False positive: when a detector flags writing a human actually wrote
A false positive is a detector error in one direction: a person wrote the text, and the tool says a machine did. The opposite error, a false negative, is machine text that reads as human. Every AI detector produces both, because detection is an estimate from texture, not a lookup against a source. Humanize 360 built its detector to show eleven named signals instead of one percentage partly so that a false positive can be examined rather than just believed.
Consider two sentences a student might write. "Furthermore, the results indicate that the intervention was effective in improving outcomes." That reads as machine to most detectors, and a diligent human wrote it after a semester of being told to sound academic. "The intervention worked, at least for the kids who turned up every week." That reads as human. Neither sentence tells you who typed it. Only the texture differs, and texture is all a detector has.
Who gets flagged most often
- Careful non-native writers, because textbook English is predictable English.
- Anyone writing in a formulaic genre: abstracts, method sections, policy summaries, cover letters.
- Writers who have been drilled to use transition words at the start of every paragraph.
- Heavily edited text, where a grammar tool has smoothed out every irregularity.
- Very short passages, where there is not enough material for the signals to settle.
Why formal English is at risk
Models learned to write from an enormous pile of published text, and published formal prose has a house style: even sentence lengths, hedged claims, "moreover" and "in addition", a list of three wherever possible. A student trained on the same conventions ends up with the same fingerprint. The detector is not wrong that the text is predictable. It is wrong about what that predictability proves.
What a fair process looks like
If you are assessing someone: never act on a score alone. Ask for drafts, notes and version history, and talk to the writer about the ideas in the piece. Our page at /ai-detector-for-teachers sets this out in full. If you have been flagged: keep every draft, run the text through a detector that shows its signals, and explain which signals your genre forces on you. A score is an opinion about texture, and you are allowed to argue with it.
What Humanize 360 does differently
The breakdown shows exactly which signals pulled the Human Score down, so a writer can point to "transition-word overuse" and say "my style guide requires those". The thresholds are published: above 70 reads human, 45 to 70 mixed, below 45 likely machine. And the tool never claims to predict another vendor's verdict, because that would be pretending to a certainty nobody has.
Common questions
How common are AI detector false positives?
Nobody can give one honest number; it depends on the tool, the genre and the writer. Every vendor acknowledges they happen. The safe assumption is that any single flagged result could be one.
Can I prove my writing is my own?
Not with a detector. Version history, earlier drafts, notes and a conversation about the content are far stronger evidence than any score in either direction.
Should I humanize my own writing to avoid a false positive?
If the detector shows two or three specific signals dragging it down, a light pass with Academic mode can fix them without changing your argument. Where your institution allows AI editing with disclosure, disclose it.
Does a false positive mean the detector is broken?
No, it means the detector is doing what it does: measuring texture. The error is in treating a texture measurement as proof of authorship.