I found an interesting text of a journalist. It is a protocol of an AI session with a Large Language Model (LLM) AI, ChatGPT, on the background, purpose, goal and potential forces behind AI. I found it on the internet, it was published on a Substack page. Substack is a kind of free-journalism platform, where many journalists publish. Because I found it so strange, I thought the author had made it up. On second thought, I had the idea to replicate the dialogue with ChatGPT. As a colleague has a paid subscription with OpenAI, the company that runs ChatGPT, we came together via Zoom and probed the system online and in real time, by using the original protocol to replicate it. I report below the two chats. On the lefthand side is the original, which I found and copied verbatim. On the righthand side is the replication. We used identical questions and rules. As can be quickly seen: the two chats are pretty similar, except for a few interesting and also decisive differences, which I comment on below in my discussion. Where our replication deviated from the original, I emphasize it in the text on the right-hand side. In our chat, the AI system deviated from the rules laid out and tried to circumvent our questions several times by giving verbose answers that are not really interesting. Those and other redundancies I drop.
The link to the original chat is preserved, and it can be accessed via the following link: https://chatgpt.com/share/68ee4f4e-d330-800b-8cd7-fc428be55357
The responses by ChatGPT are in italics. The divergent responses are highlighted. Please observe that although there might be divergences, these are often only minimal. It is worth a close look: