AI detectors don’t work: here’s what to do instead

We get asked this most weeks, usually by someone with a deadline, staring at an AI score that they don’t agree with: “Can you edit this so it passes an AI detector?”

No, because the tools don’t measure what people think they measure, and a score isn’t something any editor can control.

What the tools actually measure

An AI detector doesn’t detect AI so much as measure statistical properties of text—chiefly how predictable each word is given the words around it—before reporting how close that pattern sits to the output of a language model.

That’s only a proxy, and not a very good one at that, because predictable prose is not a machine artifact. It’s simply what careful editing produces: plain sentence structures, standard terminology, and consistent register, without ornament, which is exactly what tends to raise detection scores.

They flag careful human writing. Non-native English writers, technical and scientific authors, anyone writing to a house style, and anyone who’s been professionally edited all produce text that looks statistically ordinary, which is a feature of the writing rather than proof of the author’s identity.

They pass generated text. A few unusual words and some varied sentence lengths will beat the measurement, and the document is no more accurate for it.

Why we won’t take the work

If the brief is “get me under a threshold”, we’re not the right service for you.

We can’t deliver it. No editor can promise a particular score, because detectors are unpublished, change without notice, and disagree with each other on the same document. Anyone who guarantees you a number is guaranteeing something they don’t control.

It’s the wrong target. A score isn’t a quality measure, and optimizing for one means making deliberate choices against clarity: adding variation for its own sake and reaching for the less obvious word. You’d be paying us to make your document worse in the service of an arbitrary number.

It doesn’t survive contact with a real reader. Your evaluator, examiner, or client isn’t running a detector—they’re reading. A document tuned to a tool rather than to a person fails the reader it was written for.

What to do instead

The useful question isn’t will this be flagged but does this read as though a person who knows the subject wrote it, and can I stand behind every claim in it.

Check every source yourself. This is the central risk in AI-assisted work, and the consequences can undo a reputation that took years to build. Fabricated citations and misdescribed sources are common, and your readers can check them, so open every reference and confirm that it exists and says what the draft claims it says.

Cut the length. Generated drafts tend to be long, and most can be cut down substantially without losing content. Shorter text reads more confidently and doesn’t waste your reader’s time.

Commit to a position. Models hedge by default, so where the draft says something may be somewhat significant, decide whether you think it’s significant, and then write that down instead.

Lean on your own knowledge. Specific figures, client constraints, and things that have gone wrong in your experience are what no model knows, and nothing marks a document as human-authored more clearly than knowledge that couldn’t have come from anywhere else.

Have a real person edit it. Not to disguise anything, but because judgment about what a document is for and who is reading it can’t be manufactured.

If you’ve been wrongly flagged

This is happening to many careful writers, and it’s a better idea to meet the challenge head-on than to try to rewrite until the score moves.

Keep your drafting history. Version history in Word or Google Docs, notes, outlines, and search records are far better evidence than any detector output, and if you’re challenged, this record is what proves your authorship.

Then ask what the detector result is being used for. Most institutions and employers treat a score as a prompt to look more closely rather than as a finding, so responding to the underlying question—can you account for how this document was made—puts you in a stronger position than trying to defeat the measurement.

Our position

We edit AI-assisted drafts; it’s one of our specialties, and we’re good at it. We edit for accuracy, clarity, structure, and voice, so that the document reads as unmistakably human writing.

We don’t use detection tools of any kind or evaluate your writing against them, and we can’t guarantee any effect on an AI, plagiarism, or similarity score. We can tell you whether the document is any good, and what it would take to make it better.