A humanizer that rewrites, not one that sells scores
Most GPTZero humanizer pages sell a score. LumenWrite is an AI humanizer built for the writing itself: it turns stiff AI-drafted English into prose with natural rhythm and plain words, and it keeps every fact, number, name and citation exactly as written.
We haven't tested LumenWrite's output on GPTZero, and we have no connection to GPTZero. So this page makes no claims about GPTZero scores. Results vary by text, and no tool can promise a pass.
One name trips people up: GPTZero and ZeroGPT are different companies with different detectors. ZeroGPT is the only one we've tested ourselves, and our ZeroGPT results are public.
Where GPTZero started: perplexity and burstiness
GPTZero was founded in January 2023. According to its own story, founder Edward Tian started working on AI detection while he was a student at Princeton. It made its name with two measures, and GPTZero calls itself one of the first detectors to use them:
- Perplexity: how predictable each word is to a language model. AI text tends to pick the expected word, so it scores low.
- Burstiness: how much that predictability varies across a document. People mix short, plain sentences with long, surprising ones. AI drafts tend to run at one even pace.
That's why so much advice about GPTZero is about varying sentence length. It fit the 2023 version well. For a longer explanation with examples, see how AI detectors work.
How GPTZero works now
GPTZero has moved on. In its own words, "as of autumn 2023, GPTZero uses a deep-learning based architecture that does not directly use perplexity and burstiness." Its technology page describes a sentence-by-sentence classifier that gives a result for the whole document plus sentence highlights, and labels text as human, AI or a mix of both.
The details are in a technical report GPTZero published in February 2026. It describes a model that sorts text into finer classes, including AI text that was later paraphrased. It also says the model was trained against four kinds of attack: paraphrasing prompts, paraphrasing models, commercial humanizing services, and attacks built with access to the detector's own internals.
So tips aimed only at perplexity and burstiness target a version of GPTZero that has since changed.
Does GPTZero detect AI humanizers?
GPTZero says it does. In November 2024 it announced detection of AI humanizers and paraphrasers, with the text it catches labeled "AI-paraphrased" rather than human. The feature, which GPTZero calls Paraphraser Shield, also covers homoglyphs: letters swapped for look-alike characters from other alphabets.
These are the company's claims about its own product, not independent test results. For what outside research says, see is GPTZero accurate? Either way, that label rests on a simple idea: a rewrite changes how a text reads, not where it came from.
What a GPTZero humanizer changes, and what it can't
Advice about GPTZero is everywhere. Here is how the common tips hold up against what GPTZero says about itself:
| Common tip | What it changes | How it holds up |
|---|---|---|
| Vary sentence length | Rhythm | Aimed at burstiness, which GPTZero says it no longer measures directly. Still worth doing, because it reads better. |
| Swap words for synonyms | Single words | The AI structure stays, and odd synonyms can change your meaning. |
| Hidden or look-alike characters | Nothing a reader sees | GPTZero says it guards against them, and they add nothing to the writing. |
| Run it through a humanizer | Rhythm, structure and word choice | GPTZero trains against humanizers, so results vary. Pick one that keeps your facts. |
| Add your own examples and reasoning | The content itself | The only step that makes the text more yours. |
LumenWrite handles the fourth row: it varies rhythm, cuts stock phrases and swaps formal words for plain ones, while your argument, sources, numbers and citations stay as they were. The last row is yours. No humanizer adds your thinking.
Before and after examples
These pairs are real output, straight from the model, in Natural mode: the lightest setting. The free humanizer above uses Balanced, which changes more.
Look at the business example. The rewrite is plainer than the draft, but "pretty central" is more casual than a formal essay wants, so you would tighten it. Reading for tone like that is part of the job, and no detector score does it for you.
Use it the way your class allows, and keep your drafts
A score is only part of what a teacher sees. GPTZero's own advice to teachers is to pair a scan with draft history, and its Writing Report replays how a Google Doc was written. Process counts, so keep your outlines, notes and earlier versions.
To check a rewrite before you use it, the free AI detector marks each sentence as reading human or AI. It's a different model from GPTZero's, so the scores can differ. To fix flagged sentences by hand, see how to make writing not sound like AI.
A humanizer is an editing tool. Follow your course's AI policy, say when you used AI if the policy asks you to, and if AI drafting isn't allowed for an assignment, a rewrite doesn't change that.