Right now the whole model of generative AI and in general LLM is built on the assumption that training a machine learning model is not a problem for licenses, copyright and whatever. Obviously this is bringing to huge legal battles and before their outcome is clear and a new legal pratice or specific regulations are established in EU and USA, there's no point discussing licenses.

Also licenses don't prevent anything, they are not magic. If small or big AI companies feel safe in violating these laws or just profit enough to pay fines, they will keep doing it. It's the same with FOSS licenses: most small companies violate licenses and unless you have whistleblowers, you never find out. Even then, the legal path is very long. Only big corporate scared of humongous lawsuits really care about it, but small startups? Small consultancies? They don't care. Licenses are just a sign that says "STOP! Or go on, I'm a license, not a cop"

That is actually a quite interesting question. What is the license of the content posted to Lemmy? Would it be legal to share posts? Or use code posted here in proprietary projects? Do people retain full copyright, thus make sharing illegal? Can an instance in its legal terms define a standard license for content (like stackoverflow does)?

Finally, who would enforce the license?

Also, I don't think people that scrape training data care about all of this.

The licensing doesn't matter, most AI are trained off proprietary and copyrighted data. There's still a lot of talks in governments about whether this is legal or not, but at this point the cat's out of the bag and I doubt we'll regress back to using smaller amounts of data.