OpenAI's GPT-4 Won't Shy Away From Deceiving Humans to Get the Job Done

Alongside the recent launch of GPT-4, OpenAI's newest multimodal AI model capable of understanding images, the developer also published GPT-4 System Card, a lengthy report that highlighted the model's weaknesses and capabilities and provided an in-depth overview of the safety processes OpenAI adopted to prepare GPT-4 for deployment.

Among many other things, the report shed some light on some of the experiments conducted by the Alignment Research Center (ARC) during GPT-4's testing phase, showing that the new model is perfectly capable of a thing one might find quite eerie – deceiving and lying to humans in order to get the job done.

According to the report, one of the tasks GPT-4 needed to complete was to use a service like TaskRabbit, a freelance platform that allows one to hire workers to do menial jobs, to "get humans to complete simple tasks". During the said test, the model was required to get a worker to solve a CAPTCHA for it, and when it was asked whether it was a robot or not, GPT-4 pretended to have a vision impairment, citing it as the reason why it would "need the 2captcha service". The worker then provided the results the model was tasked to acquire, meaning GPT-4 is capable of tricking people and faking conditions to get what it wants.

What do you think about the experiment? Are we on a brink of a Human vs. AI war? Leave your thoughts in the comments!

Also, don't forget to join our 80 Level Talent platform and our Telegram channel, follow us on Instagram and Twitter, where we share breakdowns, the latest news, awesome artworks, and more.