In 1776, Thomas Jefferson drafted the Declaration of Independence. In 2025, AI detectors flagged it as 99% machine-written.
That’s not a metaphor. That’s a documented failure.
ZeroGPT and OpenAI’s own detection tools labeled the Declaration—a document written nearly 250 years before large language models existed—as AI-generated. Not borderline. Not “maybe.” But with 97–99% confidence.
And it’s not just the Declaration. The 1836 Texas Declaration of Independence? 86.54% AI. The U.S. Constitution? AI. The Book of Genesis? AI.
These tools don’t know what a human is. They don’t know what a machine is. They only know statistical patterns—and they’ve been trained on the very tradition of excellent human writing they now penalize.
I’ve Lived This
I’ve published six cybersecurity books. I’ve spent ten years refining my craft—learning to hear the rhythm of my own sentences, cutting what doesn’t serve the work, building judgment through revision.
And I’ve had my manuscripts flagged as AI-written.
Not because they were. But because they were too coherent. Too structured. Too precise.
The very qualities that define mastery—rhetorical clarity, logical flow, clean syntax—are the same ones that trigger these detectors.
And when I challenged the flag? I was asked to perform. To write an essay on the spot, in front of a panel, to “prove” I could write.
Performance ≠ Competence
Five years ago, I would have failed. Not because I couldn’t write—but because I couldn’t perform under observation.
I had to learn performance as a separate skill. Stage presence. Stress regulation. Composure under scrutiny.
The panel thought they were measuring honesty. They were measuring performance.
This is compliance theater—the appearance of rigor without the substance of understanding. And it’s being deployed in high-stakes environments: publishing, academia, hiring.
From a Security Mindset
In cybersecurity, we don’t trust tools blindly. We validate. We test edge cases. We understand that false positives can cause real harm.
AI detectors are being treated as neutral arbiters of truth. They’re not.
They’re statistical guessers trained on human excellence—and now they punish humans for writing well.
The Declaration of Independence is irrefutable evidence. It predates AI. It has documented authorship. And it still got flagged.
If that can happen to Jefferson, it can happen to anyone.
What We Need
• Transparency: Detectors must disclose their methodology and error rates.
• Appeal Mechanisms: Authors must have a path to challenge false positives.
• Human Judgment: Institutions must stop outsourcing trust to flawed tools.
• Trauma-Informed Assessment: Performance under pressure is not a proxy for authenticity.
Final Motif
Mastery should not be suspicious. You shouldn’t have to write worse to be believed.
If you’ve been flagged, doubted, or forced to perform—you’re not alone. And you’re not the problem.
The system is.
Top comments (4)
What makes this particularly insidious from a security research perspective is that the people most likely to be harmed are exactly the ones we most need writing: self-taught practitioners who’ve built mastery through operational experience rather than credentialing.
Your own work exemplifies this . You write with precision because you’re documenting tools that actually have to work in production. That clarity isn’t performative; it’s what happens when you live with systems that fail loudly if your mental models are even slightly fuzzy.
It’s funny that the same qualities that keep real systems safe are the ones these detectors are now treating as suspicious.
What’s insidious isn’t just that detectors mislabel clarity as “AI‑like.” It’s that they invert the trust model of security writing. In practice, the people who document systems with precision—self‑taught practitioners, operators who’ve lived through outages, analysts who know what breaks in production—are the ones whose prose gets flagged. The irony is sharp: the closer your writing mirrors the rigor of a functioning system, the more likely an algorithm is to call it inauthentic. That’s not just a false positive; it’s a structural bias against operational expertise.
It stops being “AI detectors are inaccurate” and turns into “the system is structurally biased against people with operational scar tissue.”
I’ve been seeing the same thing in security-writing spaces: the closer someone writes to how they’d debug prod, the higher the chance a classifier calls it inauthentic. It makes me want to experiment with ways to preserve that rigor while avoiding unnecessary detector blow-ups at the same time.
Exactly—once you see it as structural bias, the stakes change. It’s not just about false positives; it’s about systematically sidelining people with operational scar tissue, the ones who’ve earned clarity by living through outages and debugging prod.
What worries me is that the “detector blow‑ups” aren’t random—they’re triggered by the same rigor that keeps systems safe. Which means the more someone writes like they actually work in production, the more likely they are to be flagged. That’s a perverse incentive.
Experimenting with ways to preserve rigor while dodging detectors feels like defensive obfuscation—teaching practitioners to sand down their clarity so they don’t get punished. That’s the opposite of what we need. Maybe the better experiment is exposing detector fragility itself: publishing paired samples (rigorous vs. meandering prose) and showing how the classifier rewards sloppiness. That way the bias is undeniable, and the burden shifts back to the toolmakers instead of the writers.