AI writing detector that gives editors something to say to the writer

What an AI writing detector is for in a newsroom

An AI writing detector estimates how likely a passage was produced by a language model, based on measurable features of the prose. For a publisher, the useful output is not the number. It is the map: which paragraphs of a 2,000-word feature read machine-shaped, and what specifically about them does. Humanizeo produces that map. Every flagged sentence is highlighted with the reason attached, and the reasons are things a writer can act on.

The engine is our general-purpose AI detector. This page is about the editorial workflow around it: what a score proves, what it does not, and how to turn a flag into feedback that improves the piece rather than starting a fight.

What a score proves and what it does not

Worth being blunt here, because the misunderstanding causes real damage in editorial teams.

A score does not prove authorship. 80% means the passage shares a lot of measurable features with machine output. It does not mean a machine wrote it. Stylometric tests measure the surface of prose, and some people write in a flat, even, formal style by habit or training. Non-native English writers who learned the language formally are over-represented among false positives on every detector, ours included.

A score is not portable. Another tool will give a different number because it weights different signals. If your policy says "under 30% on tool X," you have a policy about tool X, not about writing.

Short samples do not carry enough signal. Under about 150 words there is not enough text to measure variance in anything. We say so in the interface instead of guessing.

What a score does show is where to look. The highlighted passages are the ones with even sentence lengths, identical paragraph shapes, stock phrases, repeated openers, or a hedged, say-nothing closing. Those are editing problems whether or not a model caused them. Treat the detector as a very fast first read that marks up the draft, and keep the judgement human.

How it works

You paste the piece, up to 30,000 characters. A bank of stylometric heuristics measures it: sentence-length burstiness, paragraph uniformity, contraction rate, stock-vocabulary hits, repeated sentence openers. In parallel an LLM judge reads each passage and gives its own opinion on whether the prose has the characteristic machine flatness. The two are combined into a per-passage probability, and each flagged sentence is mapped back to the strongest reason it flagged so that the reason is what you see next to the highlight.

No outside detection APIs. Text is stored only in your own history for 30 days, never used for training. The pipeline is ours end to end, which is the only way we could make it explain itself.

A passage-level review workflow

Here is the flow that works for desks reviewing bylined submissions, whether from staff or contributors.

1. Run it before the first read

Thirty seconds. You get the piece back with highlights. Now your first read is already focused: you know which sections to slow down on and which are clean.

2. Sort flags by reason, not by count

Ten flags for "repeated sentence opener" in one section is a two-minute fix and barely worth a note. Ten flags for "stock phrase" spread across the whole piece is a voice problem. Three flags for "hedged conclusion" on the final paragraphs means the writer did not land the piece. Same count, three very different conversations.

3. Decide what you are actually asking for

The detector cannot tell you whether the writer used a model, and chasing that question rarely helps anyone. What you can say with confidence is that the flagged passages read generic, and generic is not what you publish. Frame the note around the writing.

4. Re-run the revision

Before and after scores on every run. If the revision comes back with the same highlights in new words, the writer swapped synonyms. If the highlights are gone and the score dropped, the rhythm actually changed.

Turning flagged passages into feedback a writer can use

This is the part most detection tools ignore, and it is the part that determines whether the tool helps or just generates friction. A few notes that have worked well, based on how editors describe using the highlights.

Quote the tell, not the score. "Paragraphs four through nine all open with a topic sentence and close by restating it; vary the shape" is an edit. "This scored 74%" is not.

Give the list. If the flags are stock vocabulary, paste the phrases: "these seven phrases appear in the piece and read as filler; cut or replace with something specific to this story." Writers can act on a list. They cannot act on a feeling.

Ask for an opinion at the end. A hedged closing paragraph is almost always a writer who has not decided what they think. "What is your actual take? Say it in the last paragraph" fixes the flag and improves the piece.

Separate voice notes from fact notes. Detection is about style. Fact-checking is a different pass, and a piece that reads beautifully human can still be wrong. Do not let a clean detector run stand in for verification.

Do not accuse. Unless you have independent evidence, "this looks like AI" is a claim you cannot support and the writer cannot disprove. The highlights give you a way to have the conversation about the prose instead.

When the desk fixes it rather than sending it back

Sometimes there is no time to send a piece back. For those cases the flagged passages feed straight into the AI humanizer, which rewrites only what flagged, re-checks, and loops until the piece scores clearly human on our detector. Quotes, names, numbers and links are preserved verbatim, and that rule fails closed, so a rewrite that would touch a quoted figure is rejected. The journalistic tone preset keeps the register close to house style. You still read the result; the loop saves the hour, not the judgement.

Where this fits with the other detector pages

If the drafts you receive come mostly from freelancers working with one particular model, the ChatGPT detector page goes deeper on that model's specific habits. For editors also managing SEO or marketing batches, the AI content detector page covers pre-publish triage across many pieces at once.

And the boundary, stated plainly: Humanizeo is for editorial and professional writing you own. Academic use is prohibited under our terms, and we do not build for it.