AI detectors have become strict. A few years ago, cutting the word "delve" and swapping out a couple of stiff phrases could shift a score. That window has mostly closed. Modern detectors don't read your vocabulary and stop there. They measure how predictable your text is, how its rhythm moves, how its sentences open, how its paragraphs are shaped and how steady its tone stays from the first line to the last.
This guide covers what detectors actually look at, and then the workflow that gets the best results from EvadeGPT: draft in EveChat, humanize the whole text, and check it with the AI detector before you use it.
Key takeaways
- Detection goes far deeper than word choice. Detectors weigh predictability, rhythm, style habits, repeated phrase templates and tone, then feed all of it into a trained classifier.
- Em dashes, "delve" and "moreover" are surface signals. Removing them is the bare minimum, not the fix.
- Draft with the model that suits the job. EveChat gives you GPT, Claude, Gemini and more in one place.
- Use Ultra for the work that matters most. Standard handles everyday tasks. Rewrite is light rewording and isn't built to pass detectors.
- Send the whole text at once. In our testing, whole-text runs did much better than paragraph-by-paragraph runs.
- Detectors flag human writing too. Essays people wrote themselves get labeled as AI, especially from non-native English writers.
- Check before you submit. The AI detector shows the score before and after, and you can run a section again.
What AI detectors actually measure
Most detectors are classifiers: software trained to sort text into "human" and "AI" by learning the statistical differences between the two. None of them publish their exact recipe, but the signals they draw on are well known. Here is a short tour. Notice how little of it has to do with individual words.
Token probability and perplexity
A language model writes one token at a time (a token is a word or a piece of a word), and it tends to pick a likely next token. So AI text is built from likely words in likely orders. A detector can run your text through its own model and ask, token by token, how expected each choice was.
Perplexity is the summary of that. Low perplexity means the text was easy to predict, which is exactly what a model's own output looks like. Human writing is less predictable. People pick an odd word, cut a thought short, or start a sentence somewhere a model wouldn't.
Burstiness
Burstiness is how much that predictability, and sentence length and structure, change as you move through the text. Human writing comes in bursts: a long sentence full of detail, then a short one. Then a fragment. AI output tends to be even, with most sentences landing within a few words of each other and built the same way.
Stylometric features
Stylometry is the study of measurable writing style, the same field used to work out who wrote an anonymous text. Detectors borrow its tools:
- Function-word frequency. How often you use small words like "the", "of", "which", "that" and "it". These are nearly invisible to a reader, but their rates form a fingerprint, and models have their own.
- Sentence-opening patterns. Whether sentences keep starting the same way: with the subject, with a transition, with "This".
- Punctuation habits. How often you use commas, semicolons, colons, parentheses and dashes, and where they sit in a sentence.
- Paragraph shape. How long paragraphs run and how they are built. A topic sentence, three supporting lines and a summary, repeated all the way down, is a shape a detector can see.
Repeated n-grams and phrase templates
An n-gram is a run of n words in a row. "It is important to" is a 4-gram. Models reuse certain n-grams and sentence templates ("Not only X, but also Y", "This highlights the importance of...") far more often than people do. Detectors count them across the whole document, so one template used five times says more than one odd word.
Uniform register and tone
Register is the level of formality in your writing. People drift: a more casual line here, a sharper opinion there, a paragraph where they clearly got interested. AI text tends to hold one polished register and one even temperature from start to finish. That steadiness is itself a signal.
Classifier models trained on human and AI text
Finally, the detector puts it all together. Its classifier has been trained on large collections of human writing and AI writing, and it learns which combinations of these features separate them. It doesn't need any single signal to be conclusive. It looks at how many point the same way, and how consistently.
Why word swaps only scratch the surface
Put that list next to the usual advice. Removing em dashes, cutting "delve", "moreover" and "furthermore", swapping "utilize" for "use": all of that touches vocabulary and a sliver of punctuation. It leaves token probability, burstiness, function-word rates, sentence openings, paragraph shape, phrase templates and register almost exactly where they were.
That's why a text can look clean to you and still score as AI. Those edits are the start, not the job. Doing the rest by hand means changing the rhythm, structure and style of every paragraph while keeping the meaning, the facts and the citations intact, and keeping all of it consistent across the document. For a full essay, that's close to rewriting it. This is the work the humanizer is built for.
Detectors also flag people who wrote every word themselves
Even the best-known AI detectors get it wrong in the other direction: they label essays written entirely by hand as AI-written. This is not a rare edge case.
- Non-native English writers are flagged most. Stanford researchers tested widely used AI detectors on essays by native and non-native English writers. The detectors consistently misclassified the non-native writers' essays as AI-generated, while the native writers' essays were identified correctly.
- OpenAI withdrew its own detector. In July 2023, OpenAI took its AI text classifier offline, citing its low rate of accuracy.
- Universities have switched detection off. Vanderbilt University disabled Turnitin's AI detector in 2023, over concerns about false positives.
