QuoteThat's my point: the issue is not the input, but the output. I think it is only (or, at least, mainly) problematic if the output closely resembles "features characteristic of your expression."
The input can be problematic too if the results are too close to some sources.
Outputs from generative AI would probably fall under "fair use" in most cases but you can't assess that if the source pictures are not listed.
Since they probably use a very big number of pictures for each output, it could be in their interest to list the most significant sources with statistical weight.