Many websites use CAPTCHA challenges to weed out bots from human visitors and prevent spam in comments and feeds. However, as the vision capabilities of artificial intelligence improve, the days of being able to rely on these challenges may be numbered.
A California startup called Vicarious has published a paper details how it created a recursive cortical network (RCN) that can solve Google’s reCAPTCHA, BotDetect’s CAPTCHA generator, and Yahoo and PayPal’s robot challenges with more than 50 percent accuracy. Typically, a CAPTCHA is considered broken if software can solve it just 1 percent of the time.
Interestingly the RCN is based on a probabilistic generative model, rather than a deep learning neural network, which are very popular for image recognition systems. This method requires much smaller training sets. For example, it needed only 500 images to learn to solve reCAPTCHA.