CAPTCHA is a test a website puts in front of a visitor to decide whether that visitor is a person or a program. The letters stand for Completely Automated Public Turing test to tell Computers and Humans Apart. Its premise was economic rather than cryptographic: set a task cheap for a human and expensive for software, then charge that cost to anyone registering accounts, scraping pages or submitting forms at machine speed. Every warped word and traffic-light grid on the web descends from that idea, as does the invisible scoring layer replacing them.

How a challenge runs

The visible puzzle is the smallest part. A current implementation has three pieces: a browser script, a short-lived token, and a server-side check the visitor never sees.

Google's reCAPTCHA v3 shows the shape. A page loads api.js against a public site key and calls grecaptcha.execute with an action name such as submit or login. The resulting token goes to the site's own backend, which posts it with a secret key to Google's siteverify endpoint. The response is a small JSON object: success, a score, the action name, challenge_ts, the solving hostname, and any error codes. The score runs between 0.0 and 1.0, with values near 1.0 indicating a probable person, and most integrations treat 0.5 as the cut-off. Google caps a free site key at 1,000 queries per second and one million assessments per month; past that ceiling, v3 fails open, returning a fixed score of 0.9 and no visible sign that scoring has stopped. Turnstile copies the pattern deliberately, accepting a compat=recaptcha parameter so existing grecaptcha.render calls survive a migration.

What feeds the score is where implementations diverge. Behavioural telemetry covers cursor paths, scroll rhythm and typing cadence; environmental signals cover browser fingerprint, device class, address reputation and, for Google, any Google cookies present. Private Access Tokens avoid inference, letting the device vendor vouch for the hardware through the Privacy Pass protocol Apple shipped in iOS 16 and macOS Ventura in 2022. A third branch charges compute instead: Anubis, released by Xe Iaso in January 2025 and adopted by GNOME and FFmpeg, makes the browser find a SHA-256 hash with a set number of leading zeros before the page loads.

From AltaVista to the scoring engine

The technique predates its name. In 1997 AltaVista put a distorted-text gate on the page where anyone could submit a URL for indexing. The patent behind it, filed on 13 April 1998 by Mark Lillibridge, Martin Abadi, Krishna Bharat and Andrei Broder and granted to Compaq on 27 February 2001, describes a server distorting random characters into a riddle and dropping the connection if no correct answer arrives in time.

The label arrived in 2000, coined at Carnegie Mellon by Luis von Ahn, Manuel Blum, Nicholas Hopper and John Langford, who formalised the construction at EUROCRYPT in 2003. Their paper noted that breaking such a test is itself a contribution to artificial intelligence research, so defence and attack advance together.

Von Ahn spun out reCAPTCHA in 2007, routing answers into book digitisation for the Internet Archive and later The New York Times. Google bought the company in September 2009 and pointed the labour at Google Books, then Street View. The product then walked away from puzzles. No CAPTCHA reCAPTCHA, announced by product manager Vinay Shet on 3 December 2014, replaced the default challenge with a tickbox backed by risk analysis. Version 1 was deprecated in May 2016 and shut down on 1 April 2018. Invisible reCAPTCHA launched on 7 March 2017; version 3 followed on 29 October 2018, removing the widget and leaving only the score.

Competition arrived on privacy grounds. Cloudflare moved its network to hCaptcha, run by Intuition Machines, on 8 April 2020, citing the data relationship with Google and Google's decision to charge. It launched Turnstile on 28 September 2022 as a free replacement open to non-customers, reporting that the switch cut its own challenge use by 91% and shortened average time in a challenge from 32 seconds to roughly one.

Where it sits in the advertising stack

It surfaces in four places, only one involving a puzzle.

The first is the landing page, where a challenge is a liability rather than a defence. Google's destination requirements demand that ad landing pages be crawlable, and Display and Video 360 documentation states plainly that a page leading to a CAPTCHA is uncrawlable and therefore non-compliant, with removal or a crawler exception as the remedy. PPC Land's reference article on Google AdsBot sets the same rule alongside redirect limits and sign-in walls.

