Pangram AI Detector Under Fire After Flagging Victor Hugo as 100% AI

AI text detector Pangram sparked controversy after a string of false positives: a paper author who only polished the abstract and about 5-10 sentences with an LLM saw his 73-page thesis flagged as 63% AI-generated, and his follow-up test marked Victor Hugo's poetry as 100% AI. The cases were widely shared and reproduced in the community, raising doubts about the reliability of AI detection tools.

Confirmed

• A thesis co-author (relayed by @GolinoHudson) said only the abstract and roughly 5-10 sentences of the 73-page paper were rewritten with an LLM for readability, yet Pangram judged 63% of it AI-generated — a false positive in his view.
• In further testing, Victor Hugo's poetry was flagged 100% AI. French engineer francip reported the same case, and @JFPuget reproduced the false positive and confirmed the text was by Hugo.
• Antoine Tilloy (relayed via @birchlse) noted that multiple texts thousands of lines long, written before 2020 — before modern LLMs existed — were judged "100% AI-generated" by Pangram v4.
• Discussants argued that if Pangram is sensitive enough that removing a few line breaks flips its verdict, the burden of proof lies with Pangram, not the accused.

Why it matters

• False positives have extended from edge cases to classic literature and pre-LLM historical texts, suggesting the detector's logic is overly sensitive to stylistic features and could systematically penalize authors who use polishing tools.
• As academia and content platforms increasingly rely on AI detection, such errors fuel fears about convicting authors on detector output alone and revive the question of whether the burden of proof should rest with detector vendors.

2026-09-23 ~ 2026-09-24 · 5 related posts

Full story(10 episodes)→

• Episode 1: Award-Sweeping French Novel Flagged as Nearly Entirely AI-Written(2026-09-21, 13 posts)
• Episode 2: AI Detector Pangram Flags Decade-Old Human Text as AI(2026-09-22, 3 posts)
• Episode 3: Goncourt Favorite Accused of AI Writing, Author's Past Plagiarism Exposed(2026-09-22, 9 posts)
• Episode 4: Debate Erupts Over Pangram AI Detector Reliability: False Positives, Humanizers and an Arms Race(2026-09-23, 9 posts)
• Episode 5: Bounty Raised to $100K for French Text Falsely Flagged as AI(2026-09-23, 3 posts)
• Episode 6: Pangram AI Detector Under Fire After Flagging Victor Hugo as 100% AI(2026-09-23, 5 posts)
• Episode 7: Pangram AI detector criticized: catches crude AI text, misses polished work(2026-09-24, 4 posts)
• Episode 8: French AI-Detection Row Casts Doubt on Pangram(2026-09-24, 2 posts)
• Episode 9: Goncourt Prize Drops Novel Flagged by AI Detector(2026-09-25, 4 posts)
• Episode 10: AI-Generated Novel Shortlisted for French Literary Prize Sparks Ethics Debate(2026-09-25, 2 posts)

• Pangram v4 flags pre-2020 thousand-line texts as 100% AI-generated — birchlse · 2026-09-23
• Pangram AI detector under fire: removing line breaks triggers false positives on human writing — JFPuget · 2026-09-23
• Pangram Flags Victor Hugo Poem as 100% AI, Sparking Fear of False Accusations in Academia — GolinoHudson · 2026-09-23
• Victor Hugo text falsely flagged as 100% AI-written by detector — JFPuget · 2026-09-23
• [source] Pangram AI detector flags a Victor Hugo poem as 100% AI-generated — GolinoHudson · 2026-09-24