I don't think we know that. I agree (or anyway hope) the arts are less amenable to synthetic data than e.g. math or coding.
Replies
i've seen ai-generated garbage at markets that i guess do kind of well, but if anybody can do it, then it just spoils the selection that much more. it happened with the ghibli trend where everything wound up being yellow and then needed ai to fix it. that sort of fixing doesn't seem intuitive.
aug 24, 2025, 2:50 pm • 1 0 • view
We do know, though! We know that "synthetic data" has an extreme deleterious effect on LLMs and image generators, this is KNOWN. They've already demonstrated it.
aug 24, 2025, 2:46 pm • 2 0 • view
I disagree. When training an image-processing AI (traditional, pre-COVID kind) having to produce synthetic data is not the exception. This can be as simple as rotating the image slightly or as complex as making collages and curing the seams (is it a wonder then that GenAI images look like they do?)
aug 24, 2025, 3:19 pm • 1 0 • view