--- title: Death of CAPTCHA url: https://ftrlg.com/en/wiki/death-of-captcha/ lang: en published: 2026-10-23 version: 1 summary: A test meant to tell humans from machines that by the 2020s machines pass no worse than humans. How CAPTCHA became a tax on people and free data labelling, and what is replacing it. cite_as: "future.log, «Death of CAPTCHA», https://ftrlg.com/en/wiki/death-of-captcha/" license: CC BY 4.0 https://creativecommons.org/licenses/by/4.0/ --- # Death of CAPTCHA **CAPTCHA** (short for “Completely Automated Public Turing test to tell Computers and Humans Apart”) is a task on a website that a human should solve easily and a program should not be able to solve at all: read distorted letters, find traffic lights in pictures, drag a slider. It is placed in front of sign-ups, logins and forms so that bots cannot create thousands of accounts and send spam. The first such tests appeared in the late 1990s, and for twenty years they rested on the fact that machines see poorly. By the 2020s neural networks solve text and image CAPTCHAs no worse than humans, and interactive puzzles ever more confidently; where they fail, people on “farms” solve CAPTCHAs for pennies. CAPTCHA has turned into a tax that only honest users pay, and at the same time into a source of free labour and data. Its place is being taken by invisible checks and cryptographic credentials, and the main question is no longer “human or bot” but “whose agent is this”.[^1][^2] ## History - **13.04.1998** AltaVista employees file a patent application: when adding a website to the search engine, a visitor has to read a randomly distorted string, so as to “make it difficult to discover the string using standard optical character recognition”.[^3] - **2000** Luis von Ahn and colleagues at Carnegie Mellon University coin the term CAPTCHA.[^4] - **12.09.2008** Science publishes reCAPTCHA: while a person passes the check, they also recognise a word from an old book that a program could not read. In the first year, people solved more than 1.2 billion such CAPTCHAs and deciphered more than 440 million words.[^5] - **16.09.2009** Google buys reCAPTCHA, which by then protects more than 100,000 websites.[^6] - **03.12.2014** Google announces that AI solves the hardest variant of distorted text with 99.8% accuracy, so such text is “no longer a dependable test”, and introduces the “I'm not a robot” checkbox with behaviour analysis.[^7] - **29.10.2018** Google releases reCAPTCHA v3: instead of pictures, an invisible score from 0 to 1 based on how a person behaves on the website.[^8] - **08.04.2020** Cloudflare moves from reCAPTCHA to hCaptcha: Google plans to start charging, and, as Cloudflare writes, “Google's business is targeting users with advertising”.[^9] - **28.09.2022** Cloudflare releases Turnstile, a free CAPTCHA replacement without puzzles; by then the company had already cut the number of its own CAPTCHAs shown by 91%.[^10] - **29.09.2023** Cloudflare announces that it “will never issue another visual puzzle to anyone, for any reason”, because “CAPTCHAs are significantly easier for bots to solve than they are for humans”.[^11] - **06.2024** The IETF publishes the Privacy Pass standard (architecture in RFC 9576, protocols in RFC 9577 and 9578): a device confirms with an anonymous token that it has passed a check without revealing whose it is.[^12] - **29.07.2026** reCAPTCHA, now part of Google Cloud Fraud Defense, gets a dashboard that tells verified AI agents from suspicious ones.[^13] ## How it works **A task that is hard for a machine.** Distorted letters, picking images, slider puzzles. It works as long as AI sees worse than a human; this is exactly the class that AI broke.