Does DeAIze Work? What Our Own Numbers Show

Does DeAIze work? For the job it claims — showing you which sentences carry machine-made signal — yes, with limits we can measure. On 2026-10-04 we scored 1,504 passages from our own guides and 12 unedited model outputs. The reference sample averaged 27%; unedited output averaged 52%. Neither number proves authorship, and we flag plenty of human prose too.

Key takeaways

• DeAIze reports a document score and a per-sentence AI probability, so you can see which lines carry the signal.
• In our own 2026-10-04 run, 487 of 1,504 reference passages sat at or above the 30% flag threshold. That is a passage count, not a false-positive rate.
• Unedited model output averaged 52% against 27% for the reference sample — a gap, not a verdict.
• Detection is one instrument. Turnitin, GPTZero, Copyleaks and Sapling each define their score differently, and none of them proves who wrote a text.
• The honest test is your own draft: paste it, read the flagged sentences, decide line by line whether the wording is yours.

Why does DeAIze flag writing that a person clearly wrote?

Because detection is pattern matching, not authorship forensics. Our detector is an ensemble that scores prose for the statistical habits machine-generated text tends to share: uniform sentence length, low burstiness, predictable word choices, few concrete nouns. Human writing produced under pressure can look identical. A grant abstract written to a template, a legal memo following house style, a methods section that must be impersonal — all of these read flat, and flat is what the model is looking for.

Our own measurement makes this visible. We took 1,504 passages from DeAIze guides that were not marked as generated — metadata selection only, and we did not independently verify the authorship of every passage. Their AI scores ranged from 1% to 100%, with a mean of 27% and a median of 20%. At a 30% flag threshold, 487 of those 1,504 passages would have been flagged. Those are passages, not independent documents, and they are not a false-positive rate. But the direction is clear: a meaningful share of ordinary human prose lands in the flagging band.

If that describes your situation, the useful next step is not arguing with the score. It is reading the sentences. Our guide on AI detection false positives walks through the exact triage we recommend.

What does the DeAIze detector actually return?

Two things, at different resolutions.

Output	What it shows	What it does not show
Document score	One AI-probability figure for the whole text	Whether any specific sentence was generated
Sentence highlighting	Each sentence ranked by AI probability	Authorship, intent, or whether AI was used at all
Humanizer suggestions	Revised wording for lines you choose	Proof that every meaning survived the rewrite

The sentence view is the part that matters for editing. A document score of any value is a summary; it collapses a long text into one figure and hides the fact that most documents are mixed. In our reference sample, scores spanned the full 1-100% range — the same corpus that averaged 27% contained passages at both extremes. A single number cannot tell you which lines to fix. Highlighting can.

That is also why we do not publish a headline accuracy percentage. There is no single number that survives contact with different text types, languages and lengths. What we can publish is the distribution, the sample definition and the limits, which is what the section above does.

If you want to see the sentence-level view on your own draft rather than read about it, the AI detector runs first and shows you the flagged lines before you change a word.

How big is the gap between human and machine text in our own data?

On the same 2026-10-04 run, we generated 12 passages from fixed prompts and did not edit them. Their AI scores ranged from 11% to 91%, with a mean of 52% and a median of 49%. Compare that with the reference sample: mean 27%, median 20%.

The gap is real, and it is smaller than most people expect. Notice the overlap. Some unedited model output scored as low as 11%. Some reference passages scored at the top of the scale. A threshold drawn anywhere in that overlap will misclassify text on both sides. Twelve passages is also a small sample — enough to show a direction, not enough to characterise every model or prompt style.

One more figure from the same run: we counted AI-tell frequency at 1 per 100 words in the reference sample and 1 per 100 words in the unedited model output. The surface tells people write about — em dashes, stock transitions, “moreover” — did not separate the two corpora in our data at all. If you have been editing out em dashes hoping to change a score, that result is worth sitting with. The signal lives in rhythm and specificity, not in punctuation.

