Researchers can now tell the difference between human text, AI text, human-edited AI text, and AI-polished human text

A detection method that distinguishes between four types of text — pure human, pure AI, human text edited by AI, and AI text made to sound human — just outperformed existing classifiers. This matters because regulators care about *who wrote what*, not just whether something came from an AI, since a human polishing an AI's work and an AI polishing human work have different legal consequences.

Detection systems have been stuck at binary choices: human or AI, or at best three categories. The problem is that a regulatory rule against 'AI-generated content' doesn't mean the same thing if a human wrote it and asked an AI to improve it versus an AI wrote it and a human fixed the punchlines. The legal and ethical stakes are different, and until now nobody could measure the difference reliably. This method works because it looks at the structural patterns left by the creator (the person or model who laid down the initial logic) separately from the stylistic patterns left by the editor (whoever came in afterward). That's a genuine bottleneck removal. Whether this actually gets used in regulation depends on whether policy bodies decide they care about this distinction — right now most don't.