What if people use it to generate child porn?
Arguably, there are also lesser harms from consumption of it, but even so harmless generation of to-order child porn that collapsed the value of the harmfully generated kind by being indistinguishable from it would probably be a net win.
(Of course, with the techniques currently used, that would require a large corpus of the harmfully generated kind to work from; but one could imagine further developments in ML that allow better generalization that would allow that without any actual abuse images.)
https://en.wikipedia.org/wiki/Legal_status_of_fictional_porn... https://en.wikipedia.org/wiki/United_States_v._Handley
Is it the mere fact of someone being aroused by children that makes child pornography offensive to society, or is it our belief that the creation of such materials necessary involves some degree of exploitation? If the latter, what about the well-documented risks of exploitation of consenting adult pornographic performers? Does access to pornography increase or decrease adults' tendencies to engage in unsanctioned sexual acts?
I’m not sure what you’re trying to get at here but your post kind of reads like “child sexual abuse material is equivalent to porn made by consenting adults.” Posing that as a philosophical “what if” and then moving on isn’t very different from just stating that as an opinion.
I am not really interested in engaging on this topic, as the sort of “Debate me!”-style hot take is (in my experience) not usually offered in good faith, rather it’s more often some sort of rhetorical dick measuring contest amongst internet strangers.
The position that “things are equivalent in the case that I get to define all of the words “ isn’t the dialectical Judo throw that you think it is, buddy. It does not inspire confidence that a person is being serious.
So we should discuss it as an ethical problem because it's going to become practical and affordable far sooner than most people think. I would bet money it's already been tried and maybe done successfully, but for obvious reasons not advertised.
(This is legal in some countries, at least.)
In any case, since the training data here does not feature children, this particular model would be useless for that.
From the fact that weaboos tend to claim their favorite lolis are 9000 year old dragons or whatever, I'm guessing it's already legally fraught to trade drawings of CP. I'm guessing AI generated images would fall in the same category, though where responsibility falls I'm not as clear on. The person who input in such a way as to try to generate CP? The platform the generator is hosted on? Whoever got the training data and did the training? I feel like the person seeking CP is a clear answer but society is looking to hold facebook, twitter, and others responsible for content on their sites so I'm not as sure.
At which point the moral dimension clarifies: creating an incentive to abuse children to create better models for simulating child abuse is the exact same reason posession of CP is illegal to start with. And since it won't be possible to prove any particular model without perfect providence wasn't produced this way, their use will just be illegal.
This is, realistically, also the problem with just pornography generating models in general, and all of these models in total: if there's illegal data in the input set, what does that mean?
I suspect the next hurdle with them is going to be license compliance-enabled training - models which can prove any given output has at most a certain type of license and providence on input data.