I'm quoting the relevant bit from Valve's email in the reddit post, to indicate exactly what they're actually forbidding.
Notably: not all AI-generated content, but rather AI-generated content from models that were trained on material that's not owned by the person submitting the game.
It just has to be enough that Valves legal team can claim that they approved the game in good faith and that they can be held responsible for a game developer lying and knowingly violating the terms of service.
Most sd guis save the generational data in the final pngs as default. The metadata includes the model hash. The model can be easily checked against a data from civit or hugging face.
In court, it would be the entity claiming infringement that would have the burden of proof that their exclusive rights under copyright were violated by the assets in question, not the distributor of the asset that had the burden of showing ownership of every item in the training set of the model used in some part of the workflow.