“Can AI solve CAPTCHAs?”

Can AI solve CAPTCHAs?

Updated October 18, 2026By the CaptchaKit team

Short answer

Yes, for many types. A 2024 ETH Zurich paper reported a system that solved 100 percent of reCAPTCHA v2 image challenges, where earlier work reached 68 to 71 percent. No browser CAPTCHA is unbeatable. A CAPTCHA still raises the cost of cheap bulk abuse, so use it with rate limits and a server-side check.

What the research shows for image CAPTCHAs

Plesner, Vontobel and Wattenhofer of ETH Zurich published "Breaking reCAPTCHAv2" in 2024. They report a system built on YOLO image models that solved 100 percent of reCAPTCHA v2 image challenges, compared with 68 to 71 percent for earlier work. They also report that bots needed about as many challenges as humans did to pass.

These are the authors' own results on one CAPTCHA type. The paper says nothing about every CAPTCHA, and we found no independent replication, so read it as strong evidence about image puzzles, not a verdict on all challenges.

Solver services and human farms

Software is not the only route. OWASP's entry for CAPTCHA defeat lists automated solving by optical character recognition, matching against pre-built image databases and other machine reading, and also human farms, where people are paid to answer challenges for an attacker.

OWASP titles the threat CAPTCHA defeat and lists CAPTCHA bypass, CAPTCHA solver and CAPTCHA solving among its other names. In every case the puzzle is solved, either by a program or by a person.

What a CAPTCHA is still good for

The 2003 paper that introduced the term, "CAPTCHA: Using Hard AI Problems for Security", built the idea on problems that AI could not yet solve. That always depended on the state of AI. Today a CAPTCHA is better read as a cost control: it makes cheap, high-volume abuse more expensive, and it gives your server a signed token to check.

OWASP's authentication guidance says many CAPTCHA implementations can be solved by automated techniques or outsourced to services, and calls CAPTCHA a defense in depth control that makes attacks more time-consuming and expensive, not a preventative. Its bot management guidance says to treat visible CAPTCHAs as a last-resort step-up, not a primary defense, and lists attestation tokens, invisible risk scoring and proof of work as other options, each with limits.

What a game CAPTCHA checks, and what it does not

CaptchaKit is a game, not an image puzzle, so the image-recognition result does not apply in the same way. That does not make it unbeatable. A script that drives a real browser can be built for any browser CAPTCHA.

What the service checks is described in its code and docs. A game session starts with a single-use challenge. A completion under 1.2 seconds is rejected. A session with no pointer or key interactions is treated as a likely headless client. Issuing and completion are rate limited per site key. The pass token is a signed JWT valid for 10 minutes that your server can use once. These checks raise the effort for an attacker. They are not proof of a human.

Layers that hold up

Whatever you pick, add the checks that do not depend on a puzzle: verify the token on your server, rate limit by IP and by account, and watch solve rates for sudden jumps. See are game CAPTCHAs secure for the full list of what CaptchaKit does, and honeypot vs CAPTCHA for a free layer to add first.

Questions

Can ChatGPT solve a CAPTCHA?

General chat models are not the main concern. The cited research uses purpose-built vision models trained for the task. Treat any image CAPTCHA as solvable by software.

Are CAPTCHAs obsolete?

They are weaker than they were, and OWASP treats them as one layer, not a wall. They still raise the cost of cheap bulk abuse.

Can bots solve game CAPTCHAs?

A determined attacker can script a real browser against any browser CAPTCHA. A game changes what the attacker must build, not whether it can be done. CaptchaKit adds timing, interaction and single-use token checks to raise the effort.

What are CAPTCHA farms?

Services that pay people to solve challenges for an attacker. OWASP lists human farms among the ways CAPTCHAs are defeated.

What should I use instead of a CAPTCHA?

Rarely instead, usually alongside. Rate limits, honeypots, server-side token checks and MFA on logins all work with or without a CAPTCHA.

Do CAPTCHAs stop AI agents?

Not reliably. Google describes its newer Fraud Defense product as built to verify bots, humans and AI agents, which is a sign that a plain challenge is no longer the whole answer. See is reCAPTCHA going away.