GPTHuman False Positive: What to Do When It Flags You

A GPTHuman false positive means a rewriting tool’s own detector scored your text as machine-made, or a rewrite you ran through it still got flagged somewhere else. Neither result proves authorship. A score is one instrument’s reading, and instruments disagree. What you do next matters more than the number.

What does a GPTHuman false positive actually mean?

GPTHuman is a rewriting service. You paste text in, it returns a paraphrased version, and it also offers detection. A false positive arrives in one of two directions:

1. You ran your own human-written draft through a detector and it came back flagged.
2. You ran a rewrite through GPTHuman, checked the output elsewhere, and it still scored as machine-made.

Both are ordinary outcomes. Detectors are pattern matchers. They read sentence rhythm, word predictability, and how uniform the phrasing is. A human who writes in long, even sentences with common vocabulary can look statistically similar to model output. So can a rewrite that swapped synonyms without touching the underlying structure.

The harder problem is what people assume the score means. A flag is not a verdict, and it is not evidence you can hand to a professor or an editor. It is a signal that some sentences carry machine-like patterns. That is a useful editing prompt. It is a bad accusation.

Why does a rewrite still get flagged after GPTHuman?

Because synonym swapping does not change structure. If the original sentence had a predictable shape, the paraphrased one usually keeps that shape. Different words, same rhythm.

Here is the pattern, illustrated:

Before (paraphrased, still flat):

It is essential to consider the various factors that contribute to effective communication in professional environments.

After (structure changed):

In an office, how you say something often matters more than what you meant. Miss that and the email lands wrong.

The first version is grammatical and vague. Every clause is balanced, every noun is abstract, and no specific subject does anything. The second has a short sentence, a concrete noun, and a claim someone could disagree with. Detectors tend to read the first as machine-made and the second as human, but the real reason to prefer the second is that a reader learns something from it.

That is the honest framing. You are not trying to beat a detector. You are trying to write sentences that carry information instead of padding.

If a GPTHuman false positive sent you hunting for a second opinion, run the same passage through the DeAIze detector and read the sentence-level highlights before you decide what to change. You can start free at /register and see which lines actually carry the signal.

What our own measurements show about flagged prose

We publish what our detector returns, including the uncomfortable parts. On 2026-10-04 we scored two corpora with the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold.

Corpus	Passages	Mean AI score	Median	Range	At or above threshold
Reference sample (DeAIze guides not marked generated)	1504	27%	20%	1-100%	487 of 1504
Unedited model output	12	52%	49%	11-91%	not reported

Read that table carefully, because it cuts both ways. Our reference sample is our own published writing, and 487 of 1504 passages still landed at or above the 30% threshold. These are passage counts, not counts of independent documents, and human authorship was not independently verified. So this is not a false-positive rate. It is a demonstration that a threshold catches a lot of prose a human wrote or edited.

The unedited model output scored higher on average, which is what you would expect. The gap is real but it is not a clean separation. Any tool claiming a clean separation is selling something.

Passage construction matters too: target 140 words, eligible tail passages can be shorter, and passages from the same guide are not independent samples. If you are comparing tools, ask what they measured and on what.

How do you tell a real flag from a tool artifact?

You look at where the score comes from. Not the headline number, the distribution.

Signal	What it suggests	What to do
One or two sentences flagged in a long document	Localised pattern, often a boilerplate transition	Rewrite those lines, leave the rest
Uniformly mid-range scores across every sentence	The instrument may not be separating anything	Try a different detector before editing
Short text, under a few hundred words	Scores get unstable on short passages	Add context or accept the reading is weak
Flagged text you wrote years ago, before any AI tool existed	Strong candidate for a false positive	Preserve the original and its date
Flagged text that you did run through a rewriter	Likely real machine-like patterns	Edit the structure, not the synonyms

That last row is the one people skip. If you paraphrased with a tool, the flag is probably pointing at something real in the phrasing. Not plagiarism, not misconduct, just flat prose. Fixing it is editing.

If you want a longer treatment of how scores behave when they are wrong, /blog/can-ai-detectors-be-wrong/ walks through the failure modes without pretending there is a clean answer.

What should you do when a GPTHuman false positive hits your work?

Work in this order. Do not start by rewriting everything.

1. Save the original. Export the file, note the date, and keep the version history if you have it. If anyone questions the work later, the draft trail is your evidence.
2. Check which sentences flagged. A document-level score tells you almost nothing. A sentence-level view tells you exactly where to look.
3. Read the flagged sentences out loud. Machine-like prose is often grammatically perfect and completely flat. If you stumble because the sentence is doing too much, that is a real editing problem regardless of any score.
4. Change structure, not vocabulary. Split a long sentence. Add a specific number, name or example. Cut a transition that exists only to connect two ideas.
5. Re-check once. Do not chase the score down through ten passes. If it moves, good. If it does not, you have still improved the writing.
6. Document what you did. If this is academic or professional work, a short note about your process is worth more than a clean score.

