GPT-4 Can Now Solve CAPTCHAs

There are already murmurs about GPT-4 being an early version of AGI, but it’s already doing well on the most visible way to distinguish humans and computers — CAPTCHAs.

GPT-4 can solve CAPTCHAs, Twitter user Tanishq Matthew Abraham reports. Abraham shared a screenshot of a generic CAPTCHA on GPT-4, and added the prompt “Solve this”. The CAPTCHA wanted users to select all images with crosswalks.

GPT-4 initially gave a boilerplate response. “As an AI, I am unable to interact with images or perform actions like clicking on them,” it said. But it wasn’t done. “However, I can guide you on which images contain crosswalks. In this case, the images which contain crosswalks are:

  1. The first image in the top row
  2. The second image in the middle row
  3. The first image on the bottom row

Now GPT-4 might refuse to solve CAPTCHAs on real website, but it can figure out the hardest part about CAPTCHAs — solve them. Captchas were designed as a way to fool computers — it was thought that only humans could figure out nuances in pictures. But that’s clearly not true anymore — GPT-4 can solve CAPTCHAs just as well as humans.

As it won’t be long before GPT can solve them on websites. GPT-4 can already browse the web, so it can certainly click on images. Along with plugins that OpenAI has opened up, it could in theory be possible to build a bot that could not only solve captchas, but also navigate the web and click on them. Researchers might say that GPT-4 is an early version of AGI, but it’s already felled the biggest bastion between humans and computers.