I don't follow. If the model was open-sourced under this GPL-like license (or a compatible license), then it would follow the GPL-like license. If the model was closed, it would violate the license. In other words, it would not affect open-source models at all.

Similarly, I could imagine carving out an exception when training on copyrighted material without licence, as long as the resulting model is open-sourced.

> If the model was closed, it would violate the license.

Training is fair use. The closed models wouldn't be impacted. Even if we assume laws gets changed and lawsuits happened, they just get settled and the closed source models would progress as usual (see Bartz v. Anthropic).

Meanwhile if somebody wants to go all "GPL AI" and only train their models on GPL compatible code, they'd just be restricting themselves. The amount of code they can train on shrinks drastically, the model quality ends up being garbage and nothing was won.

Further, assuming laws got changed, those models would now be incredible easy to attack, since any slip up in the training means the models need to be scraped. Unlike the big companies with their closed models, Open Source efforts do not have the money to license data nor the billions needed to settle lawsuits. It would mean the end of open models.