Does CAPTCHA Stop Bots? The CAPTCHA Types Scrapers Meet

It stops crude bots and slows the rest. The idea comes from a 2003 paper by Luis von Ahn, Manuel Blum, Nicholas Hopper and John Langford, who defined a CAPTCHA as "an automated test that humans can pass, but current computer programs can't pass" (CAPTCHA: Using Hard AI Problems for Security, EUROCRYPT 2003). The word "current" carried the weight: every puzzle type that was hard for computers in its day has since been solved by software.

Three studies show where things stand:

Humans vs bots by CAPTCHA type (2023):

Sources Searles et al., An Empirical Study & Evaluation of Modern CAPTCHAs, USENIX Security 2023, Table 3: human figures from 1,400 participants who solved 14,000 CAPTCHAs; bot figures from the published attacks the authors cite.

In 2024, an ETH Zurich team reported that they "can solve 100% of the captchas" of reCAPTCHA v2's image challenges in their tests with off-the-shelf object-detection models, and found that the system leans heavily on "cookie and browser history data" to decide who is human (Plesner et al., Breaking reCAPTCHAv2, COMPSAC 2024). In 2025, Xiwen Teoh and co-authors from Shanghai Jiao Tong University, the National University of Singapore and Tel Aviv University presented Halligan, which treats each challenge as a search problem and solves types it was never adapted to; they call the result "a game-changer to the known practice of adopting visual CAPTCHAs", used Halligan to infiltrate human-driven CAPTCHA farms, and point to puzzle-less anti-bot alternatives (Are CAPTCHAs Still Bot-hard?, USENIX Security 2025).

OWASP's guidance draws the same line for site owners: visible CAPTCHAs are "machine-solvable by ML" and "outsourced to human solver farms for fractions of a cent per solve", so they should be "a last-resort step-up, not a primary defense"; the goal is "to raise the cost of abusive automation" (OWASP Bot Management and Anti-Automation Cheat Sheet).

So what does a CAPTCHA still do?

• It raises the cost per request. A bot has to run a model, pay a solving service or route the puzzle to people, and each of those takes time and money. That is enough to stop spam sign-ups and low-value abuse.
• It buys time for the real decision. The puzzle is the visible part; the scoring of the IP, the browser and the behavior around it is what separates people from bots today.
• It filters the careless. Scripts with default headers, datacenter IPs and machine-regular timing get challenged first and most often.

Five routes come up again and again in the studies and in site owners' guidance:

Passing a CAPTCHA proves only that a valid response came back, not that a person produced it. MITRE ATT&CK lists the practice as technique T1696.001, "Automate CAPTCHA Completion", and advises defenders: "Do not treat successful CAPTCHA completion alone as sufficient evidence that subsequent activity is human-generated" (MITRE ATT&CK T1696.001, added in ATT&CK v20, September 2026). That is why providers now weigh the signals around the puzzle more than the answer. For scrapers, the fifth route is the one to work on, and the sections below cover it.

A modern CAPTCHA works in four steps, and the puzzle is only one of them:

The name is an acronym: Completely Automated Public Turing test to tell Computers and Humans Apart. "Public" means the method is known; the security is supposed to come from the hard problem, not from secrecy.

Two newer approaches remove the puzzle entirely. Cloudflare describes Turnstile as a "smart CAPTCHA alternative" that "works without showing visitors a CAPTCHA" (Cloudflare Turnstile), and hCaptcha offers an invisible mode in which no widget appears unless needed (hCaptcha developer guide). Privacy Pass, standardized by the IETF in June 2024, lets a device prove once, "via proof of solving a CAPTCHA, checking device or hardware attestation validity", that it is trusted, and then present unlinkable tokens instead of solving puzzles (RFC 9576).

What is the most common CAPTCHA? The checkbox with an image fallback and the invisible score-based check, from the large providers, are the ones most people meet; the rest are niche. Which providers sit behind which widgets is covered in CAPTCHA providers compared.

A scraper rarely meets a classic text puzzle. It meets the challenges that bot-protection services put in front of whole sites:

• Challenge pages and Turnstile on sites behind Cloudflare: see why Cloudflare thinks you are a bot.
• Image grids and checkboxes on search engines, marketplaces, ticketing and account pages.
• Sliders and "press and hold" pages from anti-bot services on retail and travel sites: DataDome, PerimeterX and hCaptcha challenges.
• Silent scores you never see as a puzzle, only as a 403, an empty page or an endless loop of challenges.

The challenge you get depends on your score, and the score depends on everything at once: an IP with a poor reputation, a TLS handshake that does not match the claimed browser, a headless browser's traces and too many requests per minute all add up. That is why fixing the cause works better than solving puzzles: how to avoid and handle CAPTCHAs when scraping.

The pattern repeats: each puzzle type gets harder for people as software catches up, until the providers move the real test somewhere else.

1. Distorted text (from 2003): broken by character recognition, then by machine learning.
2. Image selection: harder for early software, now solved by object-detection models (2024) and vision language models (2025).
3. Checkbox with a hidden score: the click itself proves little; the risk score behind it decides whether a puzzle follows.
4. Invisible checks and challenge pages: no puzzle at all for most visitors, only signals from the browser and the network.
5. Tokens instead of tests: Privacy Pass and similar schemes let a trusted device carry proof between sites.

For scrapers, each step moved the problem away from the puzzle and toward looking like a normal visitor: a real browser where JavaScript checks run, consistent headers and fingerprints, a calm request rate and an IP that belongs to an ordinary home connection.

• Avoid rather than solve. Every challenge you trigger costs time, and solving them at scale costs money; fewer challenges come from lower rates, real browser headers and residential IPs. How the IP type changes how often challenges appear: CAPTCHA proxies.
• Solvers exist, with limits. Solver services and their costs are compared in CAPTCHA solver services. A token from a solver can still be refused if the browser and IP that submit it look wrong.
• Mind the rules. A CAPTCHA is a clear signal that a site does not want automated access to that page. Collect only public data, read the terms, and see is web scraping legal. ProxyField's acceptable use policy forbids fraud, fake accounts and access to data behind logins you are not entitled to.
• Google results as data. If you need the results rather than the pages, the Google SERP API returns the results page as JSON, at $1.20 down to $0.80 per 1,000 successful results.

Spread requests over residential IPs: ProxyField residential proxies cost $2.95/GB for 1 GB down to $1.95/GB from 1,000 GB: 412M+ residential IPs in 195+ countries, with targeting by country, region (state) and city. Start with the Easy Trial (1 GB); traffic never expires, and all purchases are final.

Checked on October 27, 2026:

• von Ahn, Blum, Hopper and Langford, CAPTCHA: Using Hard AI Problems for Security, EUROCRYPT 2003.
• Searles et al., An Empirical Study & Evaluation of Modern CAPTCHAs, USENIX Security 2023 (Table 3).
• Plesner, Vontobel and Wattenhofer, Breaking reCAPTCHAv2, COMPSAC 2024.
• MITRE ATT&CK T1696.001, Automate CAPTCHA Completion, ATT&CK v20, September 2026.
• Teoh, Lin, Li, Liu, Sollomoni, Harel and Dong, Are CAPTCHAs Still Bot-hard?, USENIX Security 2025.
• OWASP Bot Management and Anti-Automation Cheat Sheet.
• Cloudflare Turnstile; hCaptcha developer guide; RFC 9576: The Privacy Pass Architecture, June 2024.