I think that the claim that they make is that once a model is "contaminated" with GPL code, every output it ever produces should be considered derived from GPL code, therefore GPL-licensed as well.
If that theory holds - have to ensure that the models have not been trained on any code that is licensed incompatibly with the GPL, in which case the models could not be distributed at all