AI Content label errors are spreading across Instagram for the second time in roughly two and a half years, and the pattern is the same as it was in 2024: ordinary photographs are being tagged as synthetic while genuinely synthetic images sail through untouched. The Verge’s Jess Weatherbed reported on 4 September 2026 that users […]
AI-assisted writing is now the hardest thing on the internet to label, and that is the argument Pangram co-founder and chief executive Max Spero brought to TechCrunch’s Equity podcast on 2 September 2026. The 24-minute episode, produced by Theresa Loconsolo and hosted by Rebecca Bellan, was published under a blunt headline: AI detection is harder […]
Pangram, a 24-person Brooklyn startup, has become publishing’s de facto AI referee: its scores cancelled a Hachette novel, embarrassed The New York Times and cast doubt on a major literary prize, and Substack now bakes the detector into its platform. Independent NBER benchmarking really does put it far ahead of GPTZero, Originality.ai and Turnitin on false positives — yet humanizer tools evade it over 96 percent of the time, verdicts wobble on short excerpts, and bias questions hang over who gets accused. This article weighs the evidence from WIRED’s investigation, the Chicago Booth benchmark and the company’s own disclosures to answer the practical question: when should you trust a Pangram score, and when should you not?
Can you teach yourself to detect AI writing? The honest answer is maybe. There is no single giveaway — only accumulating habits: signature vocabulary such as delve, meticulous and quietly, the “not X, but Y” frame, the rule of three, claim escalation and length that ignores the situation. The evidence is more encouraging than the old consensus: annotators who use these systems daily reached 86.7% to 96.7% true-positive rates individually and 99.3% as a majority vote across 300 articles, beating every automated detector except Pangram, while occasional users managed 56.7% — barely above chance. This breakdown covers the tells that hold up, the human-versus-detector numbers, the Stanford finding that seven detectors falsely flagged 61% of essays by non-native English writers, why the tells keep expiring, and what to do instead if you commission content.
Fighting AI with AI guide for defending against deepfake phishing and cloned-voice social engineering with verification workflows, AI detection, training, and governance.
Go Humanize AI is a text humanization tool that rewrites AI-generated drafts into more natural language while preserving meaning and workflow speed.
