UnTXT documentation Back to the editor

False positives

A false positive is when a detector marks writing that a person genuinely wrote. They happen, they happen to ours, and this page is about who they happen to and what to do when one lands on you.

We would rather tell you this ourselves than have you discover it in a room where it matters.

Why human writing reads as machine writing

Models were trained on writing by people, which is the whole trick, and the writing they imitate best is the kind that is most regular. So the closer your prose sits to a well-executed standard form, the more it overlaps with what a detector has learned to notice.

Academic writing is a well-executed standard form. Hedged claims, an impersonal register, signposted transitions, a conclusion that restates and then looks forward — students are taught every one of those and marked down for leaving them out. A detector reading for regularity finds them and cannot tell whether they arrived through instruction or through a prompt.

That is not a bug we are about to fix, and any detector claiming to have fixed it is describing a benchmark rather than a room.

Who this lands on hardest

Four groups, and it is worth naming them rather than leaving it abstract.

People writing in a second or third language. Careful writing in a language that is not your first tends to be more regular, more formal and more evenly constructed than writing in your first, because you are working from rules where a native speaker is working from habit. Every heuristic detector reads that regularity as evidence, and the effect is roughly the same as turning somebody away from a job because their perfectly correct English carries an accent. It is the single worst thing about this whole category of tool.

People who over-polished. A draft revised eight times the night before a deadline loses the small irregularities that mark it as a person's, and it loses them for a reason that has nothing to do with AI.

People using grammar and paraphrase tools. Those tools smooth in exactly the direction detectors read as machine-written, so a document can pick up marks from a spellchecker's suggestions alone.

People writing to a rigid template. A lab report or a structured reflective piece has its shape imposed from outside, and predictability that came from a rubric looks identical to predictability that came from a model.

What we do about it

Not enough to make the problem disappear, and more than nothing.

Two models argue each passage out rather than one heuristic deciding on its own, which is directly because of the bias above — heuristics reading regularity are where it lives, and reviewers that can weigh what a passage is actually doing are less easily fooled by a formal register. Where they disagree, a third settles it, and you get to read the disagreement rather than a summary of it.

Anyone reviewing somebody else's work runs on a stricter threshold, so the ordinary formal register does not fill a page with marks that mean nothing.

And we mark what reads as genuinely yours as well as what does not, which gives a person who did write the essay something concrete to point at rather than only an absence of accusations.

If you have been marked and you did write it

First, breathe. A mark is not a verdict, nobody has been notified, and nothing has happened to you.

Read the reasoning attached to the passage, because it will usually tell you something true about the writing even when it is wrong about the origin — over-formal, structurally repetitive, generic where it could be specific. Those are worth fixing on their own terms.

Then decide. If the passage is yours and you are happy with it, leave it, and nothing is required of you. If you would like it to read more like you, the most reliable move is to put something specific into it: a real example, a detail only you would know to include, a claim you are willing to own in your own words rather than hedge.

If this is heading towards a formal conversation, there is a longer piece on that at If you have been accused, which is behind a free account.

If you are the one reviewing

Read the reasoning rather than counting the marks. A stack of grey on a piece written by a second-language student in a rigidly-structured assignment is close to the least informative result this tool produces, and the reasoning is what tells you so.

Ask about process before you conclude anything. Somebody who wrote an essay can generally say which source surprised them and what they cut, and somebody who did not tends to answer in generalities. That conversation is better evidence than any detector, ours included.

There is no villain here. The student is not one, and neither are you.

What we will not claim

We do not publish an accuracy figure. Every number in this market comes from a benchmark chosen by whoever is quoting it, and that includes ours, so we would rather give you the reasoning behind each mark and let you weigh it than give you a figure to trust.

Detection is probabilistic. A mark means a passage reads a certain way. It does not tell you who wrote it, and it is not going to start.

Try it — the first 1,000 words are free