Pangram and the all-clear
Michael Wagner runs Pangram on his own openly AI-assisted essays. The detector flags them all as “100% human.” That’s a false negative — and he argues it’s the more dangerous error class.
Pangram is tuned aggressively against false positives (accusing humans of being AI) because that’s what gets vendors sued. The trade-off: anything ambiguous or edited falls into the human pile. Wagner’s writing is human-drafted then Claude-polished — enough idiosyncrasy remains to clear the threshold. The detector is doing exactly what it was designed to do, which is the problem.
This creates a research methodology flaw: studies using Pangram as ground truth for “AI-assisted writing” only catch unsophisticated / raw-LLM-paste use. Skilled users land in the false-negative bucket and contribute to the “human” quality average. So the comparison isn’t human vs. AI — it’s “bad AI use” vs. everything else.
Makes a good companion to JP’s piece on the adversarial use of Pangram — users gaming the system vs. the system’s own built-in blind spots. Wagner’s case is the mirror image: the all-clear where it shouldn’t be.