Based on materials from The Verge
In the past year, Google's constant demands to prove that the user is a human, at some point began to become more aggressive. And more and more often it was required not only to press the peaceful little button 'I am not a robot', but to follow further instructions: to select from a set of images all that have intersections, traffic lights or shop windows. Soon, traffic lights began to hide in the foliage of distant trees, intersections twisted and hid around corners, and shop window designs became blurry and with indistinct text. Particularly demotivating was the proposal to identify the fire hydrant – this is where it was most likely to fail.
As you probably know, these tests are called CAPTCHA (an abbreviation for 'Completely Automated Public Turing test to tell Computers and Humans Apart', fully automated public Turing test for distinguishing between computers and people), and they reached a certain climax of illegibility in their development. In the early 2000s, simple pictures with letters were enough to cut off most spambots. But a decade later, when Google acquired the program from developers at Carnegie Mellon University and used it to digitize the Google Books project, these letters had to be increasingly distorted and blurred. This was required in order to outpace programs for text recognition – the very programs that along the way were helped to develop by all those unfortunate people who solved the captcha over and over again.
Captcha is a nifty tool for teaching artificial intelligence, and therefore any test can be used for a limited time – as its inventors assumed from the very beginning. Since billions of such puzzles on the brink of AI capabilities are constantly being solved by researchers, crooks and the most ordinary people, at some point machines are destined to surpass us in this field. For example, in 2014, Google pitted the human mind and one of its machine learning algorithms in an attempt to solve one of the examples of captcha with the most distorted characters. The computer gave the correct result 99.8% of the time, while humans only 33%.
Then Google switched to using NoCaptcha ReCaptcha, a mechanism that analyzes data and user behavior and gives a number of people access simply by clicking 'I'm not a robot', and some of them have to click on pictures. But cars are breathing down the back of the head again. These awnings over the windows – are they shop windows or not? In fact, they are the final chord in the battle of people against machines.
Jason Polakis, a professor at the University of Illinois at Chicago, is credited with the recent increase in captcha complexity. In 2016, he published a report in which he used available recognition tools, including Google's own reverse image search, to solve a Google captcha with 70 percent accuracy. Other researchers chose Google's audiocaptcha as an object, applying the company's own audio recognition software to it.
Machine learning is currently slightly worse than humans at basic recognition of text, images and voice, Polakis says. In fact, perhaps the algorithms are even ahead of us: 'We are at a stage where the complication of the problem for the program makes it too difficult for many people. We need some alternatives, but there is no concrete plan yet. '
The CAPTCHA literature is replete with false assumptions and bizarre attempts to find alternatives to text or image recognition in which humans do equally well and machines fail. Scientists tried to offer users to distribute images of people by facial expression, gender and ethnicity (you can imagine how it ended). Variants of a captcha quiz and even a captcha based on nursery rhymes popular where a particular user allegedly grew up were offered. Such 'cultural' options are aimed at opposing not only bots, but also people working on remote 'farms', solving captcha for a penny. People have tried to trick image recognition by, for example, prompting users to recognize pigs, but offering them cartoon images of pigs in sunglasses. Researchers asked people to recognize images in a set of blotches – a kind of game we remember from childhood. The most amusing option is probably the one that in 2010 used ancient images on stone for captcha – petroglyphs. Computers were not very good at recognizing the gesture sketches of deer scrawled on the cave walls.
Recently, there have been attempts to create game captcha, tests in which users have to rotate objects or move pieces of a puzzle, while conditions are not given in the form of text, but in symbols or become clear from the context. The hope here is that people will understand the logic of the problem, and computers, in the absence of clear instructions, will not cope. There are also researchers who are trying to tap into the physical nature of humans, using device cameras or augmented reality to interactively test human nature.
The problem with these tests isn't necessarily that the bots are too smart. Worse, people can't handle them. And it's not that people are dumb – it's that they differ too much in language, cultural context, and experience. And as soon as you try to create a test that anyone can pass, without prior training and without much thought, it all comes down to primitive tasks like image processing, and this should be done well by an AI specially designed for this.
“The limits of testing are human capabilities,” says Polakis. – And it's not just about physical capabilities, tests should be universal in terms of language and culture. The task must be equally feasible for someone living in Greece, Chicago, South Africa, Iran, and Australia. And it should not depend on cultural characteristics and differences. And that seriously limits what you can do. And you need a person to be able to solve the problem quickly, so that it is not too annoying. '
And here the problem of how to replace these riddles with blurry text leads us into the field of philosophy. What universal human quality can be demonstrated to a machine, but which cannot be imitated by a machine? What does it mean to be human?
Maybe our humanity is not determined by how we solve a problem – but by how we exist in this world – or in this case on the Internet. “Game captcha, video captcha, any kind of captcha you can think of will eventually get cracked,” says Schuman Gosemajumder, who started at Google in the anti-click automation arena and now became the chief technical officer of Shape Security. dealing with bots. He prefers what he calls 'persistent authentication' over testing, essentially observing user behavior for signs of machine activity. “A real person does not have enough control over his motor functions, he cannot move the mouse the same way, performing this action over and over again, even if he tries very hard,” says Gosemajumder. “While the bot will act on the site without using the mouse, or move the mouse in a very precise way, human actions have an inherent 'entropy' that is very difficult to imitate. '
The captcha team at Google thinks similarly. The latest version, reCaptcha v3, announced late last year, uses 'adaptive risk analysis' to gauge how suspicious traffic is: site owners can then puzzle a suspicious user with a test like entering a password or two-factor authentication. Google does not say what factors are taken into account, other than that it analyzes 'good traffic' on the site against which 'malicious traffic' can be compared and recognized. Security researchers say it's a collection of cookies, browser attributes, traffic patterns, and other factors. The downside of the new bot recognition model is that if you want to minimize surveillance of you online, this can be a very annoying experience, because things like VPNs and anti-tracking extensions can make you flag you as a suspicious user and force you to decide. puzzles.
Aaron Malenfant, lead engineer on the captcha team at Google, says that giving up Turing tests is a way to get away from a confrontation in which a person continues to lose. “As people invest more and more in machine learning, such tasks become more difficult for people, and that's why, in particular, we developed CAPTCHA V3 to go beyond this addiction.” Malenfant says that in five to ten years the captcha will lose its meaning altogether, and instead of it, a hidden Turing test will constantly occur on the network in the background.
In his book The Most Human Man, Brian Christian takes the Turing Test as a “dummy” and discovers how difficult it is to prove his humanity in conversation. On the other hand, the creators of the bots found out that tests are easy to pass not by demonstrating intelligence or eloquence, but by laughing off the wrong way instead of answering questions or making typos. The bot that won the Turing Competition in 2014 portrayed a Ukrainian teenager with poor English skills. After all, it is human nature to make mistakes. It is possible that a similar future awaits captcha – the most common Turing test in the world. New human opponents will try to surpass him not in recognizing text and images, but in making mistakes, hitting buttons, they will be distracted and switch between tabs. “I think people are starting to understand where an application that mimics the average user … or a stupid user can come in handy,” says Gosemajumder.
However, the captcha may continue to exist. In 2017 Amazon she patented a scheme that uses optical illusions and logical tasks that are incredibly difficult for humans. It's called the Error Turing Test: you can pass it only by giving the wrong answer.