Why Human Writing Gets False Positives
I. The Accusation
An author submits a passage to a detector. The passage is his own, drafted and revised across weeks. A verdict displays on the screen: ninety percent AI-generated. He reads the number twice, certain some error has occurred, and finds none. The tool has not misread a sentence. It has misread him.
This is becoming a common occurrence among the writers least likely to deserve it. They revise with discipline, pruning a sentence until nothing remains but its bones. They choose a word for its exact weight and reject three others that were adequate. The detectors were built to catch a language model impersonating a human. What they identify is a human who writes the way machines were trained to imitate.
II. What is Measured
No detector can read for authorship. They assess statistical regularity through patterns of sentence length, word-choice, and paragraph rhythm, among other criteria. These proxies are not proof. They are useful only insofar as what they approximate behaves in expected ways.
Generative models were trained on the finest human prose available, from essayists and stylists to writers whose sentences already exhibited precision and internal consistency. The model learned the appearance of control. A detector trained to recognize that imitation will also flag the human original. The machine copied the master. Now the tool cannot tell them apart.
III. The Irony of Revision
A writer produces a first pass, then returns to strike out every echo, every phrase exposed as filler by a more careful reading. The second pass is smoother than the first; the third more refined still. By the tenth, each word seems to belong where it stands, as though no other were ever possible.
That inevitability is what a detector reads as artificial. It cannot distinguish a passage that was born effortless from one that was carved smooth by a hand unwilling to leave a flaw standing. The irony is exact: the more an author edits, the more the writing resembles what the detector was built to catch.
Fine writers have always sought this kind of command. Anyone who has read John Milton‘s long periodic sentences, each clause bearing its exact share of weight, or Gustave Flaubert hunting the single correct word, knows that impeccable control is one of the oldest ambitions in prose. It did not arrive with the machine. The machine arrived after it, learned to reproduce a poor version of it, and now mistakes the copy for the source.
IV. The Evidence Within the Work
What a detector cannot understand is the argument’s architecture, the way an idea planted in the second paragraph is not simply repeated in the ninth but transformed by it, carrying weight it did not carry at the start. It cannot infer the writer rejecting a word because it echoes another four sentences earlier, a decision no model has any reason to make, since repetition costs a machine nothing. It cannot see the claim that could only have come from years inside a single body of material.
These are the marks that separate an assembled text from one that is authored. They are invisible to a tool that reads only the surface pattern; they are the evidence that matters.
V. What an Author Can Do When Accused
The false positive cannot be argued away by protest alone. A detector‘s number carries a false authority because it looks like measurement, and what defeats it is documentation.
Keep the drafts. Every word processor or plain text editor with autosave preserves a trail that a finished document does not: paragraphs that appeared in one form on Tuesday and another form on Thursday. That trail is not proof in a legal sense, but it is evidence no generative tool produces.
If the work passed through an editor or a collaborator, keep the correspondence. Notes exchanged over word choices demonstrate the kind of deliberation a detector cannot see. The same is true of dated notebooks, outlines, and research materials. When a detector‘s verdict is presented as a reason to doubt the work, ask what the tool assessed. It measured resemblance to a pattern, not origin. The distinction is the entire question.
VI. The Defense Is the Work
An author who writes with precision will always be vulnerable to a machine that was trained to mimic care. There is no shortcut around this, and the honest response is not to lower one’s standards in order to read as human. That would be a strange defeat, yielding to a flawed instrument rather than trusting the discipline that produced the work in the first place.
The better answer is older than any detector. Writers who produce exacting prose have never needed an algorithm’s blessing. They have needed readers willing to sit with a paragraph and ask whether it holds. Let the sentence stand on what it argues and what it earns. That is the test a machine was never built to pass.
Rio de Janeiro, the xx day of July, MMXXVI.
Readers drawn to this question may find further ground covered in a companion series: twenty-two essays examining what separates AI-generated fiction from work written by a human hand.
Click here to explore the series on Pinterest.
