Agreed, good point! Let’s say there will be non-stupid arguments in favor of content moderation. For example (top of my head):
Children need to learn using creating tools; if a video editor was able to generate hard core porn or excessively violent content, it’s tricky to leave them alone with it.
AI doesn’t just generate content, it brings in knowledge. Some knowledge is restricted from circulation for good reasons (This is basically the bio-terrorism argument).
I tend to think it’s better to limit the model’s capability so it fits the use case (e. g. no ability to create porn in software used by children) than having a Llama Guard style moderation tool in the loop supervising both user and model behavior; but I’m still very vague on my own position in the debate. I also don’t know the different approaches people are trying out. I’d really like to read up on it though.
(What I do know: I don’t want the Metas and Googles of the world being in charge of defining the control mechanisms for something that will be involved in basically all of our creative processes.)
Well, thanks for reading five of them :) I’ll try to answer your concerns:
Film:
Film is a good argument, but mostly shows that we can handle “fakes” when framed right, e. g. when they are presented with context clues marking them as a movie. Generated images will not only often lack that framing, but will be presented to us framed as if they were representing something real. I would argue that this will devalue the context clues and makes it difficult/impossible to tell which images are real and which are generated in general.
X-risk compared to human level AI:
a) Political/societal destabilization while nukes are a thing = bad. Or more general: This interferes with our ability to deal with existing X-risks (including our ability to deal with the emergence of AGI).
b) We’d need to define X-risk a bit here. If we accept really bad societal outcomes (e. g. collapse of democracy followed by something decidedly bad), then my job convincing you should be relatively easy. The confusion this will cause should systematically benefit the fringes and actors following a “tear it down”-strategy. And I don’t think we are doing great in the stability department right now anyways.
We did epistemological impressive things before photography:
True, but not having a tool is different from loosing a tool you relied on for a long time. It’s also different from that tool suddenly doing something entirely different while still appearing to do the same thing on the surface.