Pangram flags push human writers to change their style

Writers and students are removing em dashes, avoiding words such as ‘delve’ and adding deliberate typos to keep AI detectors from flagging their work. Pangram’s document analysis links these stylistic features, along with rule-of-three patterns and antithetical phrasing, more strongly to generated text than human writing. The resulting pressure is changing how people express themselves, including those whose natural style resembles model output.

A study published in June in the International Journal for Educational Integrity tested four detectors across 160 academic papers. Pangram was the most consistently accurate, identifying AI use in 37 of the 40 papers combining human and machine text, and in 37 of the 40 where humans had rewritten AI-generated material. Those findings demonstrate its ability to detect mixed or edited material, but false positives remain a concern for human writers.

Ghostwriter Sally McGraw was falsely flagged, while science blogger Janelle Shane’s own writing was classified as AI with 98.7% confidence. Writers testing Pangram also found that replacing a single em dash with a colon could change a verdict from 100% AI to 100% human. Such sensitivity raises questions about screening reliability and creates opportunities to evade detection through minor edits.