The reason is the same list above. Clear, well-organized writing with steady sentences and common transitions is exactly what many people are taught to produce, and it is also what detectors read as AI. A score is not proof of who wrote a piece of work. That is why it helps to check your own writing before anyone else does, and to keep your drafts and notes in case you ever need to show how you wrote it.
Step 1: Draft in EveChat with the right model
Good output starts with a good draft. EveChat gives you access to GPT, Claude, Gemini and more in one place, so you can pick the model that suits the job instead of forcing one model to do everything.
Each frontier model has its own strengths. As a general guide:
| If you're writing... | A model that often suits it |
|---|---|
| Long-form prose: essays, reports, reflections | Claude often writes more natural, flowing long-form text |
| Structured work: outlines, step-by-step answers, following a detailed brief | GPT is often strong at structure and following instructions |
| Something you're unsure about | Try the same prompt on two models and keep the draft that reads better |
These are tendencies, not rules. The best check is to read two drafts side by side and see which one sounds closer to what you meant.
A few habits make any draft better before it reaches the humanizer:
- Give the model real context. Paste the prompt, the marking criteria and your own notes. A draft built on your own ideas is easier to make yours.
- Ask for your structure. If you know the sections you want, say so. Ultra keeps your paragraph breaks, so the structure you set here carries through.
- Get the facts right now. Check names, figures and sources in the draft. The humanizer keeps them as they are, which is only useful if they are correct.
Step 2: Humanize with the right model
Once the draft says what you want, open the humanizer and choose a model. There are three, and they do different jobs.
| Model | Use it for |
|---|---|
| Standard | Everyday tasks: short answers, discussion posts, emails, routine work |
| Ultra | The work that matters most: big assignments and important submissions. Ultra is EvadeGPT's strongest model |
| Rewrite | Light rewording of your own text when you just want it to read a little differently. It isn't built to pass detectors |
The simple rule: if a detector score matters for this piece, use Standard or Ultra, and use Ultra when the stakes are highest. Keep Rewrite for text that is already yours and only needs a lighter touch.
Step 3: Send the whole text at once
This is the step people most often get wrong. It's tempting to humanize two or three paragraphs at a time, check them, then move on. It feels careful. It works against you.
Remember what detectors measure: burstiness, paragraph shape, repeated templates and tone across the whole document. A humanizer can only vary those things across the text it can see. Give it one paragraph and it has one paragraph of room. Give it the whole essay and it can vary rhythm and structure across the full document, the same way a detector reads it.
In our testing, whole-text runs did much better than paragraph-by-paragraph runs. We ran Ultra on complete texts and on the same texts split into paragraphs, then checked both against a strict detector, and the whole-text versions passed far more often.
You don't lose your layout by doing this. Ultra keeps your paragraph breaks, so the structure you set up in Step 1 comes back the way you left it.
Tip: Paste everything you'll submit, from the first line to the last, in one go. If a piece is very long, split it at natural section boundaries rather than paragraph by paragraph.
Step 4: Check it with the AI detector
Don't guess. Run the result through EvadeGPT's AI detector before you use it.
The detector shows the score before and after humanizing, so you can see what changed. If one section still isn't where you want it, you can run that section again instead of starting over.
Then read the whole thing yourself. You know what you meant to say, and a quick read is the fastest way to confirm every claim still says it.
What EvadeGPT does after humanizing
A few things happen automatically once the rewrite is done, so there's less to fix by hand:
- An automatic proofread pass fixes spelling and wrong-word errors.
- Em dashes are removed. They're one of the best-known surface tells, so they're handled for you.
- Names, numbers and citations are kept. The rewrite changes how things are said, not the facts you're relying on.
You should still read the final text before you submit it, the same way you'd check anyone else's edit of your work.
FAQ
Isn't removing em dashes and words like "delve" enough?
No. Those are surface signals. Detectors also measure token probability, burstiness, function-word frequency, sentence openings, punctuation habits, paragraph shape, repeated phrase templates and tone across the whole document. Word-level edits leave almost all of that in place.
Which humanizer model should I use?
Standard for everyday tasks. Ultra, EvadeGPT's strongest model, for the work that matters most, like big assignments and important submissions. Rewrite is for light rewording of your own text and isn't built to pass detectors.
Should I humanize my essay a paragraph at a time?
No. Send the whole text at once. The humanizer has more room to vary rhythm and structure when it can see the full document, and in our testing whole-text runs did much better than paragraph-by-paragraph runs. Ultra keeps your paragraph breaks, so you don't lose your structure.
Is it OK to use AI to write my essays?
That depends on your school and your course. Rules differ widely, so check your school's AI policy before you submit anything.
Put the workflow to work
Detection reads the shape of your writing, not just its words, and reshaping a full document by hand while keeping every fact in place is slow, careful work. The workflow above does the heavy part for you: draft in EveChat with the model that suits the job, humanize the whole text in the humanizer with Ultra when it counts, and confirm the result with the AI detector.
If you want more background first, our piece on AI humanizers vs paraphrasers covers why rewording alone falls short, the AI detector comparison looks at the detectors themselves, and how to make ChatGPT undetectable goes further on drafting.
Always follow your school's AI policy.
