[dredmorbius]

Most legal tests to date have been on training data rather than generated works, though the number of instances in which training works emerge from prompts of AI LLM engines suggests the latter may also be problematic.

So far there's no binding precedent, but if a work can be shown to be in training data, then there's no need to trace it through the full generative chain as you outline.

See, e.g.: <https://astraea.law/insights/ai-training-data-copyright>.