The AI ​​can recognize pedestrian crossings and traffic lights to solve CAPTCHA tests

If theartificial intelligence it is so “intelligent” that it can generate complex content (such as music, text, video, audio, etc.), as it never knows recognize pedestrian crossings and traffic lights to solve CAPTCHA tests? A question of this type was asked by a reader to the editorial staff of BBC Science Focus Magazine and, according to the answer given by Dr. Peter Bentley (computer scientist and author working at University College London), AI is actually capable of doing this, contrary to what many might think.

Bots solve CAPTCHAs with 96% accuracy

A CAPTCHA is essentially a proof based on the Turing test that is used to understand if a user is human or if it is a computer. CAPTCHAs can be of various kinds and one of these consists in showing the user images asking him to select those that show a certain subject – usually pedestrian crossing, traffic lights, stairs, house numbers, Automobilesand similar – or it could be a distorted text which must be interpreted and typed correctly.

According to research published in July 2023, the vast majority of bots, thanks to artificial intelligence, are able to solve CAPTCHA images with 96% accuracyeven surpassing the human users involved in the study, who achieved a test resolution accuracy level of between 50% and 86%.

CAPTCHA Examples |  Geopop
Two examples of CAPTCHAs

The AI ​​passes even the most complex CAPTCHA tests

As AI systems become more and more proficient, the future of CAPTCHAs appears set to change. According to reports from Dr. Peter BentleyIn fact, AIs have become so skilled that they are able not only to solve relatively simple CAPTCHAs (such as those based on images), but they are even able to imitate humans to fool bot detectors by copying characteristic traits of us humans, such as the poor precision or even the movements of the mouse pointer when we decide which boxes to click.

This will soon make even the so-called ineffective reCAPTCHAi.e. the tests that consist of the simple click of the I'm not a robot box and which analyze a whole series of behaviors performed by the user before and after this operation. In the article published on BBC Science Focus MagazineIn fact, Dr. Bentley wrote:

The “I'm not a robot” box has multiple layers of encryption, as it measures a surprising amount of data about the user: time zone, IP address, screen size, browser and plugins, keystrokes , mouse clicks, browsing history and other things we may not be aware of. Will AI soon be able to fool these too? Yes.

don't miss this article

What are Captcha codes and how do they verify that we are not robots

BBC Science Focus Magazine