Published On:
November 19, 2026
Published By:

Greg Borbonus

CAPTCHAs Don't Prove You're Human Anymore

TL;DR: Passing a CAPTCHA isn't proof that someone's human. AI can solve familiar puzzles, and paid services make outsourcing them inexpensive. Keep useful bot checks, but pair them with server-side validation, sensible limits and stronger account security.

The image puzzle on a website used to feel like a clear dividing line. A person could find the bicycles. A computer would struggle. That line has become much harder to draw.

I see it in my own business. The AI assistants I run now get through image-grid CAPTCHAs on their own, sometimes after a few rounds. That's an observation, not a controlled test, but it changes how I think about that little gate on a form.

What the research actually shows

In 2024, ETH Zurich researchers reported solving 100% of the reCAPTCHA v2 challenges presented to their system. They also found evidence that cookies and browser history strongly affected the checks.

That result describes their experiment, not every CAPTCHA on every website. It does show why recognizing objects in a picture isn't a dependable human-only skill anymore.

A UC Irvine-led study published in 2023 compared its human results with previously published automated attacks. Bot accuracy ranged from 85% to 100%, while the study's human accuracy ranged from 50% to 85%. Bots were generally faster, although image reCAPTCHA times were similar. These weren't all head-to-head tests of identical current systems.

For distorted text, the same study's comparison table lists published bots taking under a second with 99.8% accuracy, compared with people taking 9 to 15 seconds.

This reaches beyond Google's puzzles. A 2021 hCaptcha study reported 95.93% accuracy across 270 live challenges. That's historical evidence about the version tested, not a measured success rate against today's hCaptcha.

A checkbox isn't a human certificate

Browser agents can interact with these screens, too. In 2025, Ars Technica reported a user-shared example of an AI agent passing a Cloudflare verification checkbox. It wasn't an image-puzzle test or evidence of a universal bypass.

Press-and-hold checks also involve more than the visible gesture. A technical examination of PerimeterX's patents describes checks involving browser characteristics, timing and user behavior. In March 2026, browser vendor Browser Use reported automated access to sites protected by PerimeterX and other vendors. Its benchmark wasn't an independent test isolating press-and-hold challenges, so I wouldn't turn it into a solve-rate claim.

Nor does abuse require a clever AI. Some commercial solving services advertise people doing the work. Anti-Captcha advertises a "vast workforce", and 2Captcha lists reCAPTCHA v2 at $1 to $2.99 per 1,000 solves. Those are advertised prices checked in November 2026, not independently verified performance. A human solving a puzzle for someone else doesn't make the eventual request trustworthy.

Why your forms and checkout need more context

OWASP's credential-stuffing guidance says CAPTCHAs can slow automated attacks, but tools and services can break them. It recommends layers rather than relying on one control.

For your site, I'd look at each action separately: contact-form spam, fake signups, repeated login attempts and suspicious checkout activity. Passing a puzzle shouldn't automatically approve everything that follows. Ask whether the request makes sense for that account and that action.

Where the alternatives are going

Risk scores and behavioral signals. Google's reCAPTCHA v3 returns a score based on interactions instead of interrupting visitors with a puzzle. Google Cloud's assessment documentation, including Enterprise features, describes scores and reasons such as automation or unusual traffic. Your application still has to interpret the result and respond. A score is a signal, not identification.

Quieter browser checks. Cloudflare Turnstile runs checks without image CAPTCHAs and works on sites that don't route traffic through Cloudflare. Reducing customer friction is useful; it doesn't remove the need for other defenses.

Private Access Tokens. Apple's Automatic Verification documentation explains how a supported device and Apple Account can be validated, with a private token sent to a participating site. This reduces puzzles without sending the site the account details. I'd treat device attestation as another signal, not permission to approve a purchase.

Proof-of-work. Friendly Captcha uses background computational puzzles alongside risk signals. The browser does work rather than asking a person to label pictures. Because software is meant to solve that puzzle, I wouldn't treat completion as proof of humanity either. Test the experience on your customers' devices.

Practical layers I'd start with

A calmer way to protect your site

The answer isn't a harder picture puzzle every time AI improves. It's several useful checks, each doing a small job. The goal is to reduce unwanted activity while keeping things comfortable for your customers.

If you'd like a second set of eyes, I'm happy to look at your forms, logins and checkout with you and help harden the parts that need attention. An Able OWASP scan is $349 for one public site, with a written report and a recheck within 30 days. If you'd rather I set up the fixes with you, Able Rescue is $199 for up to two hours of hands-on help.

Just email greg@ableadmins.com and we can start with what your site already has.