The Detection Deception, Part 1: AI Detectors Are Snake Oil and Everyone Knows It

By John Ayers

TLDR — Key Takeaways

AI detection tools measure word predictability (perplexity), not authenticity — the better you write, the more likely you get flagged
OpenAI built and killed their own detector in 7 months due to only 26% accuracy and 9% false positive rate
Stanford research proved AI detectors are biased against non-native English speakers, systematically flagging clean, clear writing as 'machine-like'
Detection is a race detectors are structurally guaranteed to lose — every new AI model gets better at sounding human
The real question isn't 'Did AI touch this?' but 'Is the expertise behind this content real?'

Here's something that should make you uncomfortable. The tools that publishers, platforms, and universities are using to "detect" AI-generated text? They don't work. Not "they're imperfect." Not "they need improvement." They fundamentally do not do what they claim to do.

John Ayers, Co-Founder of Chapters, walks through the evidence: OpenAI killed their own detector after just 7 months (26% accuracy). Stanford proved the tools are biased against non-native English speakers. Turnitin's AI flags caused universities to reverse their own policies.

The better you write, the more likely you get flagged. Sit with that for a second.

The detectors aren't getting better — the models are getting better at beating them. That's not the same thing. The real question isn't "Did AI touch this?" It's: "Is the expertise behind this content real?"

This is a summary. Read the full article with sources and analysis on John's author page →

Ready to turn your expertise into lasting impact?

Get Started