- Who's the bunny? (Image: National Geographic/Getty Luxx Images; Rabbit: Joel Sartore)
- mg22429992.400-1_300[1].jpg (21.57 KiB) Viewed 2818 times
GOOGLE updated its captchas on Wednesday. Those fuzzy words and blurred numbers that are simple for humans (in theory) to decipher and hard for bots, are there to guard websites, email services and social networks from automated attack. But they are also fuelling the development of artificial intelligence.
As spammers and their bots get ever better at breaking captchas, Google continues to make them harder – as anyone who has ever failed to decipher a mysterious swirl of lettering can attest. Alternatives, such as audio versions for people with a visual impairment, are less secure and often equally baffling.
The latest revamp fixes some of these problems, in part by doing away with the tests for recognised users. People that Google already knows are human can tick a box to affirm that "I'm not a robot" (see "Do I know you from somewhere?"). Those who aren't automatically recognised can now pick matching images out of a grid – cats in a sea of dogs and hamsters, for instance.
Luis von Ahn, a computer scientist at Carnegie Mellon University in Pittsburgh, Pennsylvania, realised he could put our captcha-solving prowess to good use, and so find a silver lining to the internet's bot problem. Instead of having us merely prove our humanness, Von Ahn designed the system to record and model that humanness and the intelligence behind it.
Von Ahn sold his technology to Google in 2009, and the internet giant has since improved on it, building a software engine that teaches computers to recognise new things. As internet users verify their humanness, solving problems that only human minds can figure out, they build data sets that can be fed to algorithms. As well as reducing spam and blocking bots, Google's system uses the humans it protects to turn the web into training for AIs, with the goal of improving their ability to recognise real-world objects.
A team led by Dumitru Erhan at Google is working on a way to automatically caption pictures, and can already accurately add tags like "A group of people shopping at an outdoor market", for example. Other research teams are on a similar path.
Roman Yampolskiy, a computer scientist at the University of Louisville in Kentucky says the information gathered through captchas has helped machine learning algorithms to reach their current state.
As humans tear through each new generation of captchas, the data they generate can teach machine learning algorithms to solve the same problems, giving computers a handle on more and more of the world. Yampolskiy and his graduate students are working on a system that will allow any set of images on the internet to be used for captchas, enabling humans to train AI to tell different animals apart, or, say, tell cars from washing machines.
For the most part, data from captchas go to feed artificial neural networks, the computing systems that are inspired by the way the brain works.
Captcha's evolution is evidence of the progress AI is making. The first puzzles can now be solved automatically by computers 99.8 per cent of the time, according to Google. Captchas are one of the most prominent human interfaces with this progress. Face recognition and search algorithms are improving behind the scenes, but mostly we are all stuck solving those annoying little tests, man and bot.
This article appeared in print under the headline "I am not robot"
DO I know you from somewhere?
How does Google know when to give real people a free pass on to a site, without cat categorising (see main story)? When it recognises them.
Google has what is probably the largest store of personal data on the planet. So it is fairly simple for the company to match its records up with incoming IP addresses and cookies – the small piece of code that websites use to keep track of their visitors. So they can be pretty sure that the person trying to pass the captcha is human.
This has sparked some fears that Google is making it harder for non-Google users to prove their humanness around the rest of the web, as many services also use Google's captchas.