If your AI is capable of reproducing part or whole of a bit of content it was trained on, then your AI is subject to copyright. Training an AI to create derivative works, does not absolve you of violating the copyright of the authors your AI is building derivative works from. If this is too much of a legal burden for you, then don’t train your AI on work it may create derivative forms from.

It’s a pretty simple problem. The folks claiming it is not either are being intentionally disingenuous or honestly do not understand the legal definition of derivative works in copyright. It’s settled law.

You could… you know either a) don’t train on works you don’t have a license to or b) use some sort of adversarial training to ensure that the AI doesn’t replicate the work it is trained on.