A Claude detector built for the harder case.
Paste any text. Get a named verdict plus the sentences that still read like Claude.
50 free scans a day · 2,500 words per scan
Claude tells
What Claude writing looks like
Claude's prose is looser and more varied than the house style people associate with AI, and that is exactly what makes it slip past checkers tuned on ChatGPT output. Sentences run longer and lean on subordinate clauses; the register is conversational rather than institutional; the rhythm is less uniform. What remains detectable is a set of habits around stance — a persistent reaching for nuance, and a tendency to name the shape of its own argument as it goes.
Visible hedging as a value
Where ChatGPT hedges to stay balanced, Claude hedges to signal care: "it's worth being careful here", "that said", "the honest answer is". The qualification is the point rather than a reflex.
Em-dash asides
Parenthetical clauses set off with em-dashes, often stacked more than once in a paragraph, adding a qualification the sentence did not strictly need.
Structural narration
Sentences that describe the argument rather than making it — "there are two things going on here", "the useful distinction is" — signposting a piece that is not long enough to need signposts.
Symmetrical concessions
Counterarguments raised and answered in a tidy matched pair, each given the same weight and the same amount of space, which is rarer in genuine human argument than it looks.
These are patterns a careful reader can notice. The detector itself does not work from a list of banned words — it scores lexical and structural patterns across the whole document and every sentence in it, which is why rephrasing around a word like delve does not clear a document.
The hard part
Why Claude defeats older detectors
The first generation of AI detectors leaned on perplexity and burstiness — roughly, how predictable the next word is and how much sentence length varies. Those measures were built when model output was flatter and more predictable than human prose. Claude's longer, more varied, more digressive sentences score much closer to human on exactly those two axes, so a perplexity-based checker reads it as a person writing carefully. Scoring joint lexical and structural patterns instead is what keeps this case in reach, though it remains the hardest of the three models on this site.
How a scan runs
What the Claude detector actually returns
Paste the text — from about fifty words up to 2,500 in one free scan. The engine scores joint lexical and structural patterns across the whole document and every sentence in it, then reconciles those into one verdict: Human-Written, AI-Assisted, or AI-Generated. A mixed Claude draft does not collapse into a single misleading average.
You get 50 free scans a day with no account and no card. Text is scored in memory and discarded when the response returns. Nothing is retained for training. The privacy policy in the footer is the longer version of that sentence.
The percentage backs the verdict beside it — how human the passage reads for a Human-Written verdict, how likely it is that a model wrote it for an AI one. The verdict is calibrated per length band, so the same raw score on a short snippet and a long essay is not the same finding. Read the sentence shading before you act on the number — the reasoning is the part you can actually check.
No detector is proof of authorship. False flags cluster on formal prose, technical writing, and fluent English written as a second language. Missed detections cluster on heavily edited machine drafts. Treat a report on Anthropic's Claude as evidence to weigh, not as a finding about a person.
Claude is the hardest of the three major models on this site, and saying so is more useful than a rounded accuracy figure we do not have. Detectors built on perplexity and burstiness were calibrated against flatter output; Claude's longer, more varied sentences score closer to human on exactly those axes. This detector scores joint lexical and structural patterns instead, which is what makes the case tractable.
Scoring is based on writing patterns rather than a fingerprint of a specific Claude release, so the page is not pinned to one version. Confidence is lower on the newest frontier models than on older ones.
Text is scored in memory and discarded when the response returns. No account is required. Daily limits are shown on the scanner — they exist to keep the service up, not as a trial that expires.
Who this is for
When a Claude detector is the right tool
Use this page when the suspected source is Anthropic's Claude — the model whose prose looks most like a thoughtful person writing carefully. If a general checker came back human and you still have a reason to doubt, this is the case it was most likely to miss.
A document that reads too considered to be a first draft
Stacked hedges, em-dash asides, sentences that name the shape of their own argument. Quality is not the tell — caution as a rhythm is. That is the opposite of the ChatGPT problem, and it is why a GPT-tuned checker misses it.
Edited Claude, or Claude used as a copy-editor
A human draft 'cleaned up' by Claude often lands as AI-Assisted rather than AI-Generated. That is the correct description of the document, not a hedge. Sentence shading shows which passages still carry the machine stance.
A borderline result you need to argue with
Claude's register overlaps with careful human writing more than most models do. The sentence-level report exists so you can check the reasoning rather than take a number on trust. Borderline results on short or heavily edited text are evidence to weigh, not proof.
Not: a way to catch someone out
The overlap with thoughtful human prose is the risk this category has to take seriously. A false flag here falls on a person who writes carefully. Do not treat a report as the sole basis for an accusation.
Limits
What a report cannot do
We do not publish a per-model accuracy figure for Claude, or a head-to-head against GPTZero, Turnitin, Originality.ai, or anyone else. Vendor-measured numbers describe the vendor's own test set, not your document. The tells on this page are patterns a careful reader can notice; they are not how the detector works internally and not a banned-word list.
Do not use a report as the sole basis for a grade, a disciplinary outcome, or a hiring decision. Weigh it against drafts, version history, and a conversation with the writer. The FAQ at /faq covers this in more depth because it is the question that actually matters.
Short passages are genuinely hard. Below a few dozen words, any detector's output degrades toward noise. If you only have a sentence or two, the honest move is not to run a scan that will look more confident than it is.
FAQ
Claude detection FAQs
What the detector can tell you about Anthropic's Claude, what it can't, and how to read a result you're not sure about.
Can AI detectors detect Claude?
Many cannot reliably, and it is worth understanding why: detectors built on perplexity and burstiness were calibrated against flatter model output, and Claude's prose is neither flat nor especially predictable. This detector scores structural and lexical patterns jointly rather than relying on those two measures, which is what makes the case tractable — but Claude remains the hardest of the major models to call.
Which Claude versions does it cover?
Scoring is based on patterns in the writing rather than a fingerprint of a specific release, so it is not pinned to one version. As with any detector, confidence is lower on the newest frontier models than on older ones, and we would rather say that plainly than imply a guarantee.
Why did my Claude text come back as AI-Assisted rather than AI-Generated?
AI-Assisted is its own verdict, not a hedge. It means the document carries machine signals without reading as machine output end to end — typical of Claude text a person has edited, or of a human draft Claude rewrote. If you need the distinction, the sentence-level shading shows which passages carry the signal and which do not.
Does it confuse careful human writing with Claude?
That is the risk this category has to take seriously, because Claude's register overlaps with a thoughtful human writer's more than most models do. Detection is tuned to minimise false flags on genuine human writing, and the sentence-level report exists so you can check the reasoning rather than take a number on trust. Treat any borderline result on short or heavily edited text as evidence to weigh, not proof.
Is the Claude detector free?
Yes — no account and no card. Daily limits are shown on the scanner.