Pangram AI Detector Repeatedly Flags Polished Human Writing as AI-Generated
On September 4, developer aronchick posted a series of complaints about the AI text detection tool Pangram: a full page of text he wrote using standard English writing conventions—three-part parallelism, Oxford commas, and the like—was repeatedly misjudged as AI-generated. Worse, the tool pointed to no specific flagged passages or reasons, leaving him no way to make targeted edits. He blasted its verdicts as "pure vibes."
Confirmed
- aronchick submitted the last PRFAQ he wrote at Amazon 12 years ago to Pangram, and it was judged 100% AI-generated. He joked sarcastically that this document, refined through 124 rounds of human revision and weeks of polishing, "turns out to have been all written by AI"
- Byrne Hobart (via a repost from Jeremy Nguyen) noted that Pangram, which has a decent reputation, has several known failure modes: misjudgment cases he has heard about, school essays written years ago being flagged as AI, and he cautioned about its limitations
Why it matters
- The incident once again exposes AI detectors' systematic false-positive problem against human authors who "write well": following standard writing conventions ironically becomes the reason for being flagged
- Because the detection tool provides no flagged locations or reasons, those misjudged have no avenue for appeal or revision—an issue with even greater impact in academic and workplace settings that rely on detection results
2026-09-04 ~ 2026-09-04 · 5 related posts
Primary sources
- AI detector Pangram's known failure modes, including private diary entries — JeremyNguyenPhD · 2026-09-04
- Writer says Pangram AI detector keeps flagging him for using good grammar — aronchick · 2026-09-04
- [source] AI detector Pangram keeps flagging human prose over Oxford commas, user says it's "just vibes" — aronchick · 2026-09-04
- [source] Amazon PRFAQ written 12 years ago flagged 100% AI-generated by detector — aronchick · 2026-09-04
1 near-duplicate retellings: aronchick