I think that's true, but I'm not sure if this would characterize itself as "lifting ideas". I mean, it's one thing to personally take the idea and re-implement it yourself elsewhere. But I have to believe there could be an argument, that having an inference algorithm take the original source and re-implement it in the context of your engine might count as derivative work.

A particular work is deliberately supplied at inference time so the system can recreate or adapt it for another environment. It resembles commissioning a translation, port, or derivative adaptation through an automated tool, even where the output is not verbatim, so it seems it could hold a candle in court.

You have a few scenarios now up for legal debate:

    1. Is training a model using copyrighted material fair use?
    2. A model outputting chunks of identical coryrighted code from it's training material
    3. A model being given in it's prompt copyrighted source code, and asked to implement the same in a new source code.

I think #1 and #2 have open legal battles on it, but #3, which seems to be this scenario, is new to me, and I feel it could have legal ground for being considered infringing on the copyright.

Sidenote: how is your account "samatman" and created "April 24, 2010", that's a really funny coincidence