AI Generated Text Detector: Verdict and the Fix for the Rate

An AI generated text detector compares your words to patterns it learned from machine output and returns a probability, not a verdict. It can be wrong on human writing, and it can miss machine text. What you do next matters more than the number: find the flagged sentences, fix the flat ones, and re-score.

What an AI generated text detector actually measures

Most detectors do not read your meaning. They read statistical texture. Two properties do most of the work:

• Perplexity — how surprised a language model is by each word you chose. Machine output tends toward the predictable word, so its perplexity is low and even.
• Burstiness — how much sentence length and structure vary. Human prose swings between short and long; generated prose tends to sit in a narrow band.

That is why a clean, well-organised paragraph can score high while a messy, specific one scores low. The detector is not measuring quality. It is measuring predictability.

We ran our own detector over two corpora on 2026-09-28 (local ensemble, lite tier, 30% flag threshold). Human-written sample: 709 hand-written DeAIze guides, mean AI score 26%, median 25%, range 0-69%, with 239 of 709 flagged at that threshold. Unedited model output: 12 passages from fixed prompts, mean 33%, median 34%, range 4-46%. Read those two ranges again. They overlap heavily. That overlap is the honest answer to “how accurate is this?” — the score is a probability estimate, and any single reading carries real uncertainty.

GPTZero’s own documentation puts it in similar terms. Their FAQ describes the class probability as the chance the detector is correct on similar documents — 90% meaning it is right 90% of the time on documents like yours (source: gptzero.me). Their confidence categories are tuned so that, when the confidence field is high, 99.1% of human articles are classified as human and 98.4% of AI articles as AI (source: gptzero.me). Note the wording: when the confidence field is high. Low-confidence readings are a different animal, and that is exactly where most false positives live.

Why your own writing gets flagged

If you wrote every word yourself and still scored high, the cause is usually one of these, not a broken detector:

1. Uniform sentence length. A draft where every sentence runs 18-24 words looks machine-made regardless of who typed it.
2. Low-specificity nouns. “Organisations face significant challenges” is predictable. “Our Bristol warehouse missed three deliveries in April” is not.
3. Transition scaffolding. “Furthermore”, “Moreover”, “Additionally” opening consecutive paragraphs is a strong tell.
4. Template structure. Five paragraphs of equal length with a summary sentence at the end reads like an outline filled in.
5. Non-native English register. Careful, formally correct academic English from a second-language writer often scores high because it is deliberately regular. If that is you, why an AI detector flags non-native English writers covers what to document.

Our own measurement found an AI-tell frequency of 1 per 100 words in both the human sample and the unedited model output. That is worth sitting with: the tells are not exclusive to machine text. Humans use them too, which is precisely why a flag is a signal to inspect, not a finding of guilt.

If you want to see which sentences are carrying the signal rather than guessing, run the text through our AI detector — it scores the document and highlights each sentence by AI probability, so you edit the flagged lines instead of rewriting everything.

Reading the report before you touch a word

Do not start editing the moment you see a high number. First, work out what the detector is reacting to.

What the report shows	What it usually means	Your first move
Whole-document score high, one or two sentences flagged	Localised problem	Rewrite only those sentences
Whole-document score high, most sentences flagged	Structural problem	Vary paragraph length and openings first
Score high but confidence low	Borderline reading	Re-run after one editing pass
Score high on a short excerpt	Too little text	Judge on the full document
Two detectors disagree sharply	Normal	Treat both as estimates, fix the shared flagged lines

On length: GPTZero’s FAQ notes that their detector handles texts up to 150,000 characters, or a batch of up to 250 files, on a free account (source: gptzero.me). Short samples are the weak spot for every tool, because there is not enough text to establish a rhythm.

A worked example

Here is an illustration, not a real document. Take this flagged-style sentence:

Before: “Furthermore, it is worth noting that organisations must adapt their strategies in order to remain competitive in an evolving marketplace.”

Every choice is the predictable one. The opener is scaffolding, the abstract nouns carry no information, and the sentence runs to a standard length.

After: “Our 2019 pricing model stopped working when two competitors undercut us. We changed it in six weeks.”

Same idea, opposite texture. Short sentence after a medium one. A specific year, a specific number of weeks, a concrete actor. Nothing here was “humanized” in a cosmetic sense — the content got sharper, and the statistical profile followed.

That is the whole method. Detectors respond to specificity and variation because machine output lacks both.

The editing pass that actually lowers the rate

Work in this order. It is faster than rewriting from the top.

1. Score first, edit second. Get the sentence-level view so you know how many lines are involved. One or two is a twenty-minute job; most of the document is a structural problem.
2. Cut the scaffolding. Delete “Furthermore”, “Moreover”, “It is worth noting” and rebuild the sentence around its actual subject.
3. Break the rhythm. If three consecutive sentences run similar lengths, split one and merge two others. Read the paragraph aloud; if it sounds metronomic, it will score high.
4. Replace abstractions with facts you own. Names, dates, places, quantities, decisions. This is the single highest-value edit, and it is the one no rewriting tool can do for you, because the facts are yours.
5. Re-score and compare. Keep the edits that lowered the rate without changing your meaning. Discard the ones that did not.

Step five is the one people skip. Rewriting without re-checking is guesswork.

If you are working inside a longer document and do not want to rebuild it by hand, humanize AI text runs a closed loop: score, rewrite only the flagged sentences, keep a rewrite only when semantic similarity says the meaning survived, then re-score and show before and after. The site reports an average AI-score reduction of 60%+ after humanizing, with a best case of 71% — those are the product’s own predicted scores, not a third-party verdict. On the example paragraph used on the site, a 92% AI score becomes 4% AI.

When this is the wrong choice

Be clear about the boundaries before you spend anything.

Skip paid humanizing if the flag is a false positive on your own writing. If you wrote the draft yourself and it is accurate, the fix is editing, not rewriting. Rewriting your own sentences through a tool risks flattening your voice further. Run it again after a rhythm pass and see.

Skip it if the issue is structural. A humanizer works at sentence level. If your whole document is five equal paragraphs with a summary sentence at the end, no sentence-level tool will fix that. Restructure first.

Skip it entirely if the work is not yours. No tool here is for passing off someone else’s writing or for misrepresenting authorship. The ideas, data and conclusions have to be your own, and if your institution or client requires disclosure of AI assistance, disclose it. Is bypassing AI detection cheating? walks through a disclosure test that is worth reading before you decide how to handle a flagged submission.

Skip it for very short text. Captions, subject lines, a two-line bio — there is not enough signal for a stable score, and paying per word for that is poor value.

What to do when an accusation lands

A flag is not proof, and you should not act as if it is. Detector output is a probability estimate. GPTZero’s documentation says so directly, and our own numbers show why: human guides ranged from 0% to 69% on the same instrument, so a single reading anywhere in that band proves very little on its own.

Keep a record. Save your drafts, your version history, your notes and any sources you consulted. If you are a student facing a panel, what academic integrity panels assess sets out what they actually look for, and it is rarely the raw percentage. If you are a freelancer, agree on AI disclosure with the client in writing before the work starts.

Then do the editing pass above, re-score, and bring the evidence rather than the argument.

The rate comes down when the writing gets more specific and more varied. That is the fix, and it is also just better writing.

---

Measurement note: figures in this article come from our own detector run on 2026-09-28 — n = 709 hand-written guides versus n = 12 unedited model outputs, median AI score 25% and 34% respectively. Method: /methodology/.