An interesting question would be over fair use. Generally, "transformative" works are covered under fair use. I can certainly see the argument that DL output is transformative in nature.
Not a lawyer take my comments with a grain of salt.
Not a lawyer take my comments with a grain of salt.
It lives by this vagueness and is abused as a tool for laundering the source code, to wash off the license terms for free software under a suitably complex and inscrutable system that establishing this provenance is not realistic, even if the output would fall on the "non-transformative" end of the scale.
I can kind of understand the argument that GitHub itself is, perhaps, not liable for this (setting aside the fact that they use Copilot in their own works, and GitHub's own software would have to be examined under this lens), but it is certainly irresponsible of them and disrespectful of the community that it owes its success to, and it seems to me that its users would definitely be at a higher legal risk.
Though has ruled that training AI on copyright materials is fine, as long as the defendants can claim it's a card catalog: https://towardsdatascience.com/the-most-important-supreme-co...
So it's going to be interesting to see how this is viewed by the supreme court when it gets there.
Or at least that is how I understand it.
It's a thin line. Depending on how the model is trained and what size it is, it is not unreasonable to expect that it will have the ability to remember training data.