AI Takes Human Approach To Beat Captchas

Artificial intelligence bots have reclaimed the advantage in their battle with Captcha challenges – and the scary part is they’ve done it by aping what was thought to be humans’ advantage.

The Captcha challenges in question are the ones that involve identifying a string of characters despite distortions such as wavy text or cluttered backgrounds. These were developed after previous generations of Captcha were ‘beaten’ by the bots.

Now Vicarious, a AI company founded by Jeff Bezos and Mark Zuckerberg, says its developed software that can beat most popular Captcha systems the majority of the time, including a two-thirds success rate with Google’s ReCaptcha.

The system, Recursive Cortical Network, has a success rate of around 90 percent when it comes to individual characters. The lower rate for the entire challenge is because the Captcha’s use random strings rather than words, so it’s harder to guess unclear characters through context.

The key to the system is directly going after what’s meant to make Captcha’s work. Traditionally computers are great at analysing individual data points, but less skilled at spotting patterns, particularly in visual images. Meanwhile humans would take far longer to analyse every pixel in an image, but don’t need to with a Captcha because they can spot what’s important and relevant with a mere glance.

Recursive Cortical Network mirrors this ability and human experience by breaking down characters into particular features and then how those features combine into shapes. By looking for the combination of features, it doesn’t matter as much if the letters are rotated, reversed or partially obscured.

The big limitation of this for the wider field of artificial intelligence is that the system doesn’t actually understand any ‘meaning’ of the characters, instead being designed solely to solve a specific puzzle.