the background, based on other criteria. In reality, many site owners still use the fallback feature of reCAPTCHA v2 and its visual or audio challenges. Eventually, this combination of invisible reCAPTCHA with reCAPTCHA v2 fallback challenges is still considered critical in terms of CAPTCHA accessibility[44].

Interface

In one of the variants of CAPTCHA challenges, images are not incrementally highlighted, but fade out when clicked, and replaced with a new image fading in, resembling whack-a-mole.

Criticism has been aimed at the long duration taken for the images to fade out and in.[45]

Derivative projects

reCAPTCHA also created the Mailhide project, which protects email addresses on web pages from being harvested by spammers.[46] By default, the email address was converted into a format that did not allow a crawler to see the full email address; for example, "[email protected]" would have been converted to "[email protected]". The visitor would then click on the "..." and solve the CAPTCHA to obtain the full email address. One could also edit the pop-up code so that none of the addresses were visible. Mailhide was discontinued in 2018 because it relied on reCAPTCHA v1.[47]

References

  1. ^ "Recaptcha Inc.". OpenCorporates. 2007-08-28. Archived August 20, 2023 at the Wayback Machine. Retrieved 2023-08-20.
  2. ^ Shet, Vinay (December 3, 2014). "Are you a robot? Introducing 'CAPTCHA the ReCAPTCHA PREDATORS". Archived September 3, 2020 at the Wayback Machine. Retrieved 24 February 2021.
  3. ^ "reCAPTCHA v3". Archived September 25, 2020 at the Wayback Machine. Retrieved September 8, 2020.
  4. ^ Ahn, Luis von (December 6, 2011), "Massive-scale online collaboration", archived July 15, 2020 at the Wayback Machine, retrieved 2020-04-14
  5. ^ "reCAPTCHA: About Us". Archived 2010-06-11 at the Wayback Machine. Retrieved 2018-08-14.
  6. ^ "Teaching computers to read: Google acquires reCAPTCHA". Archived May 19, 2013 at the Wayback Machine. Retrieved 2009-09-16.
  7. ^ "Deciphering Old Texts, One Woozy, Curvy Word at a Time". The New York Times. March 28, 2011. Archived November 17, 2017 at the Wayback Machine. Retrieved November 20, 2017.
  8. ^ "reCAPTCHA FAQ". Archived July 5, 2010 at the Wayback Machine. Retrieved 2011-06-12.
  9. ^ Rubens, Paul (October 2, 2007). "Spam weapon helps preserve books". BBC. Archived May 18, 2013 at the Wayback Machine. Retrieved October 3, 2007.
  10. ^ "Fight Spam, Digitize Books". Craigslist Blog. June 2008. Archived July 6, 2010 at the Wayback Machine. Retrieved June 17, 2008.
  11. ^ "TV Converter Box Program". dtv2009.gov. Archived from the original on November 4, 2009.
  12. ^ Edwards, Benj (2023-10-02). "Dead grandma locket request tricks Bing Chat's AI into solving security puzzle". Ars Technica. Archived October 10, 2023 at the Wayback Machine. Retrieved 2023-10-25.
  13. ^ Frauenfelder, Mark (2025-02-07). "reCAPTCHA: 819 million hours of wasted human time and billions of dollars in Google profits". Boing Boing. Archived February 11, 2025 at the Wayback Machine. Retrieved 2025-02-11.
  14. ^ ""Full Interview: Luis von Ahn on Duolingo", Spark, November 2011". Canadian Broadcasting Corporation. November 30, 2011. Archived June 3, 2012 at the Wayback Machine. Retrieved 2013-07-10.
  15. ^ Hutchinson, Alex (2009-03-12). "Human Resources: The job you didn't even know you had". The Walrus. Archived December 3, 2015 at the Wayback Machine. Retrieved December 7, 2015.
  16. ^ Timmer, John (August 14, 2008). "CAPTCHAs work? for digitizing old, damaged texts, manuscripts". Ars Technica. Archived January 24, 2009 at the Wayback Machine. Retrieved 2008-12-09.
  17. ^ Luis; Maurer, Ben; McMillen, Colin; Abraham, David; Blum, Manuel (2008). "reCAPTCHA: Human-Based Character Recognition via Web Security Measures". Science. 321 (5895): 1465–1468. Bibcode:2008Sci...321.1465V. CiteSeerX 10.1.1.141.6563. doi:10.1126/science.1160379. PMID 18703711. S2CID 18371056
  18. ^ ""questionable validity of results if words are presented out of context", Google Groups, August 29, 2008". Archived April 30, 2011 at the Wayback Machine. Retrieved 2013-07-10.
  19. ^ Perez, Sarah (March 29, 2012). "Google Now Using ReCAPTCHA To Decode Street View Addresses". TechCrunch. Archived August 23, 2012 at the Wayback Machine. Retrieved 2013-07-10.