Does the DeAIze humanizer work, and what does “work” mean here?

The humanizer suggests revised wording for sentences you select. It does not rewrite your argument, and it does not certify the result. Treat every suggestion as an edit proposal: read it against your original, and check that claims, quantities, citations, negations and qualifiers still say what you meant. Semantic similarity scoring is a review aid, not proof that meaning was preserved.

A worked illustration, written for this article and clearly not a measurement:

• Before: “The implementation of the new policy resulted in a significant improvement in overall operational efficiency across multiple departments.”
• After: “The policy cut our approval time from days to hours in three departments. Finance saw the biggest change.”

The second version is not better because a detector likes it. It is better because it names a department, a before and an after, and a specific consequence. That is the test to apply to every suggestion the humanizer offers: does the revision say something more concrete than the original, or does it just shuffle the same abstraction?

If the answer is “shuffle,” reject it. You can try the rewrite flow on the draft editor with 200 words after email confirmation, no credit card — enough to see whether the suggestions suit your writing before you commit to anything.

Why do different detectors disagree about the same text?

Because they are different instruments with different definitions. A model score is not a share of qualifying prose, and neither is a similarity percentage. When a Turnitin report, a GPTZero reading and a Copyleaks figure point in three directions, that is not evidence that one of them is lying. It means they measured different things on different scales.

Instrument style	Typical output shape	Common misreading
Document-level model score	One probability-like figure	Treated as a probability that AI wrote it
Segment or share-based output	Percentage of text meeting a criterion	Treated as a probability
Similarity or match score	Overlap with a reference corpus	Confused with AI detection entirely

None of these outputs proves authorship. That is not a DeAIze limitation; it is a property of the whole category. We do not provide an official Turnitin readout or a Turnitin score estimator, and we do not guarantee any particular result on another vendor’s tool. If you are trying to reconcile two conflicting reports, AI Detector Accuracy explains why they diverge and what to document.

When DeAIze is the wrong tool for the job

Three cases where you should not reach for it first.

1. An authorship dispute you need to win. Detection output is a signal, not evidence. If you are facing a panel, an editor or a client, the useful material is your drafts, your version history, your notes and your sources. Start with Academic Integrity Hearing AI rather than a score screenshot.
2. Very short text. A caption, a subject line, a two-sentence reply. There is not enough signal in a handful of sentences for any detector to say anything stable, and DeAIze is no exception.
3. Text you did not write. Rewriting someone else’s work to lower a flag is not editing, and it is not what the humanizer is for. The tool assumes the ideas, data and conclusions are yours.

There is also a limit we cannot engineer away: the reference sample in our own measurement was drawn from our own guides. It is not a random sample of the English-language web, and our passages are not independent of each other. Read the numbers as a description of our instrument’s behaviour on a known corpus, not as a general claim about all writing everywhere.

A five-minute check you can run on your own draft

1. Paste the text and read the document score once. Note it, then ignore it.
2. Sort the highlighted sentences by AI probability and read the top band out loud.
3. For each flagged sentence, ask: does this name a person, a number, a place or a decision? If not, it is probably abstraction, not authorship.
4. Rewrite only the sentences where the answer is no, in your own words and with your own specifics.
5. Re-run detection to see what moved. Do not chase the score; chase the sentences that stopped sounding like you.

That loop is what the tool is built for. It is also why we publish our own weak numbers instead of a marketing figure — a detector you cannot audit is a detector you cannot use well.

Ready to see which sentences read machine-made?

Reading about measurement only gets you so far. The fastest way to understand what DeAIze does is to run your own draft through it and look at the sentence highlighting. Start with the AI detector, then edit in the draft editor if the flagged lines need rewording. The rewrite trial covers 200 words after email confirmation with no credit card, and detection limits are shown in the current interface. If you want to know what you get before you spend anything, pricing lists every plan and register takes a minute. Keep the original. Compare every suggestion. The ideas have to stay yours.