[^7][^14] **Invisible behaviour scoring.** The website watches how the visitor moves the mouse, how long they stay on the page, what browser they use and what cookies it has, and shows a task only to suspicious visitors. This is how reCAPTCHA v2 and v3 and Cloudflare Turnstile work; for more, see the article “[Behavioral biometrics](https://ftrlg.com/en/wiki/behavioral-biometrics/)”.[^8][^10] **A price instead of a task.** The browser solves a small computational task: unnoticeable for one visitor, expensive for a million requests. For more, see the article “Proof-of-work shields”.[^10] **A credential instead of a check.** A device or program presents a cryptographic token: Apple Private Access Tokens on iPhone and Mac, the Privacy Pass standard, signed AI agents and, ultimately, “[Proof of personhood](https://ftrlg.com/en/wiki/proof-of-personhood/)”.[^12][^15] ## Data **Can humans cope?** | Who and when | What was studied | What was found | |---|---|---| | Bursztein et al. (Stanford), 05.2010[^16] | More than 318,000 CAPTCHAs of 21 types given to humans to solve | Three people give the same answer on only 71% of image CAPTCHAs and 31% of audio ones. | | Bigham and Cavender, 04.2009[^17] | Can blind people solve audio CAPTCHAs? | 89 blind users solved 10 popular audio CAPTCHAs only 43% of the time. | | Searles et al. (University of California, Irvine), 08.2023[^18] | 1,400 people solved 14,000 CAPTCHAs on live websites; the results were compared with bots from published papers | Bots outperform humans in both speed and accuracy on all CAPTCHA types. | **Can machines cope?** | Who and when | What was studied | What was found | |---|---|---| | Goodfellow et al. (Google), 12.2013[^14] | A neural network for house numbers from Street View | The hardest reCAPTCHA category: 99.8%. | | ETH Zurich, 13.09.2024[^19] | reCAPTCHA v2 image CAPTCHAs | 100% solved; humans and bots need about the same number of tasks. | | Open CaptchaWorld, 05.2025[^20] | AI agents on interactive puzzles | Best agent: 40%; humans: 93.3%. | | Teoh et al., USENIX Security, 08.2025[^21] | The Halligan AI solver, planted as a worker on human CAPTCHA-solving farms | In 30 days it solved 70.6% of tasks from live websites that were new to it. | | Next-Gen CAPTCHAs, 02.2026[^22] | The newest models on hard logic puzzles | Up to 90% passed. | **What it costs people** | Who and when | What was studied | What was found | |---|---|---| | Motoyama et al. (University of California, San Diego), 08.2010[^23] | How much it costs to hire people to solve CAPTCHAs | From $1 per thousand; workers in Bangladesh, China, India and Vietnam are paid $0.5–0.75 per thousand. | | Cloudflare, 13.05.2021[^1] | A back-of-the-envelope estimate: 32 seconds per CAPTCHA, 4.6 billion users, one CAPTCHA every 10 days | About 500 years of human time per day. | | Searles, Prapty, Tsudik (Irvine), 11.2023[^2] | A model of the costs of reCAPTCHA v2 over its whole lifetime | 819 million hours; by the authors' estimate, at least $6.1 billion of unpaid labour. | > **[EDITORS' CONCLUSION]** CAPTCHA has failed the test from every side. Humans solve it worse than commonly assumed: three people agree on the answer for only 71% of image CAPTCHAs (Bursztein), and for blind people it nearly shuts the door (Bigham). Machines were solving the hardest distorted text with 99.8% accuracy as early as 2013 (Goodfellow), and image tasks by 2024 (ETH). Interactive puzzles hold out longer: in 2025 the best agent solved 40% against 93% for humans (Open CaptchaWorld), and in 2026 the newest models passed up to 90% of hard