House numbers on Google Street View can turn up as blobby, blurry things, so its engineers built a pretty crazy neural network to decipher them. Except this algorithm also turns out to be very very good at deciphering other blobby, blurry texts — like CAPTCHAs, which it cracks with 99 per cent accuracy. Take that, human.
Earlier this year, Google Street View engineers published a paper describing a neural network, modelled off of animal nervous systems, that was pretty good at identifying house numbers. “We can, for example, transcribe all the views we have of street numbers in France in less than an hour using our Google infrastructure,” wrote one of the engineers in the paper. It got the numbers right with 90 per cent accuracy.
But give it a small, black-and-white CAPTCHA with none of the lightning or colour variables, and the neural network does even better. That includes those nearly inscrutable CAPTCHAs at the top of this post.
What this all means, really, is that blobby, blurry text is not the defining test between human and bot. Last year, Google actually decided to make CAPTCHAs less distorted and easier to decipher for humans. Its system takes into account other factors like how you interact with the page in what it calls “advanced risk analysis,” the specifics of which they very purposefully do not explain. Googles says not to worry about CAPTCHAs being cracked — for now anyway. [Google Online Security Blog]