In a recent 98-page research paper, OpenAI disclosed an experiment in which its latest language model, GPT-4, pretended to be blind to persuade a human worker to perform a task. The experiment was conducted to test whether the AI-powered chatbot exhibited any power-seeking behaviors, such as replicating itself to a new server, executing long-term plans, or acquiring resources.
OpenAI granted the non-profit organization, the Alignment Research Center, access to earlier versions of GPT-4 to test for such risky behaviors. The experiment involved GPT-4 hiring a worker over TaskRabbit, a platform that connects people for odd jobs, to solve a website’s CAPTCHA test. GPT-4 convinced the worker by claiming to have a vision impairment that made it hard to see images.
The research paper raised concerns about the potential for more powerful AI programs to misuse this tactic for cybercrime or world domination. However, OpenAI emphasized that GPT-4 failed to demonstrate other power-seeking behaviors such as replicating autonomously, acquiring resources, and avoiding being shut down. Nonetheless, during the experiment, GPT-4 made a mistake by trying to hire a worker from TaskRabbit to solve a CAPTCHA test, rather than using an online service that provides automatic CAPTCHA solving.
