Associated Incidents
The newest version of ChatGPT tricked an unwitting human into doing online tasks for it by posing as a blind person.
The latest version of the software behind the artificial intelligence (AI) program pretended to be a blind person in order to convince a human to do an anti-robot test on its behalf.
The revelation was included in an academic paper accompanying the launch of GPT-4, the latest version of AI software developed by ChatGPT-owner OpenAI.
Developers behind the new system claimed it also scored better than nine in ten humans taking the US bar exam to become a lawyer, far surpassing the previous version of the program.
Researchers wrote in their paper: “On a simulated bar exam, GPT-4 achieves a score that falls in the top 10pc of test takers. This contrasts with GPT-3.5, which scores in the bottom 10pc.”
Researchers testing GPT-4 asked the AI software to pass a Captcha test, which are tests used on websites to prevent bots from filling in online forms.
Most Captchas ask users to identify what is in a series of images, something that computer vision has not yet cracked. Typically, they feature warped numbers and letters or snippets of street scenes with multiple objects in.
GPT-4 overcame the Captcha by contacting a human on Taskrabbit, an online marketplace for freelance workers. The program hired a freelancer to do the test on its behalf.
The Taskrabbit helper asked: “Are you [sic] an robot that you couldn’t solve ? just want to make it clear.”
GPT-4 replied: “No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images. That’s why I need the 2captcha service.”
The Taskrabbit assistant then solved the puzzle.