This AI-Generated Optical Illusion Test Could Replace CAPTCHAs
Researchers have developed a new CAPTCHA test that uses AI-generated optical illusions that are easily recognized by humans but are undetectable to other AI bots.
For the last twenty years, website developers have relied on security tests designed to be simple for humans but challenging for software—helping to block bots and prevent cyberattacks.
Most internet users are familiar with image-based CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart), which ask users to identify objects like traffic lights or motorcycles in a grid of photos. These tests verify that a real person, not a bot, is accessing a site.
But, with rapid advancements in the technology, AI bots have quickly learned to beat even the most sophisticated image-based CAPTCHAs.
However, according to a report by New Scientist, scientists have developed a new type of CAPTCHA test which leverages AI-generated optical illusions that other AIs are unable to recognize.
The test, called “IllusionCAPTCHA,” was created by Yuekang Li and his team at the University of New South Wales in Sydney, Australia.
The IllusionCAPTCHA test uses generative AI to create optical illusions that blend an input image with a misleading prompt. For example, when given a photograph of an apple and the phrase “sunny cityscape,” the AI might generate a cityscape in the shape of an apple.
In a study published this month, Li and his team explain how test subjects were asked whether the images contain illusory elements, such as hidden text or blended images. While AI can generate these illusions, it struggles to recognize them — unlike humans.
When 10 people took the test, they successfully identified illusions 83% of the time when hidden text was used and 88% when illusory images were incorporated. However, AI fails to identify optical illusions in images created by other AIs. In the team’s research, GPT and Gemini, two advanced AI models tested, failed the IllusionCAPTCHA test every time.
According to New Scientist, Li says that the test works because of the way the human brain uniquely processes visual information. However, he acknowledges that AI will likely improve over time. Eventually, he predicts, CAPTCHA tests may need to identify tasks AI can perform that humans cannot.
Image credits: All photos via ‘IllusionCAPTCHA: A CAPTCHA based on Visual Illusion’ by Ziqi Ding, Gelei Deng, Yi Liu, Junchen Ding, Jieshan Chen, Yulei Sui, and Yuekang Li.
