How Do AI Detectors Work When Your Text Reads Machine-Made?

How do AI detectors work? They estimate how predictable your word choices are. A language model predicts the next token from everything before it; detectors run a similar calculation and ask how surprised the model was by what you actually wrote. Predictable, evenly weighted prose scores high. Varied, specific prose scores lower. You get a probability, not a verdict.

Key takeaways

• Detectors score predictability, not authorship. A high score means “this looks like model output”, not “a model wrote this”.
• Two signals do most of the work: perplexity (how surprising each word is) and burstiness (how much sentence length and rhythm vary).
• Scores are computed per segment, then averaged. One flat paragraph can carry a whole document.
• Human writing flags too. In our own measurement on 2026-09-19, 46 hand-written DeAIze guides averaged a 27% AI score, with 17 of 46 crossing a 30% flag threshold.
• The fix is editorial, not adversarial: cut the tells, restore variation, keep your meaning.
• Never treat a detector score as proof of anything. It is a probability estimate.

What is actually being measured?

A detector does not search for a fingerprint that says “GPT” or “Claude”. It measures statistical texture.

Language models are trained to pick likely next words. So model output tends to sit in the high-probability lane: common phrasing, safe transitions, balanced clauses. Detectors approximate that same probability calculation and report how predictable your text is. Predictable text reads as machine-made because it is machine-shaped, whether a machine wrote it or not.

That distinction matters for you. If your own draft reads machine-made, the detector is not accusing you of anything. It is telling you the prose has flattened into the same statistical lane that model output occupies.

How does perplexity make a sentence look machine-made?

Perplexity is a measure of surprise. Low perplexity means each word was easy to guess from the words before it. High perplexity means the reader (and the model) had to work.

Consider two ways to say the same thing:

• Low surprise: “Effective communication is essential for success.”
• Higher surprise: “Nobody reads the second paragraph of a memo. Write the first one as if that’s true.”

The second version contains a specific image, a contraction, and an opinion. The first contains none. Detectors notice.

This is why generic filler is the most dangerous thing in a flagged draft. Every “furthermore”, every “in order to”, every three-item list of abstract nouns pushes the text toward the predictable lane. It reads smooth because it says almost nothing.

Why does burstiness matter more than vocabulary?

Burstiness is variation. Humans write in bursts: a long, clause-heavy sentence, then a short one. Then a fragment. Model output tends toward a steady middle length, because that middle length is statistically safest.

You can see it without any tool. Count the words in ten consecutive sentences of your draft. If they all land between 18 and 24 words, the rhythm is flat, and flat rhythm is one of the strongest signals a detector picks up.

Burstiness also covers punctuation and structure. Uniform paragraph lengths, identical transition words opening every section, and a tidy conclusion that restates the introduction all read as generated. Editors notice the same things before any detector runs.

If you want to see which of your own sentences carry the signal, run the draft through the AI detector first. Sentence-level highlighting turns a vague score into a short list of lines you can actually edit.

Why does my own writing get flagged?

This is the part most guides skip, and it is the part you probably care about.

In our own measurement on 2026-09-19, we scored 46 hand-written DeAIze guides with our detector at a 30% flag threshold. The mean AI score was 27%, the median 25%, and the range ran from 0 to 55. Seventeen of the 46 crossed the threshold. These were human-written articles by people who write for a living.

The same run scored 12 unedited model passages at a mean of 35%. The two distributions overlap. That overlap is the whole story: on short or formulaic text, human and machine prose can land in the same band.

Common reasons your own writing drifts into that band:

1. Template sentences. “In this section, we will explore…” is a shape, not a thought.
2. Hedging stacks. “It could potentially be argued that…” removes the writer from the sentence.
3. Even rhythm. A long career of writing formal prose trains you into uniform sentence length.
4. Second-language caution. Non-native writers often choose the safest grammatical option, which is also the most predictable one. There is a whole guide on why an AI detector flags non native English writers if that is your situation.
5. Heavy editing. Stripping every quirk in the name of polish can strip the signal too.

None of that means you cheated. It means your prose is predictable in places. Predictability is editable.

What do the different detectors actually report?

Tools differ far more in presentation than in underlying method. Here is a fair comparison of what each one shows you.

Tool	What it reports	Useful for
GPTZero	Document score plus sentence highlighting	Seeing which lines carry the signal
Turnitin	Segment-level percentage in an institutional report	Understanding what a marker sees
Copyleaks	Document score with per-section breakdown	Locating the flat stretch in a long draft
Originality.ai	Document and paragraph scores	Editorial triage at volume
Sapling	Sentence-level scores in a writing interface	Editing while you write
DeAIze	Whole-document score plus per-sentence AI probability, with cross-checks against third-party detector APIs where configured	Deciding which sentences to rewrite

Two tools can disagree on the same paragraph and both be doing their job. Each uses its own model and its own threshold. A score is a probability estimate under one set of assumptions, not a fact about your document. We wrote a longer piece on why two tools disagree if you want the mechanics.

