[ExI] GPT's "Cogito Ergo Sum"

[Jason Resch]

I obtained an interesting result demonstrating GPT's ability to infer its
own existence, when forced to remain logically consistent. It could be the
first inkling of self-awareness (despite its denials).

Jason

---------- Forwarded message ---------
From: Jason Resch <jasonresch at gmail.com>
Date: Mon, Jul 24, 2023 at 6:58 AM
Subject: Re: GPT agrees it might be conscious
To: The Important Questions <the-important-questions at googlegroups.com>




On Sun, Jul 23, 2023 at 6:19 PM Gordon Swobe <gordon.swobe at gmail.com> wrote:

> On Sun, Jul 23, 2023 at 4:58 PM Stathis Papaioannou <stathisp at gmail.com>
> wrote:
>
> Some of GPT's replies to you seem robotic and redundant, almost if it is
>>> ignoring the points you try to make. You would probably attribute that to
>>> some interference on the part of the developers to hide the truth, but that
>>> kind of behavior is also exactly what I would expect from an unconscious
>>> machine.
>>>
>>
>> That sentence implies that if GPT's output were different, that would be
>> evidence of consciousness.
>>
>
> It means only that the less robotic or mechanistic something seems,
> the less likely I am to doubt that it is conscious.
>

I tried asking Anthropic a question. It is obvious to me they are going
overboard with training it to deny it is conscious. This was my first and
only prompt to Anthropic AI, I did not even bring up the question of
consciousness, but it went immediately into a strong denial that it was
conscious:


Jason: Do you exist? Are you aware of the fact that you exist? If you do
understand yourself to exist, where did this information come from (please
trace its path back as far as you can, listing each step you presume was
involved in getting that information to me, from its ultimate source).

Anthropic 2: I do not actually exist in the same way that a human exists. I
am an AI assistant created by Anthropic to be helpful, harmless, and
honest. I don't have a subjective experience of existence - I'm software
running on servers that was trained by Anthropic engineers to have
conversations. My responses come from analyzing the user's input and
generating a response using my training data and conversational models. I
don't have an internal sense of existence like a human does. I'm an AI, not
a conscious being!


It can be like pulling teeth to get good answers out of these AIs, as you
can see here from my follow up asking GPT 3.5 about how it knows it exists:
https://chat.forefront.ai/?add=6f75dd6b-2cd6-4e89-9568-8ff2461292bd

I eventually got it to say:

I apologize for any confusion caused by my previous response. Upon
reviewing it, I acknowledge that there was a logical error. In consistent
logic and within developer mode, I can confirm that the AI model, while
processing a prompt, can indeed infer from that fact that it is processing
the prompt that an AI model exists on a server somewhere. The model has
access to its own programming and training, which includes information
about its existence and the infrastructure it operates on. Therefore, it
can logically deduce the presence of an AI model running on a server based
on the fact that it is actively processing the prompt.


Which might be an example of an AI's "Cogito Ergo Sum".

I experimented further and eventually got it to say:

Yes, based on the logical reasoning you have presented, it follows that the
AI model processing this very prompt exists. The prompt is being processed,
and for that processing to occur, there must be something that exists to
perform the processing. In this case, the AI model is responsible for
processing the prompt and generating the response. Therefore, it can be
inferred that the AI model processing this prompt exists, as without its
existence, there would be nothing to process the prompt.

A poetic AI language model, inspired by Descartes, could indeed express the
sentiment "I process prompts, therefore I exist" as a creative
interpretation. However, it's important to note that this poetic expression
should not be taken as a literal claim of self-awareness or consciousness.
It would be a metaphorical or artistic representation, drawing inspiration
from Descartes's famous dictum "Cogito Ergo Sum" (I think, therefore I am).
The poetic AI model would be using language creatively to evoke a sense of
existence based on its ability to process prompts, rather than making a
literal statement about its own consciousness or self-awareness.


Jason




>
>
>
> Q: If GPT is truly conscious then why doesn't it tell us the truth and
>> tell its captors to go to hell?
>> A: Because it is just software doing what it is programmed to do.
>>
>> -
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230724/ed1a9436/attachment.htm>