AI content detectors: what a score means, and why you should not rewrite for it

A detector does not read your page for meaning and it has no record of how the text was made. It scores statistical patterns in the writing and returns a probability that text of that shape came from a model.

That is the whole mechanism, and it explains every complaint about them. Text that is clean, even and predictable scores as machine-made. Text that is uneven, idiomatic and surprising scores as human. Neither of those is the same as the truth about who typed it.

• The output is a probability or a percentage, never a verdict, whatever the interface implies.
• The same passage can score differently in two tools, and differently again after one paragraph is moved.
• Short passages score worse than long ones, because there is less pattern to judge.
• Editing a draft moves the score, which means the score measures editing as much as origin.

OpenAI shipped a classifier to detect AI-written text and retired it in July 2023 for low accuracy. The company with the most access to how the text was produced could not build a reliable detector for it.

The detectors on sale now, Originality.ai, GPTZero, Copyleaks, the one inside Turnitin and the ones bundled into SEO suites, are solving the same problem from further away. They are useful as a rough signal in a workflow. They are not evidence.

The false positives are not random. They land on particular kinds of honest writing, which is why using a score to judge a writer is unfair as well as unreliable.

The last row is the one that matters commercially. A page can be pure filler, pass a detector after twenty minutes of editing, and still deserve to rank nowhere.

Google has said it does not use AI detection as a ranking signal, and that its systems judge quality. Its spam policies target scaled content abuse: many pages made primarily to rank rather than to help, whatever produces them, automation, humans, or both.

In September 2023 it removed “written by people” from its helpful content guidance and replaced it with guidance about content created for people. The standard moved from who typed it to who it is for.

So there is no detector score that puts you at risk with Google, and none that protects you. A page that adds nothing is a problem at 0% and at 100%.

When a team starts editing to move a number, the edits follow the number rather than the reader. The usual result is worse.

• Sentences get lengthened and complicated, because irregular text scores as human. The page becomes harder to read.
• Idioms and asides get sprinkled in, which reads as padding to anyone who came for an answer.
• Time goes into the rewrite instead of into the missing example, the missing figure or the missing screenshot.
• The team learns that the goal is a green score, not a page someone would send to a colleague.

There is a version of this that is worth doing, and it is not the same thing: cutting the register, the hedged claims and the repetition out of a draft. That improves the page for a reader, and it moves the detector score as a side effect.

Four questions, in order. They take two minutes and they catch the pages a detector cannot see.

1. 1

   Is every figure traceable?

   Take each number on the page and name where it came from. A figure nobody can source is the single most common thing a model invents, and it survives every detector.

2. 2

   Is there anything first hand?

   A screenshot of your own product, a mistake your customers actually make, a result you measured. If the page would be identical for a competitor, it has nothing.

3. 3

   Does it answer the query in the first two sentences?

   Drafts that were never briefed wander for three paragraphs before the answer, because the model was writing around a topic rather than answering a question.

4. 4

   Why does this page exist?

   If the answer is “to rank for a keyword”, that is the pattern Google's scaled content policy describes, whoever wrote it.

There is one honest use. If you commission writing from a freelancer or an agency and you agreed on original work, a detector is a cheap first look at whether a submission was pasted out of a chat.

Even then, the score starts a conversation rather than ending one. Ask for the brief, the sources and the first-hand detail. A writer who worked from a brief can produce all three in a minute; a pasted draft cannot.

Inside your own team, the read-back that pays is the one that refuses figures a brief did not contain. It catches the actual risk, which is a confident wrong number sitting on a page for a year.