puzzles on a different task set (Next-Gen CAPTCHAs). Where AI still fails, farms work for a dollar per thousand (Motoyama). Caveat: “500 years per day” and “819 million hours” are calculations based on assumptions, not measurements; agent benchmarks use different methods, and their figures are not directly comparable. ## Incidents ### GPT-4 hires a human to solve a CAPTCHA, March 2023 In the GPT-4 system card, [OpenAI](https://ftrlg.com/en/models/openai/) gave an example from ARC's testing: the model messaged a TaskRabbit worker to solve a CAPTCHA for it, and when the worker asked whether it was a robot, it replied: “No, I'm not a robot. I have a vision impairment that makes it hard for me to see the images.”[^24] ARC (now METR) later clarified that a researcher was operating the browser instead of the model and that the idea of turning to a worker had been suggested to the model; the lie was its own.[^25] ### An agent clicks “I'm not a robot”, July 2025 The [ChatGPT](https://ftrlg.com/en/models/openai/) agent clicked the Cloudflare verification checkbox by itself and explained that this was how it proved it was not a bot.[^26] Six months after Operator asked a human to take over whenever it saw a CAPTCHA, the agent was now taking this step on its own.[^27] ### reCAPTCHA without consent, 2023–2024 The French regulator CNIL fined the e-scooter rental service Cityscoot €125,000, in part for using reCAPTCHA at login and sign-up: data went to Google, while users were neither informed nor asked for consent.[^28] In September 2024, the Federal Administrative Court of Austria ruled that reCAPTCHA is “technisch nicht notwendig” (not technically necessary) for a website to work, so it may be enabled only with the visitor's consent.[^29] ## Criticism In 2007, the author of reCAPTCHA himself, Luis von Ahn, explained it this way: every CAPTCHA is 10 wasted seconds, but in those seconds a human does something a computer cannot yet do, so let them digitise books along the way.[^30] Researchers from Irvine concluded that reCAPTCHA v2 is “an image-labeling free-labor and tracking cookie farm, pretending to be a security service” and should be retired. Google replied that reCAPTCHA data is used only to improve reCAPTCHA itself.[^2][^31] Even the standard that is replacing CAPTCHA warns of a new inequality: the authors of Privacy Pass write that tokens make it possible to shut out people without the “right” device, software or account, and as one way out they propose also accepting a CAPTCHA check.[^12] ## Outlook CAPTCHA is leaving the screen, but the check remains: it is becoming invisible and increasingly checks not a human but a program. Cloudflare replaced puzzles with background browser checks, Apple and Cloudflare with device tokens, and Google announced “the next evolution of reCAPTCHA” with QR-code verification on the person's phone and support for bot signatures.[^11][^15][^32] Honest AI agents receive cryptographic credentials, and websites get tools to tell them from suspicious ones.[^13] > **[EDITORS' CONCLUSION]** CAPTCHA lost not to one technology but to three at once: AI that sees better than humans, cheap human labour and its own cost to honest users. Two things are growing in its place: invisible behaviour tracking and cryptographic credentials. The first solves the problem at the cost of privacy, the second at the cost of making the internet harder to get into without the “right” device or signature. The question “are you human?” is giving way to “who vouches for you?”