Artificial intelligence programs can create optical illusions that other AIs are unable to recognise, creating a useful CAPTCHA test to differentiate humans from bots.
A cat-and-mouse game has played out聽for almost two decades between website developers who want to keep bots out of their sites and the hackers who want to bypass those protections. Websites have long deployed tests that are designed to be easy for humans to pass, but that trip up software.
Over time, these CAPTCHA tests 鈥� which stands for Completely Automated Public to tell Computers and Humans Apart 鈥� have become more advanced and so gradually trickier to crack. But the fast pace of AI development means that bots have quickly gained the ability to solve any new test that gets deployed.
Advertisement
Now, at the University of New South Wales in Sydney, Australia, and his colleagues have developed a new test they call IllusionCAPTCHA. The test uses generative models to create optical illusions that combine an input image and a prompt. For example, given a photograph of an apple and the prompt 鈥渟unny cityscape鈥�, the AI might create an image of a city in the shape of an apple.
Test subjects are shown these images and asked if they feature an illusory aspect, such as combined images or hidden text. While AIs can easily make these images, they are unable to detect such illusions when they examine these images.
When 10 people attempted the test, they passed 83 per cent of the time when illusionary text was added to an image, and 88 per cent of the time when illusory images were incorporated. But neither GPT nor Gemini, the two AI programs assessed, successfully passed any tests.
Li says the test relies on the unique way the human brain processes visual information. The gap between what AI and humans are capable of is useful to inform researchers looking to improve AI, he says. 鈥淲e鈥檙e trying to make AIs closer to human, closer and closer, and the more similar to humans they are, the harder it is for us to differentiate,鈥� says Li.
He expects AI to eventually become more capable than humans even at visual tests, which will force CAPTCHA tests to look for things that AI can do but humans cannot. However, AI will then learn to pretend to be unable to do these tasks, and the cat-and-mouse game will continue. 鈥淚 think it鈥檚 going to be a problem forever,鈥� he says.
arXiv
