Simply put, I think AI Generation for the creation of art and images is more or less a bad thing and not one that's approached with the critical lens it should be.
It'd be one thing if AI image generation at all credited any of the artists it was pulling its work for or at all compensated them for it. It'd also be another thing if AI image generation didn't pose a larger problem for smaller work spaces in the near future. It'd be another thing if the training data wasn't literally just pulling from Google images. However, as the current state of things are at the moment, I just find it really hard to be excited for this. I am an artist and there's probably a chance in the future that smaller creators will be locked out of smaller art jobs that can often put them on the map because why pay for someone to do something like this?
I think in essence AI generation has been held back by its blurriness and bizarre warping- it's not really able to produce things with the same sort of lucidity a normal person would normally approach images. With new models of AI thought present in something like DALLE 2, there's less of that and it's where things begin to ramp up considerably into an endgame where this starts to look really bad for artists.
Ultimately it is a tool. End of the day it is. There are plenty of people who use AI Generation as their main praxis of generating their artwork and are able to do creative things with it. Perhaps they are able to fill their training data with entirely public domain images. I think that's fine- but ultimately if the endgame is to create assets from text prompts and get a close enough approximation to the desired thing, the less need there will be to talk to an artist about capturing it, getting their input and output. Will the consumer even care to know in the future what was behind the thing they consumed- whether it mattered at all if a human hand was involved in the creative process?
The hypocrisy probably doesn't look great on me, who partially uses AI generation for writing- but in essence I'm using it more of an aid at the moment than a full-fledged tool in the process. I still have to rewrite a majority of the lines that the AI spits out, and I don't expect that to change because writing is so context-based and bare. There's not really a way to "improve" on AI writing tools nor attribute their work to any particular author in the same way.
This is why I'm largely against the usage of AI generation for the creation of images- I speak as a 25 year old woman who went to art college for four years, but it's not that I think AI Image Generation really has the opportunity to "replace" us in any strict sense aside from a gimmick at the moment even with something like DALLE-2, but casual usage is going to ramp up in the future and it's not going to have the same approach as an AI artist who uses the program as a tool. The abstract thoughts of a machine trained to plagiarize from Google Images replacing smaller art jobs is an issue and one that will probably be alarmingly approaching. The broader implications of this seem to be largely ignored in discussions of these tools.
Google Images is often the basis for training data for these AI, and earlier I talked about how the training data needs to be cleaner and from artists who can be compensated for their work being used. I think even putting aside from how artists are often exploited through Google Images and through the training data, pulling from such a large net is going to still be unable to prevent biases of hate speech, racist and sexist imagery, and more unpleasant things hiding behind an AI generated image of a cat in a beret. AI generation isn't sophisticated enough to not pull from problematic material if it's in its training data, and this can become such a massive part of it that simply removing it can't fix it.
The ballad of AI Dungeon echoes across something like DALLE-2 or Imagen. For those unaware, AI Dungeon was a machine dungeon master of sorts and would take prompts from the user and play off it, to somewhat varying results. AI Dungeon had materials in its training data that were very NSFW, which turned out to be kind of an issue when it was put to task at moderating user-generated content that would have to pull from these specific bits of training data. More can be read here
, although a fair bit of warning that this content is heavily disturbing.
In conclusion- unless there's better efforts to keep humans within AI tools like this and to not blindly incorporate images without thought, don't support these projects. Yes, there is potential here that in its own right can interesting for people to dabble in, but as the consideration of AI image generation software in more actual use cases begins to become more rampant, there are problems inherent with how things are done right now that make it something I'm not willing to support.