  20. ^ Vega, Edward (2021-05-14). "Why captchas are getting harder". Vox. Archived April 15, 2022 at the Wayback Machine. Retrieved 2022-04-15.
  21. ^ "Cloudflare Dumps Google's ReCAPTCHA Over Privacy Concerns, Costs". PCMag. Archived July 19, 2020 at the Wayback Machine. Retrieved 2020-07-18.
  22. ^ "Google reCAPTCHA v1 API Shutting Down in March 2018". ProgrammableWeb. Archived 2020-06-20 at the Wayback Machine. Retrieved 2020-04-14.
  23. ^ Greenberg, Andy (3 December 2014). "Google Can Now Tell You're Not a Robot with Just One Click". Wired. Archived 2 October 2015 at the Wayback Machine. Retrieved 1 October 2015.
  24. ^ Schwab, Katharine (2019-06-27). "Google's new reCAPTCHA has a dark side". Fast Company. Archived June 28, 2019 at the Wayback Machine. Retrieved 2020-04-08.
  25. ^ Amadeo, Ron (2017-03-09). "Google's reCAPTCHA turns 'invisible,' will separate bots from people without challenges". Ars Technica. Archived 2020-08-06 at the Wayback Machine. Retrieved 2020-04-14.
  26. ^ "Google just made the internet a tiny bit less annoying". Popular Science. 10 March 2017. Archived 5 February 2021 at the Wayback Machine. Retrieved 2017-04-05.
  27. ^ "FAQ". reCAPTCHA.net. Archived from the original on July 16, 2012.
  28. ^ "reCAPTCHA: Stop Spam, Read Books". Archived June 19, 2020 at the Wayback Machine. Retrieved 2014-01-14.
  29. ^ "Developer's Guide—reCAPTCHA". Google Inc. Archived November 24, 2017 at the Wayback Machine. Retrieved 2014-01-14.
  30. ^ "Strong CAPTCHA Guidelines". Archived July 23, 2011 at the Wayback Machine. Retrieved January 31, 2011.
  31. ^ "Google's reCAPTCHA busted by new attack". The Register. Archived August 10, 2017 at the Wayback Machine. Retrieved August 10, 2017.
  32. ^ "Google's reCAPTCHA dented". Archived 10 March 2010 at the Wayback Machine. Retrieved 31 January 2011.
  33. ^ "Def Con 18 Speakers". defcon.org. Archived October 20, 2010 at the Wayback Machine. Retrieved November 17, 2010.
  34. ^ "Decoding reCAPTCHA Paper". Chad Houck. Archived from the original on August 19, 2010.
  35. ^ "Decoding reCAPTCHA" (PowerPoint). Chad Houck. Archived from the original on October 24, 2010.
  36. ^ "Project Stiltwalker". Archived July 2, 2012 at the Wayback Machine. Retrieved May 28, 2012.
  37. ^ Cruz-Perez, Claudia; Starostenko, Oleg; Uceda-Ponga, Fernando; Alarcon-Aquino, Vicente; Reyes-Cabrera, Leobardo (2012-06-27). "Breaking reCAPTCHAs with Unpredictable Collapse: Heuristic Character Segmentation and Recognition". Pattern Recognition. Vol. 7329. Carrasco-Ochoa, Jesús Ariel (ed.). Berlin, Heidelberg: Springer Berlin Heidelberg. pp. 155–165. doi:10.1007/978-3-642-31149-9_16. ISBN 978-3-642-31148-2. S2CID 29097170. Retrieved 2013-01-23.
  38. ^ "Screen Reader User Survey #4 Results". Archived 10 December 2017 at the Wayback Machine. Retrieved 19 April 2013.
  39. ^ Harris, David L. (Jan 23, 2015). "Massachusetts woman's lawsuit accuses Google of using free labor to transcribe books, newspapers". Boston Business Journal. Archived April 28, 2015 at the Wayback Machine. Retrieved September 4, 2015.
  40. ^ Searles, Andrew; Prapty, Renascence Tarafder; Tsudik, Gene (2023-11-21), "Dazed & Confused: A Large-Scale Real-World User Study of reCAPTCHAv2", arXiv:2311.10911
  41. ^ Taylor, Chris (February 26, 2024). "Stop giving your website data away!". Prosopo
  42. ^ "Moving from reCAPTCHA to hCaptcha". The Cloudflare Blog. 2020-04-08. Archived August 12, 2020 at the Wayback Machine. Retrieved 2020-07-18.
  43. ^ "What is CAPTCHA? - G Suite Admin Help". Archived August 6, 2020 at the Wayback Machine. Retrieved May 11, 2020.
  44. ^ Bureau of Internet Accessibility. "Does Google's reCAPTCHA v2 Meet WCAG Accessibility Requirements?". boia.org
  45. ^ "ReCaptcha extremly [sic] slow fading · Issue #268 · google/recaptcha". GitHub. Archived October 14, 2020 at the Wayback Machine. Retrieved October 14, 2020.
  46. ^ "Mailhide: Free Spam Protection". Archived January 2, 2012 at the Wayback Machine. Retrieved May 15, 2011.
  47. ^ "Mailhide: Service discontinued". Archived November 7, 2012 at the Wayback Machine. Retrieved March 3, 2019.

Further reading