I Checked If My Contractor Used AI. Then I Checked My Own Writing. Now I Trust Nothing.

AI detectors are not truth machines. They are statistical guessing tools — and the moment you stake a professional or academic decision on one, you are flying on vibes dressed up as data.

Here is what actually happened to a content manager I know. She hired a freelancer for five blog posts. When the posts arrived, something felt off. Too smooth. Too uniform. She ran them through three popular AI detectors. Four out of five came back flagged — 80%, 87%, 91% AI probability. She fired the freelancer by email that afternoon.

Then, out of curiosity, she ran her own writing. Her weekly newsletter. The one she has written by hand every Thursday for three years. It came back 68% AI on one tool. 74% on another.

She called the freelancer back. It was an awkward conversation.

What Does "Check If AI Wrote This" Actually Measure?

AI detectors analyze text for two statistical signals: low perplexity (predictable word choices) and low burstiness (uniform sentence length). These appear frequently in AI-generated text because language models optimize for likely word sequences. Here is the problem — these same patterns appear in clear professional writing, ESL writing, technical documentation, and anything produced to a style guide.

A detector does not know who wrote the text. It only knows how predictable the text is. Those are not the same thing. If you want the technical breakdown of why this gap exists, our explainer on how AI detectors work goes deeper into the mechanics — and why fixing this at the model level is harder than anyone admits publicly.

Why the Score Feels Certain (Even Though It Isn't)

The number is the trap. When a tool outputs "91% AI," that reads like certainty. It is not. It is a confidence score from a probabilistic model trained on data the company likely will not publish. The same sentence scores differently across tools. The same document, lightly reformatted, can swing 30 percentage points on the same tool.

False positives are not rare exceptions. They are a documented, ongoing problem. Non-native English speakers are flagged constantly. Academics are flagged. People who write in a formal or structured register are flagged. The evidence on AI detection false positives is genuinely troubling — and it is not improving as models get better, because better models produce text that is statistically indistinguishable from careful human writing. The better the AI, the worse the detector's job gets.

So What Should You Do When You Suspect AI?

If you are a content manager, editor, or employer trying to verify a contractor's work, the honest answer is: a detector alone cannot confirm AI use. What you can actually do:

• Run the free AI detector on your own writing first. If it flags you, you will understand immediately why the score is not the final word.
• Ask for a revision with a specific new constraint. Real writers adapt in unexpected ways. AI-generated content tends to produce structurally similar output even under different prompts.
• Look for factual specificity. AI defaults to generality. A human who knows their subject includes specific details, counterarguments, and examples that were never in the brief.
• Request process evidence. A rough first draft, a voice note walking through the thinking, a working outline with dead ends. Process is something AI cannot fake in real time.

If you are the writer being accused, the situation is different but equally frustrating. Understanding how to prove your writing is human matters whether you are defending a dissertation chapter or an invoice for $600.

What If You Did Use AI — And Need to Fix the Score?

Let's be direct. A significant share of people searching "check if AI wrote this" used AI and want to understand their exposure before submitting. That is a legitimate thing to want to know.

The practical answer is humanization — restructuring the text so its statistical patterns shift toward human writing. WriteMask posts a 93% pass rate across major detectors including Turnitin, GPTZero, and Originality.ai. The goal is not evasion for its own sake. It is making AI-assisted writing read the way it should have been written from the start: with genuine voice, varied rhythm, and the kind of specificity that signals a real person thought this through.

The Question Nobody Asks (But Should)

Forget "did AI write this." There is no reliable answer. The better questions are: Is the text original? Is it accurate? Does it actually serve the reader? A post that clears every detector but says nothing useful is worse than a well-humanized draft that solves a real problem.

AI detectors measure a proxy for quality, not quality itself. And proxies fail — sometimes loudly, sometimes quietly, almost always at the worst moment. Before making any decision based on a detection score, use the AI detection risk quiz to understand what is actually driving the flag and whether it reflects a real issue or just a statistical accident that caught the wrong person on the wrong Thursday.