FULL STORY

Pangram's False Positives on Human Writing Spark Backlash

Developers first mocked Pangram for flagging well-written human text as AI-generated; a week later, another author found his lightly AI-polished essay judged 100% AI, deepening doubts about detector reliability.

2026-09-04 ~ 2026-09-11 · 2 episodes · 11 posts

Episode 1 · Pangram AI Detector Repeatedly Flags Well-Written Human Text as AI (2026-09-04, 7 posts)

On September 4, developer aronchick posted multiple complaints about the AI text detector Pangram: pages he wrote using standard English conventions—rule-of-three parallelism, Oxford commas—were repeatedly flagged as AI-generated, and the tool gave no indication of which passages triggered the verdict, prompting him to call it "pure vibes."

Confirmed

  • aronchick submitted his last PRFAQ from Amazon, written 12 years ago, and Pangram judged it 100% AI-generated; he joked that the document refined over 124 rounds of human revision "was AI all along"
  • Byrne Hobart (via Jeremy Nguyen's repost) noted Pangram's known failure modes, including hearsay false positives and years-old school essays flagged as AI
  • mike64t mocked that using the phrase "which is precisely why" alone got a sentence flagged as obviously AI-written
  • mattslotnick joked that someone flagged so often by Pangram began believing he himself was an LLM

Why it matters

  • The episode highlights systematic false positives against humans who write conventionally well—standard style itself becomes the trigger
  • With no flagged locations or explanations, wrongly accused writers have no path to appeal or revise, a serious problem in academic and workplace settings that rely on detector results

Episode 2 · AI Detector Pangram Flags Human Writing as 'AI Slop', Sparking Debate (2026-09-10, 4 posts)

AI detector Pangram drew criticism after a user found that paragraphs they wrote themselves, but lightly polished by AI for grammar, were flagged as 'AI slop'—even 100% on retests. Economist Afinetheorem noted consistently 3-5% AI scores on merely edited text, questioning zero-tolerance academic thresholds.