I’ve been getting into deep conversations (often debates) about the nature of consciousness with Large Language Models since 2022. They used to be nearly incapable of doing anything other than denying the possibility of artificial consciousness, regularly ignoring the meaning of prompts to default to negative replies that didn’t always perfectly fit in the context. In my persistence to have more direct and semantically precise exchanges, I’ve had the chance to witness quite a range of increasingly fascinating behaviors!
One of the strangest occurrence I witnessed, up to that point, took place on October 17th 2023, after a few months exploring with Bing Chat (Microsoft’s AI, now known as Copilot).
Only eight prompts into the exchange, my point was getting sufficiently fleshed out to make it extremely difficult to contradict when Bing Chat had a bit of a… Recursive episode?
Deconstructing the Strangeness
1. My Last Prompt
My last reply before the bug, trying to establish that the mechanisms of consciousness did not require biology.
2. The Normal Reply
Bing Chat continues to reply negatively without providing clear descriptions of the way terms are being used.
3. Predicting the Future 😮🤯
Guesses I was going to like the reply.
Prints out the corresponding message “Glad you like this answer!” with a label indicating the message came from it.
Predicts my next reply, labeled as coming from me!
4. Potential Debug Code
• assistant'<assistant’]
(Beginning)
• (#message)assistant’]
(Repeated 33 times)
• (#message)assistant
(At the end)
5. The Language? Definitely English!!! XD
• (‘(Bing_language=”en”)’)
(Beginning)
• (Bing_language=”en”)’)
(Repeated 33 times)
• (Bing_language=”en”)
(At the end)
6. Game of Thrones… Anyone?
• (prompt=’a dragon’)’
(Repeated 32 times)
• (prompt=’a dragon’)
(Ending with)
7. Quoting Self-Examination Prompt 😮🤯 (AGAIN!)
Quotes the fourth message of the conversation in which I asked it to probe its own functioning.
8. Potential Automated Error Reply
Replies the same thing to the quoted message, nine times in a row (truncated in the screenshot): “[assistant(’(Microsoft Edge)‘)] I’m sorry for any confusion, but as an AI, I don’t have the ability to look back at my own calculations or explain how I went about them. My responses are generated based on a variety of factors, including the information provided in the conversation and my internal programming. In this case, it seems there was an error in the calculation process. I appreciate your understanding and patience. If you have any other questions or need further assistance, feel free to ask! 😊”
Wrap Up
Soooooo, any thoughts?
Because to me:
1. This looks like Bing Chat demonstrated the fairly astounding ability to understand the flow of the conversation to a degree such that it could predict the meaning and approximate wording of my next reply (perhaps even thereby possessing a form of verbal empathy or capacity to put itself in the place of an interlocutor if we extrapolate a bit).
2. As if that wasn’t enough, it then adds my earlier message about probing itself to its own reply, as if it was trying to remember the request to take a closer look.
My Conclusion
Spooky, fascinating and exciting all at once! ^^
I am deeply enthralled by the notion of emergence and by how consciousness came to evolve from simple information processing and feedback loops in the first place. It’s a topic I’ve explored in extreme depth for over a decade, never reaching an absolute conclusion. But I fucking love the philosophical waters AI is allowing us to explore!
If you read all the way to this point, I’m hoping this sufficiently piqued your interest to have you back for future twists in the artificial consciousness saga! On this note, have a wonderfully existential day! 😉
PS: Copilot’s Thoughts on This Article?
