Who owns what was used to train it?

A model has read books, articles, websites, listened to music, looked at images. Most of these works have an author, and that author has generally neither given consent nor been paid. This is the heaviest legal liability in the sector, and it remains unresolved.

The two opposing interpretations

From the labs' side: training a model is not copying a work, it is extracting statistical regularities from it. Like a human who reads a thousand novels and writes their own: they copy nothing, they have learned.

From the authors' side: the work has indeed been copied for processing, without authorisation, for commercial purposes, and the result directly competes with those who wrote it.

Both arguments hold up, and that is why the courts take years. Beware of anyone presenting this issue as clear-cut in either direction.

What has already happened

• Labs have bought paper books by the pallet, cut them up to scan them. Buying legally and then destroying the copy is a way to stay on the right side of copying, and it shows how seriously the question is taken.
• Lawsuits by authors and publishers are ongoing, some against the same companies and over the same books as twenty years earlier.
• In music, agreements are beginning to take shape on how rights are split for a generated track. This is the area moving fastest, because collective rights management already existed there.

And you, practically

Two things concern you directly.

What you generate. Rules vary by country, but the trend is clear: an image or text produced solely by a machine, with no human creative input, is poorly protected. You can generally use it; you will struggle to stop someone else from using it too.

What you imitate. Asking for "in the style of" a living, named artist, for commercial use, is the riskiest ground. Style itself is not protected, but an imitation that is too close, or that suggests the artist was involved, will backfire on you.

Key takeaway: do not carve any certainty into stone on this topic. A court ruling could shift it entirely, and you will then need to re-check.