ChatGPT detector that highlights the actual tells

What a ChatGPT detector can and cannot tell you

A ChatGPT detector estimates whether a passage carries the stylistic fingerprint of OpenAI's models. It cannot prove who typed the words, and it cannot see a chat log. What it can do, when it is built properly, is point at the sentences that read like ChatGPT and say why. That second part is what editors actually need, and it is the part most tools skip.

Humanizeo's detector is a general AI detector; this page is about the ChatGPT-shaped signals it weights most heavily, because in editorial work that is where most machine drafts come from.

The tells ChatGPT leaves in a draft

Every model has habits. ChatGPT's are well documented by now, partly because so many people have stared at its output for so long. From what we see across drafts checked on our side, these are the ones that fire most.

Its favourite vocabulary

There is a drawer of words ChatGPT reaches for at many times the rate a working writer does. "Delve." "Tapestry." "Testament." "Multifaceted." "Navigate the complexities." "In today's fast-paced world." One of these in a 1,000-word article proves nothing. Nine of them in the same article, spread evenly through every section, is a pattern, and the detector lists each hit so the writer can see the drawer they have been pulling from.

Symmetrical paragraphing

Ask ChatGPT for an article and you get sections of near-identical length, each with a topic sentence, two or three supporting sentences, and a closing line that restates the topic sentence. Human writers do this when they are following a strict template. Otherwise they ramble in one section and get terse in the next. Our paragraph-uniformity measure catches the symmetrical version and highlights the run.

The signposting phrases

"It is important to note that…" is the famous one. Also "it is worth mentioning", "this highlights the importance of", "a key takeaway is". These are sentences that announce that a point is coming instead of making the point. The stock-phrase list flags them, and the fix is usually deletion rather than rewording.

Hedged conclusions

ChatGPT ends things softly. "Ultimately, the right choice depends on your specific needs and circumstances." Balanced, inoffensive, and empty. A human who has spent 1,200 words on a topic usually has an opinion by the end and states it. The LLM judge in our ensemble is good at spotting a closing paragraph that could be pasted onto any article on any subject, which is what a hedged conclusion is.

Even sentence length

This one is not ChatGPT-specific, but it is the most reliable signal across the board. Sentences that all land in the 15-to-22-word band, paragraph after paragraph, with no fragments and no long winding ones. The burstiness measure scores the variance and flags low-variance stretches.

How it works

You paste the draft (up to 30,000 characters) and two things happen in parallel. A set of stylometric heuristics measures the document: sentence-length variance, paragraph shape, contraction rate, stock-vocabulary hits, repeated openers. At the same time an LLM judge reads each passage and gives an independent opinion on whether the prose has that flat, over-balanced quality. The two are combined into a per-passage probability.

Then the part that matters for an editor: each flagged sentence is mapped back to the strongest reason it flagged and highlighted with that reason attached. You get a draft you can scan in a minute, not a number you have to interpret. We call no outside detection API. Everything runs on our own pipeline, which is why we can show our reasoning.

A workflow for editors receiving freelancer drafts

This is roughly how content leads who use the tool daily tend to run it.

1. Check before you read. Paste the submission and look at the highlight density first. A draft with two flagged sentences in 1,500 words needs a normal edit. A draft with forty needs a conversation.
2. Read the reasons, not the score. A 60% document score made up entirely of "low contraction rate" on a technical piece is probably just a formal writer. The same score made of stock phrases and hedged closes is a different situation.
3. Send the highlights, not the verdict. Writers respond far better to "these eleven sentences open the same way and these four phrases are filler" than to "this looks like AI." One is editing. The other is an accusation you cannot fully back up.
4. Re-check the revision. Before and after scores on every run make it easy to see whether the writer actually fixed it or just swapped synonyms.

Why a detector alone is not enough

Finding a ChatGPT-shaped paragraph is half the job. The other half is fixing it without wrecking the facts, and that is where most teams lose time. Manual rewrites of forty flagged sentences take an hour. A synonym-swap tool changes the words and leaves the rhythm, so the passage flags again.

The reason our detector lives inside Humanizeo is the rewrite loop. Flagged passages go to the AI humanizer, which rewrites only what flagged, re-runs detection, and repeats until the draft scores clearly human on our detector. Numbers, brand names, links and facts are preserved verbatim, and that check fails closed: a rewrite that would change a figure is rejected rather than shipped. Tone presets (professional, conversational, journalistic, technical) keep the register your publication uses.

That loop is what turns a detection tool into something an editor can use at 5pm on a deadline.

Honest limits

A ChatGPT detector cannot distinguish ChatGPT from Claude, Gemini or a fine-tuned open model with any real reliability; the stylistic overlap is too large, and that is fine, because for editorial purposes the question is "does this read machine-written," not "which vendor." Human writing in rigid formats flags. Short samples under 150 words are too thin. Other detectors will give different numbers because they weight different signals. Our score is a probability with its evidence shown, nothing more.

If you are dealing with a broader mix of outsourced SEO content rather than individual freelancer pieces, the AI content detector page covers batch review before publishing. And for a look at how the field compares, including where GPTZero and Originality.ai do things we do not, there is the most accurate AI detector roundup.

One more thing, stated plainly: this is not a tool for checking coursework. Academic use is prohibited under our terms. Editorial and professional content you own, only.