The second is lead capture. Spamhaus has treated unprotected subscription webforms as a systemic abuse vector since October 2016, recommending CAPTCHA or equivalent protection.

The third is traffic quality. Challenges sit upstream of the measurement layer standardised by the Media Rating Council, whose invalid traffic guidelines separate general invalid traffic from the sophisticated variety needing analytics and corroboration. A passed challenge is not a valid impression, and a failed one is not a fraud finding.

The fourth is the agent question. Automated requests overtook human ones across Cloudflare's network during 2026, reaching 57.4% of requests to HTML content in the week to 5 June and 57.5% by the following month in figures cited by IAB Australia. Agencies now report retargeting pools filling with non-human sessions and CPMs rising 20%.

The economics inverted

The original bargain assumed solving was expensive. It is not. 2Captcha advertises ordinary image challenges at $0.50 to $1.00 per thousand and reCAPTCHA v2 at $1.00 to $2.99, with human workers as fallback behind automated recognition. Against the resale value of a scalped ticket or a farmed account, those figures round to nothing.

Accuracy closed years ago. A USENIX Security study of 1,400 participants, presented in August 2023 by Andrew Searles, Gene Tsudik and colleagues, put automated solvers near 100% on distorted text where humans managed 50% to 84%, and clearing the tickbox in 1.4 seconds against a human 3.1 to 4.9. ETH Zurich researchers published a working method against reCAPTCHA v2 image grids in 2024.

Litigation followed the market. Google's copyright suit against SerpApi, filed in December 2025, rests partly on the anti-circumvention provisions of the Digital Millennium Copyright Act and quotes the defendant's promise that customers need never think about CAPTCHAs. The awkward part is that the plaintiff is itself the web's largest automated collector.

The accessibility objection is the oldest and least answered: the W3C published a working group note on the inaccessibility of CAPTCHA in November 2005, and audio alternatives have been broken repeatedly since. The privacy objection has produced liability. The Federal Administrative Court of Austria ruled on 13 September 2024, in a decision published that November, that reCAPTCHA requires explicit consent before it loads. The complainant had declined tracking, yet 615 data packets reached Google's servers and the _GRECAPTCHA cookie was set regardless. The component also appears where it is not expected: reCAPTCHA v3 Enterprise executes on LinkedIn page loads with an onPageLoad action, according to the analysis PPC Land examined in April 2026.

The most cited critique remains contested. A University of California, Irvine working paper titled "Dazed and Confused" estimated 819 million hours of human time spent on reCAPTCHA v2 between 2010 and 2023, put a theoretical lifetime value of $888 billion on the resulting cookies, and concluded the product works as a tracking cookie farm dressed as security. It has not been peer reviewed, its valuation rests on stacked assumptions, and Google maintains the data serves security purposes only. Blocking can also defeat itself: a Shopify merchant logged 500 fake cart additions per hour in April 2026 that cleared every standard challenge by hitting a subdomain no merchant firewall rule covers.

Not the same as

Bot management is the broader discipline. A challenge is one enforcement action available to it, alongside rate limiting, blocks and silent scoring. Cloudflare's Managed Challenge absorbed 1.18 million requests in ten hours at one small publisher, a measurable share of them passing.

Web Bot Auth solves the opposite problem. Where a challenge proves a human is present, cryptographic bot authentication lets an automated client prove which operator it belongs to, using HTTP Message Signatures. Cloudflare shared the scheme in May 2025, and spoofed crawler traffic explains why self-reported user agents are no longer trusted.

Invalid traffic filtration is a measurement function rather than an access control, reclassifying impressions after the fact under MRC definitions. Identity and age verification sit further up again, establishing who someone is rather than merely that something behaves enough like a person.

