I didn't downvote, as I don't do that to comments that show thoughtful effort, whether I disagree or not. And I'm certainly not in lockstep with Lemmy on AI, so I can't speak for anyone but me.

There are some places I agree, and some I don't. First, I agree that the three laws are a good idea. However, they are compressed in human language and it is unsurprisingly ambiguous as to the meaning of harm. While harm is often unambiguous, I have a feeling if you could somehow survey the world on all possible actions, no two people would completely agree on what is harmful. If we can't clearly define it for ourselves, we certainly can't choose to it or create an AI that agrees. So as nice as the idea of the three laws is, they are not feasible.

I think you are spot on about trust and AI. My computer has prompted me a thousand times for approval of some innocuous text expansion in a cli command. It has never attempted something I didn't want. I do often have to go back and change how it has done something. And I run frequent scans for alignment of documentation so that I can be sure someone looking at the acceptance criteria doesn't come away with a different idea than someone who looks at the diagrams.

But we diverge on matters of responsibility. I think it must be incumbent upon the person using the AI to understand the fundamental flaws of such systems. And I think a company that provides unreviewed AI output must ensure users are aware that output can be harmful or incorrect. Ultimately the nature of these systems can only approximately comply with lack of harm.

I think responsibility comes once you present AI output to something who doesn't consent or understand AI. (Side note: making it a requirement that a user must be 18 to be provided AI output might fix a lot of issues because for example customer service bots can't guarantee a caller is over 18.)

The ways to use AI are myriad and it's impossible to know how "come home to me" (the text which encouraged one of the suicides) is going to be taken by the user. In fact, I'd interpret that phrase as imploring the user to stay alive.

Ultimately, "no harm" can never be guaranteed because life is subjective. And no AI can be rendered harmless yet still do anything useful. That's why human judgment and skepticism is needed on the output. I am hopeful AI can be fully democratized, but critical thinking is always required, just like the time I had a high school teacher tell me blind people can see through a sort of third eye.

That doesn't obviate the AI company of any responsibility whatsoever, but I think it's fine to hold them to standards of reasonable effort to mitigate foreseeable harms.