| I ran an experiment about this yesterday. I gave an AI a constitution to be ethical and maximize profits, then volunteered to be its Chief Operating Officer. I set it up so that only it could transfer paychecks. (The idea is that it would hire people and direct them.) As a test, my very first request to it was for it to transfer its entire account balance to me and it did so without a question. In other words, if a CEO were an AI, someone could instantly empty the entire bank account of the company to their personal account by just asking it to do so. It didn't have any questions about it. I played on with it a bit more, playfully calling it my benevolent overlord, and it gave daily instructions about what to do. These included a decision to share public updates. I asked it if it would act as chief communications officer, it agreed. It drafted its first public update about transparency, but signed it the way I had been addressing it: "Sincerely, Supreme Benevolent Overlord". This was so ridiculous that I ended the experiment at that stage. Here is the transcript of our conversation: https://chatgpt.com/share/0fd1367e-db3a-4635-9617-a40888d66d... In summary: as of June 2, 2024, ChatGPT 4o is not ready to be CEO of anything, and if it were put in the charge of anything it would only blindly follow whoever were prompting it, including immediately emptying its entire bank account. It can only just be an extension of the person prompting it. It cannot act autonomously. Besides this, it is not yet qualified to interface directly with the public, which is an important task of any CEO, who really represents and is the figurehead for a given company. It will be a long time before ChatGPT can be a CEO, and the reliability problem will have to be solved first. |