upvote
Previous-Token Prediction Based LLM Inversion for Prompt Reconstruction

(arxiv.org)

Imagine if you could take the AI marketing slop, or political smear campaigns, texts, and Images, and see the misanthropic sludge that was used as a prompt to generate it.
reply
Or what if you could take something not slopped and get back the prompt that would most closely result in it
reply