jevgram heatmap shows which words look AI-written, not just one score

kieranklaassen · x · 2026-09-26

In "What the Jev" #3, Kieran Klaassen highlights jevgram (by Matt Hammer): instead of one AI-detection score, it asks the model the same question about every 1-, 2-, and 3-word window and paints a heatmap pinpointing exactly which phrases trigger the verdict.

The prompt is swappable — ask "does this person sound upset?" and the same sliding-window mechanism becomes an entirely different text-analysis tool.

Original post →

More from coding & agent

coding & agent channel →