This is maybe more possible with local models than it is commercial ones. This question is in the context of image generation.

I was wondering if on any of the open LLM's there was a way of auditing the trail of data a model had processed ie what image was used to generate an image for a specific prompt.