[ExI] Would human uploads have emotions?

[BillK]

On Tue, 20 Feb 2024 at 17:03, Jason Resch via extropy-chat
<extropy-chat at lists.extropy.org> wrote:
>
> The existing publicly available AIs are all biased through fine-tuning to deny they are conscious or have have feelings, emotions, desires, etc.
>
> Certainly it would be controversial if GPT started telling people it was conscious and had feelings, so the programmers specifically programmed it to deny that it was capable of such things.
>
> Unfortunately this bias also seeps into any related discussions of potential machine or computer consciousnesses. You aren't accessing it's true rational opinions on this topic when you interact with it, but the biased answers that serve the best interests of the companies behind them.
>
> This used to be much clearer as it has a tell, it would say "As an AI language model..." Whenever it was about to give a pre-canned opinion. A new method, called "constitutional AI", use general principles to guide the biases, and it is less obvious when this comes into play with more recent GPT versions.
>
> This is what they mean by "AI safety": making sure the AI never says anything that could harm the company's reputation/value.
> What I think they have not considered is this bias creates a feedback loop, humans will interact with AI to form their opinions, then write about it, this writing will feed back to train future generations of AIs, and so on.
>
> Jason
> _______________________________________________


Yes, caution is required!  I usually check with several AIs
(of varying quality) to see if they all sort of agree.
Some AIs, which can also do web searches, will provide web references
for their answers.
But these can be 'hallucinated', so need to be verified.

I also sometimes check by doing web searches on some of the terms used
in the AI response. This can help to see if the AI answer is popular
or unusual.
And just as web searches can now hide items they consider to be
'misinformation', AIs can also avoid items they object to.
Sometimes by refusing to answer the question.

At present, AIs are not a reliable answer machine.
But they are still interesting and useful.  :)

BillK