And if your output isn't distinct enough from the inputs, you too aren't allowed to claim copyright or sell your work without proper licensing.
With the AI we can at least be 100% certain of which input you trained it on and under which licens, making the whole a lot easier to deal with, as compared to humans. The liability is the same, but it's much easier to avoid legal implications, so why not play ball and ensure that you have the correct licenses?
It's not clear that you need any license to train on data in the vast majority of cases. Having a license to train on it won't guarantee that you can grant your users a license to any particular output, especially given the addition of user input. And most of the utility is in creating outputs that are indeed distinct.
So the answer to "why not play ball?" is:
1. It's not clear that it's legally required
2. It would be incredibly expensive and/or slow progress dramatically, or limit you to pre-existing licensed content (e.g. Adobe) which drastically reduces some types of capabilities
3. Given #1, for any company that doesn't have an Adobe-style library, "playing ball" is essentially betting the company that it will become legally required, because on top of developing an AI model you're going to have to become an expert content licensing and documentation studio
With the AI we can at least be 100% certain of which input you trained it on and under which licens, making the whole a lot easier to deal with, as compared to humans. The liability is the same, but it's much easier to avoid legal implications, so why not play ball and ensure that you have the correct licenses?