That’s the argument I would be making, but it certainly isn’t Microsoft’s (Copilot), OpenAI’s (Codex), etc’s position: they think the output is sufficiently laundered from the GPL training data so as not to constitute a derivative work (which means none of the original licenses – “open source” or otherwise – would apply, and the recipient could do whatever they want).

Edit: actually, to be more clear, I would take either of two positions:

1. That the presence of GPL (or in general, copyleft) code in the training dataset requires all output to be GPL (or in general, copyleft).

2. That the presence of both GPL code and code under incompatible licenses in the training dataset means that the AI output cannot legally be used at all.

(Position #2 seems more likely, as the license for proprietary code would be violated, too. It’s just that I don’t care about that; I only care about protecting the copyleft parts.)

i am vaguely familiar with software licensing is GPL type of open source?

You could say that, LOL. It’s the OG of “copyleft” licenses (the guy that made it invented the concept), although “permissive” licenses (BSD, MIT) existed before.

“Copyleft” and “permissive” are the two major categories of Free Software (a.k.a. “open source”, although that term has different connotations) license. The difference between them is that “copyleft” requires future modifications by people other than the copyright holders to be released under the same terms, while “permissive” does not. In other words, “copyleft” protects the freedom of future users to control their computer by being able to modify the software on it, while “permissive” maximizes the freedom of other developers to do whatever they want with the code (including using it in proprietary apps, to exploit people).

See also: https://www.gnu.org/philosophy/free-sw.en.html