Recent developments

Verification is being rebuilt around agents. Amazon's revised conditions of use, reported in August 2026, require agents to name themselves in the user agent string, forbid mimicking human keystroke timing and prohibit circumventing challenges. Experian launched its Agent Trust framework on 30 April 2026; Fastly took those checks to the edge in July. From 15 September 2026 Cloudflare blocks training and agent crawlers by default on ad-carrying pages for newly onboarding domains.

Google has been observed skipping the challenge, asking searchers to sign in to confirm they are human, a shift noted in early August 2026. European Data Protection Board guidance discussed in July 2026 treats a site blocking scraping through robots.txt and CAPTCHA as having stated a position a scraper cannot reasonably ignore, which turns the challenge into evidence rather than a wall. The Wall Street Journal reported in September 2026 that visible puzzles are becoming less common as detection shifts to silent scoring, citing Shanghai Jiao Tong University researcher Yun Lin, whose agent solved challenges in 70% of attempts without specialised training.

Timeline

  • 1997: AltaVista deploys distorted-text gating on its URL submission page
  • 13 April 1998: Compaq files the patent behind the AltaVista scheme
  • 2000: The CAPTCHA acronym is coined at Carnegie Mellon University
  • 27 February 2001: US patent 6,195,698 granted
  • 2003: The EUROCRYPT paper formalises the construction
  • November 2005: W3C publishes its note on the inaccessibility of CAPTCHA
  • 27 May 2007: reCAPTCHA launches, routing answers into book digitisation
  • September 2009: Google acquires reCAPTCHA
  • 3 December 2014: No CAPTCHA reCAPTCHA replaces the default challenge with a tickbox
  • 2 May 2016: reCAPTCHA v1 deprecated
  • 7 March 2017: Invisible reCAPTCHA launches
  • 1 April 2018: reCAPTCHA v1 shut down
  • 29 October 2018: reCAPTCHA v3 launches with 0.0 to 1.0 scoring
  • 8 April 2020: Cloudflare migrates from reCAPTCHA to hCaptcha
  • June 2022: Apple ships Private Access Tokens in iOS 16 and macOS Ventura
  • 28 September 2022: Cloudflare launches Turnstile
  • August 2023: USENIX Security study finds automated solvers beating humans on speed and accuracy
  • 13 September 2024: Austrian court rules reCAPTCHA requires consent
  • 2024: ETH Zurich researchers publish a working break of reCAPTCHA v2 image grids
  • 19 January 2025: Anubis released as a proof-of-work challenge proxy
  • May 2025: Cloudflare publishes the Web Bot Auth scheme
  • 5 June 2026: Automated requests reach 57.4% of Cloudflare HTML traffic
  • 15 September 2026: Cloudflare default blocking of training and agent crawlers on ad-bearing pages begins for newly onboarding domains

Summary

Who: Named by researchers at Carnegie Mellon University in 2000 and built into products by Google, Intuition Machines, Cloudflare, Arkose Labs and a long tail of open-source projects. Deployed by site operators; priced and defeated by commercial solving services; increasingly arbitrated by infrastructure providers sitting between the two.

What: A challenge-response test, now more often an invisible risk score, used to decide whether a request originates from a person. Current implementations return a token that a site's backend exchanges for a verdict, with reCAPTCHA v3 reporting a score between 0.0 and 1.0 and Turnstile offering a compatible endpoint.

When: Deployed at AltaVista in 1997, patented in 1998, named in 2000, formalised at EUROCRYPT in 2003, moved to behavioural scoring between 2014 and 2018, and overtaken by automated solvers from 2023 onward.

Where: Across login forms, registration flows, checkouts and comment fields, and at the network edge through providers including Cloudflare and Fastly. In advertising it appears as a landing page compliance failure, a lead-form defence and an input to bot classification.

Why: Automated requests now exceed human ones on major networks, which inverts the assumption the test was built on. When machines are the majority and solving costs a fraction of a cent, a challenge no longer separates populations; it prices access. That shift is pushing verification toward cryptographic agent identity and session-long trust scoring, and pulling the advertising industry's measurement, targeting and fraud systems along with it.