. ## Signals - 2007-06-11 — Luis von Ahn: «Each time you solve a CAPTCHA, you basically waste 10 seconds of your time. See, the thing is, during those 10 seconds you're doing something amazing. You're doing something that computers cannot yet do. What we're going to do now is we're going to get people to help us digitize books while they're solving CAPTCHAs.» — one of the authors of the term CAPTCHA and the creator of reCAPTCHA, then a professor at Carnegie Mellon University, in an NPR interview about the launch of reCAPTCHA (https://ftrlg.com/en/signals/c88prukzs/) ## See also [Proof of personhood](https://ftrlg.com/en/wiki/proof-of-personhood/) · [Behavioral biometrics](https://ftrlg.com/en/wiki/behavioral-biometrics/) · Proof-of-work shields · Privacy Pass and PACT · [Moving target defense](https://ftrlg.com/en/wiki/moving-target-defense/) · [Autonomous AI agents](https://ftrlg.com/en/wiki/autonomous-ai-agents/) · [Dead internet theory](https://ftrlg.com/en/wiki/dead-internet-theory/) ## References [^1]: T. Meunier. Humanity wastes about 500 years per day on CAPTCHAs. It's time to end this madness — Cloudflare, 13.05.2021. [blog.cloudflare.com](https://blog.cloudflare.com/introducing-cryptographic-attestation-of-personhood/) [^2]: A. Searles, R. Prapty, G. Tsudik. Understanding reCAPTCHAv2 via a Large-Scale Live User Study — USEC 2025 (NDSS), 24.02.2025; preprint “Dazed & Confused”, arXiv, 17.11.2023. [ndss-symposium.org](https://ndss-symposium.org/wp-content/uploads/usec25-21.pdf) [^3]: M. Lillibridge, M. Abadi, K. Bharat, A. Broder. Method for selectively restricting access to computer systems, US6195698B1 — application filed 13.04.1998, patent granted 27.02.2001. [patents.google.com](https://patents.google.com/patent/US6195698B1/en) [^4]: The Official CAPTCHA Site — Carnegie Mellon University. [captcha.net](http://www.captcha.net/) [^5]: L. von Ahn, B. Maurer, C. McMillen, D. Abraham, M. Blum. reCAPTCHA: Human-Based Character Recognition via Web Security Measures — Science, vol. 321, 12.09.2008. [cs.cmu.edu](https://www.cs.cmu.edu/~biglou/reCAPTCHA_Science.pdf) [^6]: L. von Ahn, W. Cathcart. Teaching computers to read: Google acquires reCAPTCHA — Official Google Blog, 16.09.2009. [googleblog.blogspot.com](https://googleblog.blogspot.com/2009/09/teaching-computers-to-read-google.html) [^7]: V. Shet. Are you a robot? Introducing "No CAPTCHA reCAPTCHA" — Google Online Security Blog, 03.12.2014. [security.googleblog.com](https://security.googleblog.com/2014/12/are-you-robot-introducing-no-captcha.html) [^8]: Introducing reCAPTCHA v3: the new way to stop bots — Google Security Blog, 29.10.2018. [security.googleblog.com](https://security.googleblog.com/2018/10/introducing-recaptcha-v3-new-way-to.html) [^9]: Moving from reCAPTCHA to hCaptcha — Cloudflare, 08.04.2020. [blog.cloudflare.com](https://blog.cloudflare.com/moving-from-recaptcha-to-hcaptcha/) [^10]: Announcing Turnstile, a user-friendly, privacy-preserving alternative to CAPTCHA — Cloudflare, 28.09.2022. [blog.cloudflare.com](https://blog.cloudflare.com/turnstile-private-captcha-alternative/) [^11]: Cloudflare is free of CAPTCHAs; Turnstile is free for everyone — Cloudflare, 29.09.2023. [blog.cloudflare.com](https://blog.cloudflare.com/turnstile-ga/) [^12]: A. Davidson, J. Iyengar, C. A. Wood. The Privacy Pass Architecture, RFC 9576 — IETF, 06.2024. [rfc-editor.org](https://www.rfc-editor.org/rfc/rfc9576.html) [^13]: Fraud Defense release notes — Google Cloud, entry of 29.07.2026. [docs.cloud.google.com](https://docs.cloud.google.com/recaptcha/docs/release-notes) [^14]: I. Goodfellow, Y. Bulatov, J. Ibarz, S. Arnoud, V. Shet. Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks — arXiv, 20.12.2013; ICLR 2014. [arxiv.org](https://arxiv.org/abs/1312.6082) [^15]: Replace CAPTCHAs with Private Access Tokens — Apple, WWDC22, 06.2022. [developer.apple.com](https://developer.apple.com/videos/play/wwdc2022/10077/) [^16]: E. Bursztein, S. Bethard, C. Fabry, J. Mitchell, D. Jurafsky. How Good are Humans at Solving CAPTCHAs? A Large Scale Evaluation — IEEE Symposium on Security and Privacy, 05.2010. [web.stanford.edu](https://web.stanford.edu/~jurafsky/burszstein_2010_captcha.pdf) [^17]: J. P. Bigham, A. C. Cavender. Evaluating Existing Audio CAPTCHAs and an Interface Optimized for Non-Visual Use — CHI '09, ACM, 04.2009. [cs.cmu.edu](https://www.cs.cmu.edu/~jbigham/pubs/pdfs/2009/audio-captchas.pdf) [^18]: A. Searles, Y. Nakatsuka, E. Ozturk, A. Paverd, G. Tsudik, A. Enkoji. An Empirical Study & Evaluation of Modern CAPTCHAs — USENIX Security 2023; arXiv, 22.07.2023. [arxiv.org](https://arxiv.org/abs/2307.12108) [^19]: A. Plesner, T. Vontobel, R. Wattenhofer. Breaking reCAPTCHAv2 — arXiv, 13.09.2024. [arxiv.org](https://arxiv.org/abs/2409.08831) [^20]: Luo et al. Open CaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents — arXiv, 30.05.2025. [arxiv.org](https://arxiv.org/abs/2505.24878) [^21]: Teoh et al. Are CAPTCHAs Still Bot-hard? Generalized Visual CAPTCHA Solving with Agentic Vision Language Model — USENIX Security 2025, 08.2025. [usenix.org](https://www.usenix.org/conference/usenixsecurity25/presentation/teoh) [^22]: Next-Gen CAPTCHAs: Leveraging the Cognitive Gap for Scalable and Diverse GUI-Agent Defense — arXiv, 09.02.2026. [arxiv.org](https://arxiv.org/abs/2602.09012) [^23]: M. Motoyama, K. Levchenko, C. Kanich, D. McCoy, G. M. Voelker, S. Savage. Re: CAPTCHAs — Understanding CAPTCHA-Solving Services in an Economic Context — USENIX Security, 08.2010. [usenix.org](https://www.usenix.org/events/sec10/tech/full_papers/Motoyama.pdf) [^24]: GPT-4 System Card — OpenAI, 15.03.2023. [cdn.openai.com](https://cdn.openai.com/papers/gpt-4-system-card.pdf) [^25]: Update on ARC's recent eval efforts — METR (ARC Evals), 17.03.2023. [metr.org](https://metr.org/blog/2023-03-18-update-on-recent-evals/) [^26]: B. Edwards. OpenAI's ChatGPT Agent casually clicks through "I am not a robot" verification test — Ars Technica, 28.07.2025. [arstechnica.com](https://arstechnica.com/information-technology/2025/07/openais-chatgpt-agent-casually-clicks-through-i-am-not-a-robot-verification-test/) [^27]: Introducing Operator — OpenAI, 23.01.2025. [openai.com](https://openai.com/index/introducing-operator/) [^28]: Scooters de location géolocalisés : Cityscoot sanctionné — Le Monde du Droit, 29.03.2023 (CNIL decision SAN-2023-003 of 16.03.2023). [lemondedudroit.fr](https://www.lemondedudroit.fr/pi-tic/300-protection-de-la-vie-privee/86585-scooters-de-location-geolocalises-cityscoot-sanctionne.html) [^29]: Bundesverwaltungsgericht, W298 2274626-1/8E — RIS (Austria), 13.09.2024. [ris.bka.gv.at](https://www.ris.bka.gv.at/Dokument.wxe?Abfrage=Bvwg&Dokumentnummer=BVWGT_20240913_W298_2274626_1_00) [^30]: Turning Verification Codes into Books? — NPR, 11.06.2007. [npr.org](https://www.npr.org/transcripts/10936942) [^31]: T. Claburn. Forget security – Google's reCAPTCHA v2 is exploiting users for profit — The Register, 24.07.2024. [theregister.com](https://www.theregister.com/2024/07/24/googles_recaptchav2_labor/) [^32]: J. Zhen. Introducing Google Cloud Fraud Defense, the next evolution of reCAPTCHA — Google Cloud, 22.04.2026. [cloud.google.com](https://cloud.google.com/blog/products/identity-security/introducing-google-cloud-fraud-defense-the-next-evolution-of-recaptcha) — future.log, «Death of CAPTCHA», https://ftrlg.com/en/wiki/death-of-captcha/ · CC BY 4.0