AI detector that tells you which sentences flag, and why

What an AI detector actually measures

An AI detector is a tool that estimates how likely a piece of text was produced by a language model rather than a person. It does not read minds and it does not look up the text anywhere. It measures style: how regular the sentences are, how predictable the word choices are, how evenly the paragraphs are shaped. Humanizeo runs an AI detector built for people who publish for a living, and the one thing we do differently is show the evidence. Every sentence that pushes the score up gets highlighted, with a note saying what pushed it.

That matters because a single percentage is almost useless to an editor. "72% AI" tells you nothing about which of the 40 paragraphs to reread. A highlighted draft does.

Who this AI text detector is built for

Content teams reviewing freelancer submissions. Agencies signing off on client deliverables. Publishers with a house style that machine drafts keep flattening. SEO leads checking outsourced batches before anything goes live. If your work involves reading other people's drafts and deciding whether they are ready, this is the AI checker we built for you.

It is not for academic work. Our terms prohibit checking or rewriting coursework of any kind, and we mean it. Editorial and professional content you own, nothing else.

How it works

Two systems look at every draft, and they disagree often enough that running both is worth it.

1. Stylometric heuristics

A set of measurements that do not care what the text is about. Sentence-length variance across the document. How similar each paragraph is to its neighbours in shape and size. The rate of contractions. Hits against a list of stock phrases that models overuse. How many sentences open the same way. Each one is fast, explainable, and produces a number you can argue with.

2. An LLM judge

A language model reads the passage and gives its own opinion on whether the prose has the flat, hedged, over-balanced quality that machine output tends toward. It catches things the heuristics miss, like a paragraph that has decent rhythm but says nothing a human would bother saying.

3. Ensemble score, per passage

The two signals are combined into a probability for each passage and for the document. Then the highlighting layer maps every flagged sentence back to the strongest reason it flagged, so the note you see on screen is "five sentences in a row between 18 and 22 words" rather than a bare colour.

We do not call any third-party detection API. The whole pipeline is ours, which is also why we can explain it.

What the detector looks for

These are the five heuristic families, roughly in the order they tend to fire on a typical marketing draft.

Sentence-length burstiness

People write in bursts. A 31-word sentence, then a 4-word one. Models drift toward the middle, and a paragraph where every sentence lands between 15 and 22 words is the single most common tell we see. Low burstiness on its own is not damning (legal copy is uniform by design), which is why it is one input rather than the verdict.

Paragraph uniformity

Six paragraphs of three sentences each, every one opening with a topic sentence and closing with a soft summary. Humans rarely manage that kind of symmetry unless they are filling a template. The detector measures paragraph size and internal structure across the document and flags the run when it is too tidy.

Contraction rate

"Do not" versus "don't". Models default to the formal expansion far more than working writers do, especially in conversational registers where a person would never say "it is" out loud. A near-zero contraction rate in a casual piece is a strong signal; in a technical spec it is barely one at all, and the judge model accounts for that.

Stock vocabulary

Every model has a favourite drawer of words. We maintain a list of the phrasing that shows up at many times the human baseline, and count hits. One hit is noise. Twelve in 800 words gets highlighted with the offending phrases listed so the writer knows what to cut.

Repeated sentence openers

"This means that…", "Additionally, …", "By doing so, …". When more than a handful of sentences start with the same construction, the passage lights up. This one is also the easiest for a writer to fix by hand in two minutes, which is why we surface it prominently.

What the score means, and what it does not

The number is a probability. 85% means the text looks a lot like machine output by our measurements, not that a machine wrote it, and not that another tool will agree. Detectors disagree with each other constantly because they weight different signals. Treat ours as a well-argued second opinion with its reasoning shown, and treat any detector that claims certainty with suspicion.

False positives, honestly

Some human writing flags. Boilerplate-heavy formats (press releases, product descriptions, legal notices) are uniform by nature and read as machine-like to any stylometric test. Non-native English writers who learned formal register in school often produce exactly the even, hedged prose a model produces. Very short samples under 150 words do not contain enough signal for a reliable read, and we say so in the interface rather than guessing.

The fix is the same in every case: read the highlighted passages, not the score. If a flagged paragraph is plainly a person writing in a stiff format, ignore the flag. Per-passage explanation exists so you can make that call yourself instead of trusting a number.

Where the generic detectors differ

Different sources of machine text leave different fingerprints, and there are specialised pages for the ones that come up most in editorial work. A ChatGPT detector pass looks hardest at the phrasing habits and symmetrical paragraphing that model in particular produces. For outsourced SEO articles, agency deliverables, and programmatic pages, the AI content detector page walks through a pre-publish review flow and what Google has actually said about machine-assisted content. Newsrooms and publishers running passage-level review with writers get their own treatment on the AI writing detector page, including how to turn a flag into feedback a writer can act on. And if you are weighing us against GPTZero, Originality.ai, Copyleaks and the rest, the best AI detector comparison lays out word limits, highlighting, and who each tool is really for, without pretending we win every column.

When a draft flags: the rewrite loop

Detection on its own just gives you a problem. The reason the detector sits inside Humanizeo is that the flagged passages feed straight into the AI humanizer, which rewrites only what flagged, re-runs detection, and repeats until the draft scores clearly human. Facts, numbers, names and links survive verbatim; that constraint fails closed, so a rewrite that would alter a figure gets rejected rather than shipped. You see the before and after score on every run.

Most teams end up using both tools in one motion: paste, read the highlights, rewrite the flagged bits, re-check. Five minutes a draft, roughly, once the habit sets in.

Plans and limits

Every plan includes a monthly detect and humanize word allowance that resets on the 1st; see pricing for the numbers. Up to 30,000 characters per run. Your text is stored only in your own history and auto-deleted after 30 days, and nothing you paste is used for training. Works in English, Spanish, Turkish, German, French and other languages the input happens to be in.