AI Writing Signal Analyzer

Find the stylistic markers of AI-generated writing — without pretending it is proof

Paste a passage and this tool highlights, in place, the specific features in it that appear more often in language-model output than in general human prose: the invisible characters, the unspaced em dashes, the “it’s worth noting” constructions, the delve-tapestry-testament vocabulary, and the unnaturally even sentence rhythm. Every highlight is clickable and every one tells you why it was flagged and what would legitimately produce it anyway.

What it does not do is give you a verdict. There is no “87% AI” number here, because no tool of this kind can produce one honestly, and pages that do are selling you a confidence they do not have. If you have arrived here trying to decide whether a student cheated or a contractor used ChatGPT, read the next section before you do anything else.

What a stylistic analyser cannot tell you

Three limits are structural, not fixable with a better model:

• Nothing here detects a watermark. Statistical watermarking schemes are keyed — verification requires the secret held by the model vendor. A third party cannot check for one, so any tool claiming to detect AI “watermarks” from the outside is describing something it is not doing.
• Every marker has an innocent explanation. “Robust” and “leverage” are ordinary English words. Curly quotes come free from Word and Google Docs. Bold-lead-in bullet lists are house style at half the companies in the world. Density across categories is the only thing that carries any weight, and even that is circumstantial.
• The false positives are not random — they land on specific people. The strongest statistical signal available without model access is sentence-length variance, and it is systematically lower in non-native English writing, in translated text, in technical documentation and in anything edited to a house style. That is the mechanism by which commercial detectors end up accusing ESL students at a far higher rate than native speakers. The tool is built to make that visible rather than to hide it.

So the honest framing is: these are signals to look at, not evidence to act on. Used to audit your own draft before publishing, or to understand why a piece of text reads as machine-written, it is genuinely useful. Used against a person, it is not.

What it actually measures

The analysis is a marker catalogue plus three whole-text statistics. Markers are pattern matches; each has a weight and a cap, so a single word repeated forty times cannot dominate the result. They fall into five categories:

The three structural statistics are computed across the whole passage rather than counted as events:

• Sentence-length variance (burstiness). The coefficient of variation of sentence lengths. Human prose usually sits above 0.50; this flags below 0.42. Sampling from a probability distribution smooths out the mix of very short and very long sentences that people naturally produce. This is the best-studied marker available without access to the model — and the one with the worst false-positive profile.
• Paragraph-length uniformity. Models tend to emit paragraphs of near-identical size. Weak on its own; templated and CMS-authored writing looks the same.
• Contraction rate. Contractions per thousand words. Formal model output avoids them more consistently than people do. Genre dominates this one entirely — academic and legal writing is legitimately contraction-free.

The 120-word floor, and why the score is deliberately conservative

Below 120 words, no score is produced at all. Individual matches are still highlighted so you can look at them, but the number is withheld, because ordinary human writing trips the same markers by chance at that length. Refusing to score short text is the single most important guard in the tool and the thing most consumer detectors get wrong.

Above the floor, the result is a bounded index out of 100, split into a marker component worth up to 60 and a structural component worth up to 40. Two design choices matter:

• Markers are scored as a rate, not a count. Weighted hits are normalised per thousand words. Without this, a long document accumulates markers simply by being long.
• Structure corroborates; it does not accuse. If the structural signals fire but no lexical or phrasing evidence supports them, their contribution is scaled down sharply. This came out of testing against a human-written incident postmortem that tripped all three structural signals — flat rhythm, even paragraphs, zero contractions — with not one lexical marker present. Formal human writing genuinely looks like that, and letting structure alone drive the verdict is exactly how detectors end up accusing technical authors and ESL writers.

The result is presented as a band — few, some, many, heavy — with the arithmetic shown underneath, including when the structural component was reduced and why. It is not a probability. A 40 does not mean “40% AI”, and the interface says so on every result.

Fixing a draft, not judging one

The most defensible use of this tool is on your own writing. Every flagged span that has a sensible replacement is grouped, so one click fixes every occurrence of the same term at once: unspaced em dashes become commas, curly quotes become straight ones, zero-width characters are deleted outright, and overused vocabulary gets plain-language alternatives. There is an apply-all for the mechanical categories and a multi-step undo, and once you make an edit the score shows what it was before alongside what it is now, so you can see which changes actually moved it.

One category deliberately offers no fix: structure. There is no find-and-replace for flat sentence rhythm. Varying it means rewriting, and offering a mechanical substitution would just be teaching the tool to defeat its own strongest signal. If the structural signals are firing on your draft, the fix is to break up the long sentences and let a short one land on its own.

Model family attribution, and why it is weak

Some markers lean toward one model family — unspaced em dashes and reflexive qualifier openers toward Claude, emoji-decorated headings, assistant pleasantries and “landscape” clichés toward GPT. The tool reports a lean only when the family-specific evidence is both plentiful enough and lopsided enough to say anything, and it refuses far more often than it commits.

Treat any lean it does report as the weakest output on the page. Both families have converged heavily on the same register, and these attributions shift with every model revision — they are the part of the marker catalogue that ages fastest. It is a curiosity, not an identification.

Questions people ask before trusting a result

• Can I use this to prove someone used AI? No, and you should not try. Markers are removed by an afternoon of editing, and they are produced by plenty of human writing. A high band is a reason to have a conversation about a piece of work, never a finding on its own.
• Why did my own writing score badly? Almost always one of three things: it is formal, so the contraction and rhythm signals fire; it went through Word or Google Docs, so the typography markers fire; or it is technical, so several of the overrepresented words are simply the correct words for the subject. Look at which categories are lit up rather than at the total.
• Does editing the text to lower the score make it undetectable? It makes these markers go away, which is the point when you are polishing a draft. It says nothing about how the text was produced, which is the honest reason a low score is not evidence of human authorship.
• Is my text uploaded? No. Analysis, highlighting and every replacement run entirely in the browser.
• How current is the marker list? The shared markers — the assistant register both families converged on — are the stable ones. The family-specific attributions drift with each model revision and are the first thing to distrust.

What the highlights are really for

The output that matters is not the number, it is the marked-up text. Hovering a highlight shows the marker’s name and the reason it exists; each marker also carries its own caveat explaining what innocently produces it. Invisible characters, which have no glyph, are rendered as a visible stand-in so you can actually see where they are — usually the single most useful thing on the page, because a zero-width space pasted into a CMS or a code block causes real problems entirely separate from where the text came from.

Read the flags, decide whether each one is genuinely odd or just the subject matter, and form your own view. That is the whole design. Everything runs in your browser — the text is never sent to a server — so you can paste an unpublished draft, a candidate’s cover letter or an internal document without it leaving your machine. Load the built-in example if you want to see what a heavily-flagged passage looks like before committing your own text.