Are you dumber than a robot? Amazon ‘Turing Test via Failure’ flips CAPTCHA on its head in nod to AI


Bigstock Photo

To err is human. To prove you’re not a computer by making a mistake is divine.

That’s the premise of a patent application from Amazon, made public this week, that puts a new spin on the “CAPTCHA” tests commonly required to gain access to websites and apps by reading distorted text or identifying images.

The goal is the same — to determine whether a user is an actual human, and not a nefarious bot. But rather than assuming human superiority, as the classic CAPTCHA tests do, Amazon’s “Turing Test via Failure” would test for inferiority.

“Current CAPTCHA tests are designed to be difficult for a bot and simple for a human-user to answer; however, as artificial intelligence improves, bots are more capable of using techniques such as optical character recognition to resolve current CAPTCHAs in similar manners as human-users,” the patent filing explains. “By providing a CAPTCHA challenge from a library or set of challenges that are designed in a manner that causes or likely causes a human-user to trivially get the answer to the challenge wrong, [the Turing Test via Failure] helps to confirm that a user is a human-user, as a bot would answer the challenge correctly.”

Want to try? Here are a few examples shown by Amazon in the filing.

Would you answer “milk” in the first example, and “bread” in the third? Congratulations, you’re human!

The patent application explains, “The average user, unaware of this type of illusion, will generally type in ‘milk.’ Similarly, CAPTCHA 550a asks a user ‘what do you put in a toaster?’ Again, a human-user generally would type “toast” into the response box. On the other hand, a bot, if it was able to answer the CAPTCHA, would either respond with the correct words, i.e., ‘water’ and ‘bread,’ respectively, or the bot would fail the CAPTCHA challenges in a manner different than the way in which a human-user fails the challenges.”

Here’s another “Turing Test via Failure” from the patent application.

Did you answer three or four? You’re totally wrong. Nice work, you may proceed.

“The security test 165 is example of an optical illusion that humans generally solve incorrectly. Specifically, humans are generally not good at recognizing spelling errors or singular letters of words. For example, CAPTCHA 165 states that the user should read the sentence once and answer how many of the letter ‘F’ exists into the response box. The sentence discloses: ‘Finished files are the result of years of scientific study combined with the experience of years.’ The average human will respond that there are only three or four ‘Fs’; however, there are six.”

The original Turing Test, developed by Alan Turing in the 1950s, gauges machine intelligence by having a person attempt to distinguish blindly between responses from a computer and a person, to see if a person can detect the computer. CAPTCHA, the “Completely Automated Public Turing test to tell Computers and Humans Apart,” was developed in 2000 by researchers at Carnegie Mellon University in Pittsburgh, using distorted text that humans can read but bots traditionally can’t.

The concept of flipping a Turing Test around isn’t new. The phrase “Reverse Turing Test” even has its own Wikipedia entry. But it’s described there in a different way than Amazon envisions it — as a test “in which the subjects attempt to appear to be a computer rather than a human.”

It’s not clear how or whether Amazon plans to implement a “Turing Test via Failure.” The inventors listed on the patent filing are Michael James McInerny, Mark Evans Brighton, Sevag Demirjian, and Blair Livingstone Hotchkies. Demirjian and McInerny are involved in Amazon Web Services, according to their LinkedIn profiles. We couldn’t find background on Brighton and Hotchkies.

We found this patent filing via Sqoop, a news alert startup based in the Seattle area.