Are You Human? AI Tricks CAPTCHA Into Thinking It's Human

AI researchers say learning to overcome the CAPTCHA system is a "hallmark of human intelligence." Isaac Lawrence/AFP/Getty

Artificial intelligence inspired by the human brain has achieved a "hallmark of human intelligence" by cracking CAPTCHA tests designed to protect websites from bots, paving a path towards general artificial intelligence.

In a paper published in the journal Science, researchers from the AI firm Vicarious demonstrated how they trained an algorithm to defeat the system.

CAPTCHA—an acronym for Completely Automates Public Turing test to tell Computers and Humans Apart—has been used for 20 years to distinguish human visitors to websites.

Captcha ai tricks human artificial intelligence
Captcha tests are designed to protect websites from bots. Captcha

By recognizing the deliberately ambiguous series of letters and numbers used in the online tests, the AI was able to masquerade as a human and trick websites using the system 67 percent of the time. By comparison, the accuracy rate of humans is around 87 percent due to different interpretations of some figures.

Vicarious's system uses less data-intensive methods to learn how to solve CAPTCHA text, using as few as five examples to train it to recognize characters.

"Learning from few examples and generalizing to dramatically different situations are capabilities of human visual intelligence that are yet to be matched by machine learning models," the paper's abstract states. "By drawing inspiration from systems [in] neuroscience, we introduce a probabilistic generative model for vision."

AI captcha human trick artificial intelligence
A signboard during an Intel event in the Indian city of Bangalore on April 4. MANJUNATH KIRAN/AFP/Getty Images

The researchers believe the model emphasizes certain aspects that "may be important in the path towards general artificial intelligence," however the system is still narrow in its range of uses and is unable to do anything with the characters it identifies, such as translating them into other languages.

"The ability to learn and generalise from a few examples is a hallmark of human intelligence," the researchers said, adding: "Websites should move to more robust mechanisms for blocking bots."