New AI algorithms beat CAPTCHA tests 90% of the time
San Francisco-based artificial intelligence startup Vicarious has announced that it has developed software algorithms which can solve CAPTCHAs up to 90 percent of the time. Though CAPTCHAs are any automated test which differentiates between humans and computers, they often take the form of strings of partly distorted letters and numbers which many websites use to check that a visitor is human, the idea being that a computer cannot read the disguised text while a human (hopefully) can. It's this type of CAPTCHA that Vicarious's algorithms are designed to beat, and the high success rate renders the current standard of text-based CAPTCHAs ineffective, the company claims.
The main purpose of a CAPTCHA, which stands for Completely Automated Public Turing test to tell Computers and Humans Apart, is to prevent automated software bots accessing web services where they might get up to all sorts of nuisance, such as signing up for spam accounts, or adding or editing information in a directory or wiki.
Vicarious isn't the first to develop CAPTCHA-beating algorithms, but the company claims that its system, which uses its Recursive Cortical Network technology, can beat current text-based CAPTCHAS up to 90 percent of the time, including those developed or used by Captcha.com, Google, Yahoo and Paypal, defeating their very purpose in the process.
To put the success rate into perspective, Vicarious cites a 2011 paper out of Stanford, which deemed a CAPTCHA system "broken when the attacker is able to reach a precision of at least 1 percent."
Vicarious talks only in vague terms as to how the algorithms solve the problem, but it does claim that its system uses "relatively minuscule" amounts of data and computing power compared to other modern AI systems such as IBM's Watson, operating at "a level of effectiveness and efficiency much closer to actual human brains."
Commercial applications of the Recursive Cortical Network are still years away, according to Vicarious, and so it would be premature to conclude that CAPTCHA systems have been rendered obsolete over night. Rather, the remarkably high success right would seem to mark this out as a significant step in the field of AI.
Please keep comments to less than 150 words. No abusive material or spam will be published.
HA! I just HAD to pretend my software had beaten gizmag's captcha.
As they saying goes, "artificial intelligence is no match for human stupidity".
It's a lot better at CAPTCHA than I am...
Haha, me too. Delicious irony.
I think this has good potential to be applied to OCR tech.
With regards to the accuracy, most websites will give you several attempts before lockout, so if it is 90% first go, it will likely defeat most systems after about three goes.
And if Captcha tries to be more tricky by adding multi coloured lines or placing the text sideways, this kind of algorithm would likely be able to be updated in time to accommodate.
The only way I can think this algorithm would get unstuck is if the letters are dashed, because the vectoring algorithm may get confused about the flow of the line, where as a human brain might fill in the gaps.
In addition, the visitor is asked a question about the form data on the page that appears when he/she clicks the confirmation link in the email. If his/her answer does not match the previously submitted form data that data is deleted.
Visitors always know what data they entered into the form, but bots wouldn't know what information the page is requesting even if they were to click the link in the confirmation email -- which they also never do because bots do not receive email.
Bottom line: There have always been more effective ways to prevent bots than by using captcha, and this is one of them. My clients will never have to worry about this new captcha-reading software.
Sincerely, owkaye at gmail dot com