Problem with the first two is that, to my understanding, there's zero guarantee that DALL-e will not generate copyrighted material which exists in some form or the other in the training data set. And you can't really check if the image generated holds it.
A custom shirt printing company would offer the generator. It's up to the customer to decide whether the output can be legally sold (or, perhaps, they're not selling it).
Anything you wish can be printed today, any logo of any brand. Only selling it is as a product restricted.
Whatever it covers, it never stopped anybody I know from ordering whatever shirts they wanted, even if it had like 100 logos of all possible brands on it (yeah we really tried to trigger some reaction - couldn't), and nobody ever had any problem with us ordering it or wearing it (or any other t-shirt we """illegaly""" made).
And I can murder someone in the forest and nobody will have a problem with it, because those that would care would not know. That shirt company would almost certainly be sued if interesting parties knew, as has happened before [1], just as if someone saw an obvious derivative work coming from an AI.
For a real world example, I've had real company logos appear on my GAN generated images. I used labels that included terms specific to a single companies rendering technology/pipeline, and the vast majority of images that had those labels had the company logo on it. They were somewhat distorted, but it was very obvious which company it was.