What is the fastest way to fix a flagged draft?

Do not paraphrase everything. That is how meaning gets lost and how a draft ends up sounding like a thesaurus had a bad day.

Work like this instead:

1. Score the draft and look at sentence-level highlighting, not just the headline number.
2. Fix the top flagged sentences first. Usually a handful of lines carry most of the signal.
3. Break the rhythm. If three flagged sentences in a row are all the same length, make one of them five words.
4. Replace abstract nouns with concrete ones. “Effective communication” becomes the actual thing you did.
5. Delete the scaffolding. Most “it is worth noting” constructions can simply vanish.
6. Re-score. Check that meaning survived, not just that the number dropped.

Here is a worked illustration. The before line is generic; the after line keeps the same claim but adds a specific, checkable detail and a shorter rhythm.

• Before: “Regular revision significantly improves the overall quality of written work.”
• After: “Revise twice. The second pass is where the argument usually appears.”

Same point. Different texture. The second version is harder to predict word by word, which is exactly what the detector is measuring.

If you would rather have that pass done for you, DeAIze’s draft editor scores the text, rewrites only the flagged sentences, checks semantic similarity so meaning survives, then re-scores and shows you before and after. On the example paragraph on the site, a 92% AI score becomes 4% AI. Across the site’s own runs, the reported average reduction is 60%+, best case 71% — those are our predicted scores, not a verdict from any third-party detector, and no tool can promise you a particular number on someone else’s system.

Where do AI tells come from in the first place?

Surface habits give a draft away long before statistics do. In the same 2026-09-19 measurement, we counted tell frequency: 3 per 100 words in the human sample versus 1 per 100 words in the unedited model output. That number is worth sitting with, because it inverts the usual assumption. The human-written guides contained more of the classic tells — em dashes, “moreover”, tidy triads — than the raw model output did.

The lesson: tells are a style habit, not a species marker. If you write formal English for a living, you have probably absorbed several. The em dash tell is a good example of a habit that gets blamed on models but shows up everywhere in professional prose.

Read your draft for these, in order:

• Openers that announce instead of say something.
• Transition words doing grammatical work but no logical work.
• Three-item lists where two items would do.
• Paragraphs of identical length.
• A closing paragraph that repeats the opening.

Fix those and the statistical texture usually follows, because the two are the same phenomenon seen from different angles.

What should you not do?

Do not paste text into a detector and treat the output as evidence about a person. Detector scores are probability estimates. Institutions that treat them as proof create false-positive cases, and there is a documented pattern of exactly that.

Do not rewrite until the meaning changes. If a sentence no longer says what your data says, you have traded one problem for a worse one. Researchers and analysts should read how to humanize text without changing meaning before touching a technical draft.

Do not use any of this to pass off someone else’s work. The ideas, data and conclusions have to be yours. Detection tools exist to reduce false positives on your own writing and to clean up your own AI-assisted drafts, and that is the only frame in which they make sense.

And do not assume that a low score means you are safe from scrutiny. A reader who finds your draft thin will say so whether or not a detector ever ran.

Which sentences should you rewrite first?

Start where the signal is concentrated. Sentence-level highlighting exists precisely so you do not have to guess.

A practical priority order:

1. Sentences flagged high and sitting in your thesis statement or conclusion. Those carry the most weight with a reader.
2. Runs of three or more consecutive flagged sentences. That is a texture problem, not a word problem.
3. Any flagged sentence that is also your longest. Length plus predictability is the strongest combination.
4. Anything you did not write yourself. If a paragraph came from a model, it needs an editorial pass regardless of its score.

Leave the low-signal sentences alone. Rewriting text that was never a problem is how drafts lose their voice.

Ready to see which sentences read machine-made?

Understanding how detectors work is useful. Seeing it on your own draft is what changes your editing.

DeAIze is a web app with two halves: a detector that scores the whole document and highlights each sentence by AI probability, and a humanizer that rewrites only the flagged lines, keeps a rewrite only when semantic similarity says the meaning survived, then re-scores and reports before and after. Uploads handle .txt, .docx and .pdf, and rewriting a Word document preserves fonts, styles and tables. Detection and rewriting cover English and 9 other languages, the interface ships in 10 languages, and your text is processed for the current task only — never used for training, never resold.

You can start on the free tier with 200 words on signup, no credit card. If you want to see the numbers first, pricing is pay-as-you-go: Starter at $20.99 for 1,000 words, Standard at $44.99 for 3,000 words, Pro at $99.99 for 10,000 words, and Max at $199.99 for 30,000 words. Credits never expire and there is no subscription. Run your draft through the AI detector, open it in the draft editor, and fix the lines that actually carry the signal.