An AI-detection tool once flagged the U.S. Constitution as machine-written. That's not a funny glitch — it's the entire problem in one example, and it's why "AI detectors" are closer to astrology than science.
Schools, hiring teams, and editors are reaching for tools that promise to tell human writing from AI writing. Here's why that promise is one the technology can't keep.
There's no signal to detect
A detector has to answer "did a human or a model write this?" from the text alone. But a fluent model is trained specifically to produce text that looks like human writing — that's the whole objective. There is no reliable, intrinsic fingerprint separating "AI-fluent" from "human-fluent" prose, because the model was built to erase exactly that difference. So detectors fall back on crude proxies like how "predictable" or "smooth" the text is.
That backfires immediately. Clear, well-structured human writing looks "too predictable" and gets flagged as AI. This is why polished documents — and famously, formal texts like the Constitution — trip detectors. It's also why they disproportionately misflag non-native English writers, whose more measured phrasing reads as "machine-like" to a dumb statistical proxy. A tool that punishes people for writing clearly is worse than useless.
The base-rate trap
Even a detector that's "99% accurate" is dangerous at scale. Run it on 10,000 genuinely human essays and a 1% false-positive rate means 100 real students falsely accused of cheating. When the cost of a false positive is someone's grade, job, or reputation, "usually right" is not a standard you can build on. The confident percentage on the marketing page hides the human cost in the error bar. Understanding where a metric quietly lies is a discipline I apply across everything at my work.
It's an unwinnable arms race anyway
Suppose a detector actually worked. Anyone who wanted to evade it would paraphrase the output, run it through another model, or edit a few sentences — and the fragile signal is gone. Detection is inherently adversarial: the moment it works, it's trivial to defeat, which is why even the companies building these models have quietly retired their own detection tools rather than stand behind them.
What actually works
The answer isn't a better detector — it's abandoning detection for provenance. Don't try to guess a document's origin after the fact; establish it at creation with signed metadata and content credentials. And in education and hiring, the durable fix is process, not policing: assess in ways that value the thinking and the drafts, not just a final block of text a machine can also produce. I write more about these AI-trust questions at www.divyakush.com.
The uncomfortable truth: if you're relying on an AI detector to make a consequential decision about a real person, you're trusting a coin flip wearing a lab coat. The technology cannot deliver what it promises, and the confident ones are the most dangerous.
Related reading
- Guardrails: keeping LLM systems from going off the rails — the broader problem of trusting AI systems safely.
Top comments (0)