We know that stuff generated from AI content is generally not your copyright, but there isn't any current ruling on whether or not you're free to use copyright-protected content to train a model in a legal way (e.g. fair use since it's been 'learned on', not directly copied). Some companies are using OpenAI gpt stuff, which is almost certainly trained on tons of copyrighted content and academic content, while other companies are being more cautious and soliciting models specifically trained on public domain/licensed content.