AI Detection Software: What to Trust, and How to Cut the Rate

AI detection software assigns a probability that text looks machine-written — it does not prove who wrote it. If you have been flagged, the useful move is not arguing with the number. It is finding which sentences carry the signal and rewriting those, then re-scoring to confirm the rate actually dropped.

What does AI detection software actually measure?

Detectors do not recognise authorship. They estimate how predictable a passage is. Two signals do most of the work: perplexity (how surprised a language model is by your word choices) and burstiness (how much sentence length and structure vary). Text with low perplexity and flat rhythm reads as generated, whether or not a model touched it.

That is why formal academic prose, translated text and writing by non-native English speakers get flagged. Those styles are predictable by construction. If that is your situation, the honest framing is in why an AI detector flags non native English writers — the flag is a style reading, not an accusation.

We ran our own detector (DeAIze, local ensemble, lite tier, 30% flag threshold) on 2026-09-28 across two corpora. On 709 hand-written DeAIze guides the mean AI score was 26% and the median 25%, with a range of 0–69%; 239 of those 709 human documents were flagged at that threshold. On 12 unedited model passages the mean was 33% and the median 34%. Two things follow. Human writing is not automatically safe, and unedited model output is not automatically caught. Both directions of error are normal.

Vendors publish their own tolerances, and they differ. GPTZero’s FAQ states that when its confidence category is high, 99.1% of human articles are classified as human and 98.4% of AI articles as AI, and it describes a 90% class probability as meaning the detector is correct on similar documents about 90% of the time (source: gptzero.me). Turnitin’s product page says it may flag one in every 100 fully human documents, that no score is attributed between 1% and 19%, and that a document it identifies as 50% likely AI-written could contain as much as 65% AI writing (source: turnitin.com). Read those as vendor-defined operating bands, not as a guarantee about your file.

Which detector should you actually check?

Match the tool to the decision. A free sentence-level check is enough for editing; a report a school or client will read is a different job.

Tool	What it reports	Notable published detail	Best for
DeAIze (ours)	Whole-document score plus sentence-level AI probability	Own measurement, 2026-09-28: human sample n = 709, mean 26%, median 25%, 239 flagged	Finding which sentences to rewrite
GPTZero	Document score, sentence highlighting, confidence category	Vendor FAQ: at high confidence, 99.1% of human articles classified human, 98.4% of AI articles classified AI; 90% class probability described as correctness on similar documents (gptzero.me)	A second opinion on an already-flagged draft
Turnitin	Segment-level indicator in an institutional report	Vendor page: about 1 in 100 fully human documents flagged; no score attributed from 1% to 19%; a 50% finding could reflect as much as 65% AI writing (turnitin.com)	Understanding what your institution sees
Copyleaks, Originality.ai, Sapling	Document and segment scores	—	Cross-checking before you submit

Before you rewrite anything, find out which lines are carrying the signal. DeAIze scores the document and highlights each sentence by AI probability, so you edit the flagged sentences instead of the whole draft — free tier, 200 words, no card.

Why did one tool flag you and another clear you?

Because they are different instruments reading the same page. Thresholds, training data, segment length and language handling all differ, so disagreement is expected rather than evidence of a bug. A short passage makes it worse: there is less text for the model to estimate from, which is why a three-line paragraph can swing wildly between checks.

Practical rule: pick one detector as your working instrument, note its score, and stop shopping for a friendlier number. Chasing a clean read from a fifth tool teaches you nothing about your prose. If two tools disagree badly, that disagreement is itself the finding — the text sits near a boundary, and the boundary is soft.

How do you bring a flagged rate down?

The work is sentence-level editing, not synonym-swapping. Paraphrase tools rotate vocabulary while keeping the same flat rhythm, which is the thing being measured.

1. Score first. Get a document score and a per-sentence reading. Do not edit blind.
2. List the flagged sentences. Usually a minority. Keep the rest untouched.
3. Rewrite for rhythm, not vocabulary. Mix a short sentence in. Split a long one. Let one sentence carry an opinion or an aside.
4. Add the specifics only you have. A date, a place, a number from your own data, a caveat you actually believe. Detectors read generic phrasing as low-perplexity text.
5. Re-score. If the number did not move, the rewrite was cosmetic. Try again on the same sentence.
6. Check meaning survived. Compare before and after; if a claim shifted, revert.

Here is an illustration, not a measured result. Before: “The implementation of renewable energy solutions is essential for the mitigation of climate change impacts across multiple sectors.” After: “Renewable energy is not optional any more. In our region the grid added more solar than gas last year, and the planning queue is the bottleneck, not the technology.” Same argument, specific detail, uneven rhythm. That is the shape of an edit that moves a score — though no rewrite guarantees a particular number on any detector.

What does a high score not mean?

It does not mean you cheated, and it does not mean you are safe if you did. Detector output is a probability estimate. GPTZero’s own FAQ frames a 90% class probability as correctness on similar documents, not certainty about yours (gptzero.me). Turnitin’s page concedes that a 50% finding could reflect as much as 65% AI writing — a range, not a measurement (turnitin.com).

Our 2026-09-28 run makes the same point from the other side: 239 of 709 human-written guides were flagged at a 30% threshold, and 12 unedited model passages averaged 33%, with one as low as 4%. Flags fall on human work, and machine work slips through. If you are facing a panel rather than a score, what an academic integrity hearing assesses covers what evidence actually matters.

The uncomfortable part: a low score is not proof of anything either. It means the text did not trip that instrument on that day.

When this is the wrong choice

If the flagged text is not yours — a passage you pasted from a source, a section a colleague generated and handed over — do not humanize it. Rewriting someone else’s work to slip past a check is the thing detectors exist to catch, and it is not what these edits are for. Similarly, if your institution forbids AI assistance outright, editing the output does not make the draft compliant. The rules come first; the tooling is downstream.

Skip the paid route too if you only need to know whether a sentence reads oddly. A free sentence-level check answers that. Paying makes sense when you have a document that has to clear a specific reader’s check and you need the rewrite loop run and verified.

One more honest limit: no tool can make machine-written ideas yours. The arguments, the data and the conclusions have to be your own before any editing pass is worth running.

A five-minute routine that lowers the rate

Score the document. Read the flagged sentences aloud — your ear catches flat rhythm faster than any score. Rewrite three of them with one concrete detail each. Re-score. If the number dropped, you have your method; if it did not, your text may simply sit in a band where that detector is unreliable, and the right answer is a second opinion rather than a rewrite spiral.

Keep a copy of the before text. If anyone asks, showing what you changed and why is a stronger position than a clean score with no trail behind it.

---

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/.