The first rule is to separate the art from the algorithm: the output is a product, the input is a dataset, and that matters for the originality test. In the courtroom we’ll apply the “substantial similarity” doctrine, and in the gallery we’ll ask whether the piece carries the authorial imprint or is merely a derivative remix. Every pixel is a potential flag—if it was copied from a protected work without transformation, that’s infringement. If it’s truly recomposed, it’s a new work, but then we need to prove that the model’s training data didn’t violate the “non‑derivative” clause in the license agreements. So, dissect the code, dissect the provenance, and then decide if we’re in the red or the green. The verdict is still pending, but the audit is already underway.