IMO, saying "Pangram thinks Echo’s writing is human more often than frontier models." is only a testament to how unreliable AI detectors are.
Unfortunately, for people trying to bypass those detectors, that metric probably matters quite a bit. I've seen students
firsthand checking their AI-generated documents over and over again until they got them just right to pass.
A better, albeit not perfect, approach is to look at the writing process itself. For essays, a few tools let you examine and replay Google Docs revision history. I built one of them, Turingo (www.turingo.net), which tracks suspicious pastes throughout the revision history. This approach is not without problems, such as autotypers, which I'm hopefully going to address soon, but it beats a detector's guess.
Lately, I've been thinking about having a similar proof-of-process for other content as well (blog posts, code, etc). If anyone is working on such a thing, or knows any interesting projects, please let me know!
Unfortunately, for people trying to bypass those detectors, that metric probably matters quite a bit. I've seen students firsthand checking their AI-generated documents over and over again until they got them just right to pass.
A better, albeit not perfect, approach is to look at the writing process itself. For essays, a few tools let you examine and replay Google Docs revision history. I built one of them, Turingo (www.turingo.net), which tracks suspicious pastes throughout the revision history. This approach is not without problems, such as autotypers, which I'm hopefully going to address soon, but it beats a detector's guess.
Lately, I've been thinking about having a similar proof-of-process for other content as well (blog posts, code, etc). If anyone is working on such a thing, or knows any interesting projects, please let me know!