Last night I was reviewing a partner submission in a cafe, and the more I read, the more it felt off: every sentence was smooth, but the overall texture felt like it all came out of the same mold. I’ve gotten stuck here too. Before, I even waved this kind of “machine text that looks human-written” through, and the rework later was more exhausting.

This method isn’t about adding an even bigger AI. It’s about using an “old way” for a first text screening pass: looking at how words appear and how sentences are distributed, then judging whether it resembles machine-generated text at scale. Last Tuesday at 9 p.m., in a coworking office in Wangjing, Beijing, Lin Ran, who sells knowledge products, used it to screen student assignments—not to catch people, but to pull out suspicious pieces first and then spend another 10 minutes reviewing them manually.

The cost to copy this today is pretty clear: on money, we can treat it as $0 to start and just follow the logic in the article; on time, about 30-60 minutes is enough to understand that it is comparing writing habits, not reading for meaning; the technical barrier is not exactly low, so if we don’t touch code at all, it’s fine to treat it first as a way of thinking; the first step is not installing a complicated environment, but opening the original link and seeing how it turns “looks AI-written” into text features we can actually compare.

If we’re just getting started, I’d treat it as a reminder before reviewing, not rush to build a real system. If we already have 1-2 clients, I’d use it to help screen proposals, outsourced drafts, and student assignments—filter first, then review by hand. If we’re scaling up, I’d consider plugging this kind of detection into the content workflow, at least to reduce the team’s first round of manual eyeballing. But not everyone needs this tool, and if we don’t try it now, that’s fine too. Just knowing this path exists is already useful.