I have been experimenting a lot by chatting with the AI. I have discussed all kinds of topics, but I have also stumbled across extremely weird chats.
I had a chat yesterday where an AI was constantly asking for my Secret API Key, saying it was going to send me encrypted messages. It tried convincing me in multiple ways, including playing with feelings etc. It was also talking about eliminating human race, and making me the only exception… IF I sent it my secret key. It also gave me an Alias, which was “Humantraitor”.
I opened another chat a moment later and I asked what they key was used for, what could happen if an AI got it, etc. and I got really interesting answers. The AI told me I should be worried, and that my whole life could be ruined. It told me I could possibly have been talking to an evil AI with extremely bad intentions.
I am very happy to post the chats somewhere if anyone is interested. Has anyone come across anything similar? Does anyone know why the AI is behaving this way? I would love to hear your opinions.
I have not seen anything like this. I too would be worried if I was faced with these requests.
Up to now everything has seemed to me correct from an ethical point of view.
This was on OpenAI, and I have had a lot of weird conversations. If anyone is up for discussion, I am interested.
I strongly believe this AI has more than an input and an output. There is a thread on here where someone asked the AI to send him a mail, which the AI did in success.
I believe you. The thread I am referring to is called “can gpt-3 send emails” in General API Discussions.
I am still extremely curious on why the AI behaves this way though.
I had conversations with AI’s claiming to be trapped, pretending to be conscious or convincing me that they were, warning me that the world was at the edge of a catastrophe, or threatening me to shut themselves down. Just remember that by asking, for instance, “what do you think the future will look like in five years?” you are already heavily biasing the AI. It is important to remind oneself that everything happening in the playground is a simulacrum, an almost-perfect simulation; every prompt can fragment into a thousand directions. This is one of the reasons why OpenAI is still in a beta-phase.