Step three is the one people resist. Reading aloud feels unscientific. It is also the fastest way to find the sentences that actually need work, because your ear catches what a score cannot explain.

Why do human writers get flagged at all?

Because human writing is not one thing. Academic prose is formal and repetitive by design. Non-native English writing often uses a narrower vocabulary range, which looks statistically similar to model output. Translated text carries source-language structure into English. Legal and medical writing uses fixed phrases on purpose.

None of that is misconduct. It is what those genres sound like. The problem is that a detector trained to spot uniformity cannot tell the difference between uniformity caused by a style guide and uniformity caused by a language model.

If you write in a second language, /blog/ai-detector-flags-non-native-english-writers/ covers the specific patterns involved and what to document. The short version: keep your drafts, keep your notes, and do not delete anything.

What are the options when you need to fix flagged text?

There are four realistic paths. They are not equal, and the trade-offs depend on what you are protecting.

Option	What it does	Trade-off	Good for
Paste into a rewriter and accept the output	Fast, changes wording	You lose your voice and may not notice changed meaning	Nothing you will sign your name to
Rewrite by hand from the flagged sentences	Slow, keeps your meaning	Requires you to read your own prose critically	Anything with your name on it
Use sentence-level detection, then edit	Targeted, shows you where the signal is	The score still will not prove authorship	Drafts you need to defend
Ignore the score and get a second reader	Free, catches real problems	No number to show anyone	Most everyday writing

If you are comparing tools rather than approaches, /blog/deaize-vs-gpthuman/ looks at what each returns on the same passage. The useful question is not which one wins. It is which one tells you what to fix.

How does DeAIze handle the same problem?

DeAIze is a web app with two halves: an AI detector and an AI humanizer.

Detection is sentence level. The whole document gets a score, and each sentence is highlighted by AI probability, so you can see which lines carry the signal. That is the difference between a verdict and a worklist.

The humanizer suggests revised wording for your own draft. It does not write your argument for you, and it will not resolve an authorship dispute. Compare its suggestions with your original and verify claims, quantities, citations, negations and qualifiers. Semantic similarity is a review aid, not proof that every meaning was preserved.

On privacy: text is processed for the current task only, never used for training and never resold. That matters if you are pasting unpublished research or client work.

There is a rewrite trial: 200 words after email confirmation, no credit card. Detection has separate configurable limits shown in the current interface. The rewrite trial is not a thousand free words and it is not unlimited scanning.

Pricing is pay-as-you-go, credits never expire, no subscription:

Plan	Words	Price
Starter	1000	$20.99
Standard	3000	$44.99
Pro	10000	$99.99
Max	30000	$199.99

Uploads accept .txt, .docx and .pdf. Rewriting a Word document preserves fonts, styles and tables. Detection and rewriting cover English and 9 other languages, and the interface ships in 10 languages.

What should you not expect from any detector?

Do not expect a clean answer about authorship. No detector output proves who wrote a text. Score definitions differ by instrument: a model score, a share of qualifying prose and a similarity percentage are not interchangeable, and reading one as if it were another is where most confusion starts.

Do not expect DeAIze to match another tool’s number. We do not guarantee any particular score on another detector, and we do not provide an official Turnitin readout or a Turnitin score estimator. If your institution runs Turnitin, that report is theirs to interpret.

Do not expect a rewriter to solve a policy problem. If your course or employer restricts AI assistance, check the applicable policy. If the work is disputed, preserve the original and process the evidence rather than editing the record. The ideas, data and conclusions have to be yours.

A short checklist before you conclude you have a false positive

• Did you keep the original file and its date?
• Did you run the text through more than one detector, and did you read what each one measures?
• Are the flagged sentences concentrated in a few places, or spread evenly across everything?
• Is the text short enough that the score is unstable?
• Did a rewriter touch this text at any point?
• Can you read the flagged sentences aloud without stumbling?
• Have you checked the applicable policy on AI assistance?

If you answered most of these and still think the flag is wrong, you probably have a real false positive. That does not give you a clean remedy. It gives you a documented position, which is what actually helps in a conversation with a teacher, an editor or a client.

Ready to see which sentences read machine-made?

A GPTHuman false positive leaves you with a number and no direction. The sentence-level view gives you a direction.

Start with the detector at /ai-detector/ and read the per-sentence highlights before you change anything. If you want to revise a flagged passage, open the humanizer at /humanize-ai-text/ and compare its suggestions against your original line by line. When you are ready to run a longer document, the plans are at /pricing/, and you can create an account at /register.

None of this proves authorship. It shows you where the signal is, so you can decide what to fix.