Claude detector for writers

Submitted text is analysed on our server and the result is prose: observable properties of the writing, with the limits of the reading stated alongside them. The analysis brief forbids claiming certainty about authorship, plagiarism or which model produced a text, and asks instead for observable signals and limitations. The output format follows from that constraint rather than hiding behind it.

Envelope: 5 runs per rolling 24 hours, 1 to 12,000 characters, an optional 300-character instruction, live character and word counts, no account, no retained history. There is no scoring subsystem on this route and no span-labelling layer.

Findings are only worth having if you can falsify them, so build that into how you run it.

Step 1

Submit a bounded passage you wrote, or edited, and can therefore verify against. An unfamiliar text gives you no way to test what comes back.

Step 2

Phrase the instruction as a hypothesis. "Are the paragraph openings uniform?" produces a checkable answer; "analyse this" produces a description.

Step 3

Falsify each finding in turn. The value of the run is the subset that survives, and a finding you cannot check is not evidence of anything.

The strongest argument against attribution comes from inside the category. AIDetego publishes a Claude detector page that describes 16 modules analysing statistical fingerprints, and on that same page states: "Exact attribution to a specific model (Claude vs. ChatGPT) is currently not possible", adding "no tool can reliably do this". Detector-checker.ai's Claude page says its tool "does not identify the source model" and cannot separate Sonnet, Opus and Haiku.

Plagly.ai goes further by undercutting the signals themselves. Having listed hedging, balanced framing, heavy structure and a warm measured tone, it concludes: "None of these is proof on its own, and careful human writers legitimately share several of them." Three independent vendors, three admissions, and the headline product in each case is a verdict built on the thing just disclaimed.

The published lists overlap enough to be worth taking seriously as descriptions of a style. As evidence of origin they fail; as a checklist for tightening prose they hold up.

Qualifier density

"It is worth noting", "generally", "in many cases", "it seems", "tends to". Measurable, removable, and a genuine drag on a paragraph's authority.

Clause-first openings

Sentences that establish conditions before asserting anything. The habit delays the claim and flattens the rhythm across a whole section.

Imposed structure

Numbered lists, headers and summaries applied to material with no natural divisions, plus a repeated thesis, counterargument, synthesis shape.

Variance suppression

Described by one vendor as consistently high readability throughout. The absence of a weak sentence is itself the pattern, which is why editing toward evenness makes it worse.

A percentage travels. It gets copied out of the tool that produced it, away from the caveat printed underneath, and into a message where it functions as a finding of fact. Nothing about the number resists that journey, and every vendor caveat in this result set lives on a page the number will leave behind.

Prose resists it better. A finding that says the middle section hedges every claim is checkable in situ and useless as an accusation, which is the correct combination for this class of inference. The trade is that the output is slower to read. That is a price worth paying on a route whose misuse is somebody's grade or job.

overchat.ai asserts that Claude models embed "an invisible statistical pattern" by "subtly influencing token selection during generation". That is a claim about another company's internals which we have no way to test, so we neither assert nor deny it, and nothing on this route detects anything of the kind.

The same page then makes the distinction its own competitors usually blur. Its scanner looks for zero-width characters, hidden Unicode, punctuation style and formatting residue, and it states: "Finding one of these artifacts therefore does not prove that text was written by Claude, and removing them does not necessarily remove Claude's watermark." Surface artifacts and a model-level pattern are different objects, and conflating them is how a formatting check gets sold as proof of origin.

Valid: on your own draft, where every finding can be checked against a document you understand, and where the output is an editing instruction rather than a conclusion about a person.

Not valid: as input to a decision about somebody else. The signals overlap with competent writing, the vendors disclaim attribution, and no output from this route is designed to survive being quoted at someone. That question is answered by process evidence and a conversation, which is slower and is the only method that actually addresses it.

No verdict, percentage, confidence score, probability band or per-sentence labelling. No model attribution of any kind. No accuracy or false positive rate, since none has been measured on this engine and importing a competitor's 99% figure would describe a different instrument.

No watermark detection, statistical or character-level, and no hidden character scan on this remote route. No language list, where one competitor names five. No file upload, no unlimited scans, no API, no integration, no account and no stored history.