CAPTCHA and Artificial Intelligence: How a Website Verifies a Human Without Clicking on Images

Click on all the squares with traffic lights. Then click on the crosswalks. And one more time, because one square was missing. CAPTCHA image tests were meant to distinguish humans from programs, but with the development of artificial intelligence, their original logic is falling apart. How do websites attempt to verify visitors today, and what’s truly new about it?

Why AI Has Cracked Image-Based Tests

CAPTCHA is a test that is supposed to be easy for humans and difficult for computers. However, recognizing objects in photos has become one of the tasks machines have mastered best.

In 2024, researchers from ETH Zurich demonstrated in their study “Breaking reCAPTCHAv2” that an object recognition model (YOLO), trained on approximately 14,000 annotated images of traffic scenes, could pass reCAPTCHA v2 image tests with 100% accuracy. It outperformed humans in previous studies. An important detail: to ensure the bots were indistinguishable from humans, the researchers also simulated mouse movements and used a VPN during the attack. Thus, the test was not only cracked by the image model but also by the mimicry of human behavior.

It is important to distinguish between the two. This result applies to a specific type of test. The claim that AI has “killed” all CAPTCHAs is an oversimplification. Rather, their role is changing: they are no longer a standalone barrier, but merely one of several signals that websites combine.

How clickless device verification works

One alternative is to verify that a request is coming from a real, undamaged device. This is achieved using so-called Private Access Tokens (PATs), built on the Privacy Pass protocol, which is being standardized by the IETF.

The story of a single token

Simply put: a device requests confirmation from a trusted “verifier.” Using certificates stored in the device’s security chip, the verifier checks that the device has not been compromised and that the user account is valid. It then issues a token that the website can verify, but cannot use to determine which device it comes from. Technically, blind RSA signatures are used for this, not “Zero-Knowledge Proofs,” as is sometimes claimed.

What the website sees and what the verifier sees

The website does not receive your identity or serial number. The verifier—that is, the platform provider—does, however, know about the device and the account, so this is a transfer of trust, not a removal of it. Anonymity vis-à-vis the website itself is preserved, but you’re relying on the platform not to misuse the token.

Who’s using it today

Apple introduced PAT in iOS 16 and macOS Ventura. According to available sources, one major web service uses them to skip CAPTCHA for visitors in Safari. Sources also indicate that outside this manufacturer’s ecosystem, the tokens are used only sparingly, as the value of PAT lies in device attestation—and only platform developers can reliably provide that. We have therefore not been able to verify the claim that the entire internet is “definitely switching” to such a solution.

We were unable to confirm the connection to the FIDO Alliance mentioned in the source during our research. FIDO focuses primarily on passkeys, which is a different technology.

Limitations of Device Verification

• It verifies the device, not the user’s intent. An attacker with the actual device and a regular account can obtain the token, so this protection is typically supplemented by limiting the number of requests.
• It’s not a solution for everyone. It only works where both the platform and the server support it. Other users may still see traditional authentication prompts.
• Dependence on platforms. The strength of the protection depends on the trustworthiness of the attestation chain and on who decides on it.
• It doesn’t reveal who you are. The token does not disclose whether an authorized person is using the device.

What This Means for Users

For the average visitor, this change means fewer interruptions during login or checkout, provided their device is supported. At the same time, websites will increasingly combine multiple signals: visitor behavior, device characteristics, and cryptographic tokens. CAPTCHA may not disappear entirely, but it will appear less frequently and may serve as a fallback option.

Conclusion

Image-based CAPTCHAs have lost their ability to reliably distinguish humans from machines, so the focus is shifting to device evidence and a combination of signals. Token-based authentication is interesting because it protects privacy from the web, but it is not a universal replacement and requires trust in the platforms. A reasonable perspective is this: CAPTCHA is not going away, but it is no longer the sole gatekeeper.