In your hypothesis, it is assumed that the model is trained in some small part on unethical content, but where your logic deviates is then labeling the product of that training unethical when said exploitation has not been perpetuated or renewed. By the same judgment, ethical actions are impossible after the historical first (and “original sin” is systemic crowd control, not cultural metric).