A tool like Copilot can more or less automatically improve via telemetry, I’d expect the same thing as image models catch-up. I’d also expect the human signals to get further and further downstream of the actual creation process (e.g. Gameplay tester reports visual bug versus artist manually edits character)
An aside, but people use and pay for Copilot, it is out of the research lab phase.
None of this is solved algorithmically.
The current situation where you'd download billions of free images off the Internet only works once, and only if you somehow justify it as a research endeavour. Once this thing is monetized intellectual property laws will kick in.
They won't be able to compete just on data because lots of people are producing custom models, and they can be blended together like a stylistic and thematic pallet. Plus, if you own the pipeline you can use embeddings, dreambooth in specific elements and set it up in batch mode doing a random walk through the latent space for cheap. This stuff is not hard to set up and run, and with money and expertise, you can create something that is both unique looking relative to other AI art, and more optimized for your workflow than a service.
Image gen services will compete on ease of use, general quality and access to models that are larger than can fit in 24-48g vram without a big up front cost. There will probably be some services that provide specific features that people use even if they own their own pipelines, but the core customers will be smaller shops who don't use it enough to justify a real investment.
AI can make training datasets too.
For example to replace the human generated stuff from stable diffusion you could have some random-ish image generator coupled with some sort of image classification AI. As long as you have a good enough classification AI (or even more than one) that tells you what images are, you can focus on random-ish image generator algorithms to generate training data for another AI to generate images from descriptions.
(this is obviously with lots of handwaving and there will be problems that need to be solved - e.g. to avoid 99% of the generated training data be stuff like "noise on noise" but have some form of variety :-P), but the point is AIs generating data for other AIs is something that isn't far fetched and you don't need to think in terms of a single AI either)
(though https://commons.wikimedia.org/wiki/Commons:Freedom_of_panora... may matter here)
SD is open source, the paid stuff is going to be the paid models like this I think.
b) Training datasets can only be made by humans.
c) A paid tier of Stable Diffusion is obviously coming. It will be differentiated by a better (and more custom) training dataset.
d) No serious developer would be caught dead using the stock free tier Stable Diffusion.
e) Big studios will most certainly hire closely-guarded artists to curate and expand their proprietary training datasets.