Updated August 2026 · Anthropic's marking policy read first-party
Claude AI detector: what a Claude humanizer changes and what a Claude AI checker actually reads
Nothing currently sold as a Claude detector can tell you that Claude wrote a passage. In August 2026 Anthropic added the one signal that could, and then did not ship the reader for it.
2 free runs a day, up to 200 words each. We save your run so you can get back to it, and a delete button appears with the result. See privacy.
This is our own AI-pattern score, measured here on sentence rhythm, template phrases, vocabulary variety and passive voice. It is not a GPTZero, Turnitin, Originality.ai, Copyleaks or ZeroGPT result, and it does not predict one. Worth knowing: we also ask the rewrite to vary sentence length, drop template phrases and prefer the active voice, so some of the drop is built in. Read the two panels below, not just the number.
Before ·
After ·
Plans planned from $9/month billed yearly · Your runs saved to your history, delete any time
The short answer
No tool sold as a Claude AI detector identifies Claude. Every one of them is a general AI-likeness classifier that returns a probability about writing style, and a style score cannot name the model that produced the text. One thing did change in August 2026: Anthropic now embeds an invisible watermark in text from Claude models launched on or after August 2, 2026, and attaches signed C2PA metadata to generated files. That mark is the only Claude-specific signal that exists, Anthropic has not yet published the tool that reads it, and Anthropic says a detected mark "is not fully conclusive" about authorship anyway.
Last updated August 20, 2026. Every quotation on this page was read on the vendor's own documentation on the date shown beside it.
The comparison people actually need
What each "Claude AI detector" is really measuring
The column that matters is the third one. It is "no" for every commercial tool on the market, including the expensive ones.
| Tool | What it measures | Can it identify Claude? | Can you run it today? |
|---|---|---|---|
| Anthropic itselfChecked Aug 20, 2026 | Since August 2026 Claude embeds an invisible watermark in generated text, plus C2PA signed metadata on generated files. | Yes, in principle. This is the only mechanism that is actually tied to Claude rather than inferred from style. | Not yet. Anthropic says it will "share details on detection mechanisms in forthcoming technical documentation". There is no public reader today. |
| GPTZeroChecked Aug 20, 2026 | A statistical classifier over prose. Returns HUMAN_ONLY, MIXED or AI_ONLY with a probability and a confidence category. | No. Its FAQ says it "works robustly across a range of AI language models, including but not limited to ChatGPT, GPT-5, GPT-4, GPT-3, Gemini, Claude". Coverage is not attribution. | Yes, free tier plus paid plans. Tells you AI-likeness, not authorship. |
| TurnitinChecked Aug 14, 2026 | An institutional AI writing indicator inside the similarity report. Scores prose sentences in paragraphs only. | No. Turnitin has never claimed to name the model, and it will not even name the rewriting tools it says it detects. | Not by you. Turnitin sells to institutions, and the AI indicator is visible to instructors and administrators, not to students. |
| Copyleaks, Originality.aiChecked Aug 10, 2026 | Commercial classifiers aimed at publishers, agencies and academic integrity teams. | No. Both market model names in their copy. Both output an AI probability, not a model label. | Yes, paid. Originality.ai from $12.95 a month annually, Copyleaks from $13.99. |
| Sites branded "Claude AI detector"Checked Aug 20, 2026 | Generic AI-likeness checkers behind a Claude-specific landing page. Several are one page in a set that also has a GPT page and a Gemini page. | No. The page is segmented by keyword, not by classifier. The same text scored on the ChatGPT page of the same site typically returns the same number. | Yes, usually free. Treat the Claude branding as marketing. |
| QuillBot, Grammarly free checkersChecked Aug 20, 2026 | Free web detectors from companies that also sell a rewriter, so they are paid on both sides of the question. | No. Same general AI-likeness score. | Yes, free. |
There is a simple test for any site branded as a model-specific detector. Paste the same paragraph into its Claude page and its ChatGPT page. If the number comes back identical, and it usually does, you are looking at one classifier behind two landing pages. That is not fraud, it is keyword segmentation, but it does mean the word "Claude" in the product name carries no technical meaning.
The reason no classifier can do model attribution is structural rather than a gap someone will close next quarter. Detectors read finished text for statistical properties: how predictable each word is given the ones before it, how much sentence length varies, how often stock connectives appear. Those properties describe a writing style. Several models trained on overlapping data produce overlapping styles, so a style measurement cannot resolve to a vendor. We took the mechanism apart in how AI detectors work, including the widely misread meaning of a percentage score.
The accuracy question sits on top of that one. Independent testing has repeatedly put real-world detector performance well below vendor marketing, and a peer-reviewed study of 190 US undergraduates found detectors misread genuine student work about 1.3 percent of the time while human graders misread it 5.0 percent of the time. The full evidence base, including the studies people cite without reading, is on the AI detector false positive rate page.
New in August 2026
Anthropic now watermarks Claude text, and it matters less than it sounds
Read from Anthropic's own help documentation on August 20, 2026. Every answer below is quoted or paraphrased from that page.
Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and it is implementing that commitment globally rather than only for European users. In Anthropic's wording, when a supported Claude model generates text "it weaves an imperceptible watermark directly into the text itself", and because the watermark is part of the text "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing".
That is a genuine change to this subject, and it is the first time any major vendor has shipped text watermarking at scale. OpenAI has said it developed a text watermarking method and continues to weigh alternatives without shipping one, which we covered in whether ChatGPT watermarks its text. Google watermarks Gemini output with SynthID. Anthropic has now joined the second group, and the practical consequences are narrower than the headlines suggest.
Which Claude models carry a mark?
Models "launched on or after August 2, 2026 support marking at launch". For anything released before that date, Anthropic says "Existing models are in progress" under a legal transition period.
Where does it apply?
Everywhere. Anthropic lists "Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag", plus access through AWS, Google Cloud and Microsoft Foundry, and says marking applies "wherever Claude is offered, worldwide".
Does a detected mark prove AI authorship?
No. Anthropic states that a mark "provides a signal that content was processed by Claude, but is not fully conclusive", and gives proofreading, translation and summarizing as cases where your own words carry a Claude mark.
Does no mark prove human authorship?
No. Anthropic lists four reasons a mark can be missing: an older model, heavy editing or paraphrasing, a passage too short for a reliable signal, or a surface that does not support marking.
Can anyone read the mark today?
Not publicly. Anthropic says it will "support users and other third parties to detect Claude's marks" and will publish details later. Until then, no checker you can run verifies a Claude watermark.
Is this the same as Google SynthID?
Same idea, different key. Both bias word choice so a pattern is recoverable later, and in both cases verification needs the vendor. Neither is readable by a third-party detector site.
The limitation Anthropic states most plainly is the one nobody quoting this story has picked up. A mark means the text "may have been processed by Claude", and Anthropic gives the obvious counterexamples itself: people use Claude to proofread, translate and summarize, so "the output can carry a Claude mark even if the underlying ideas, text, or data originated from another source". Your own paragraph, pasted in for a grammar pass, comes back marked. Anyone who treats a future watermark hit as proof of AI authorship will be wrong on a large fraction of cases.
The reverse error is just as easy. Anthropic writes that a lack of a detected mark "doesn't mean the content wasn't AI-generated or processed", and lists the reasons: the model predates marking, the passage is too short for a reliable signal, the surface did not support marking, or the text "has been heavily edited, paraphrased, translated, or mixed into other writing".
We should be direct about that last clause, because every competitor in this category is about to start advertising against it. We do not claim to remove Anthropic's watermark, we have not tested against it, and no reader exists for anyone to test against. What our rewriting does is change prose: rhythm, register, sentence construction. If you want a defensible position on provenance, the answer has never been a tool. It is a record of how the document was written, which is the case we make in how to prove you did not use AI.
Working with a Claude draft
Four steps, in the order that actually works
Step 1
Separate the two questions before you do anything
One question is whether a classifier will call your text AI-like. The other is whether Claude touched the file at all. They have different answers, different evidence, and only the first one is affected by how you write.
Step 2
Score the draft on a general detector, not a Claude-branded one
A Claude-branded checker runs the same classifier as its ChatGPT-branded sibling. Use GPTZero or Copyleaks directly, read the confidence category rather than the headline number, and remember the score is about style.
Step 3
Rewrite for rhythm and register, then read it aloud
Claude drafts tend to be long, evenly paced and heavily qualified. Uneven sentence length and fewer hedges do more for how the piece reads than any score, and reading it aloud finds the flat stretches faster than a tool does.
Step 4
Add what the model could not know
One real number, one thing you saw happen, one example from your own work. No rewriting tool can supply this, and its absence is what an experienced reader notices long before a detector says anything.
Step three is where a Claude draft differs from a ChatGPT one. Claude tends to produce longer, more even, more heavily qualified prose, with a caveat attached to claims that do not need one. Cutting the hedges is usually the single highest-value edit, and it makes the argument stronger as well as less uniform. The model-by-model differences are set out in which AI is least detectable, and the ChatGPT version of this page is the ChatGPT humanizer.
Who this is for
Where a Claude humanizer earns what it costs
And where it does not. If none of these describe you, editing the draft yourself does the same job for nothing.
Writers drafting in Claude
You use Claude because the long-form output is good, and you want the published version to read like you rather than like a careful assistant.
Agencies with a disclosure policy
Your client contract says what AI involvement has to be declared. The August 2026 marking change is a compliance fact you should know before it comes up in a review.
Anyone told "the detector says this is Claude"
No commercial detector can support that sentence. Knowing exactly why is the most useful thing on this page, and it is the part people get wrong in both directions.
Teams standardizing on one model
Switching drafting models to chase a lower score is the wrong lever, and it is now a worse lever than it was, because the marking landscape differs by vendor.
Editors reviewing submitted copy
You want to know what a score actually licenses you to say. Mostly: less than the number suggests, and nothing at all about which model was used.
Researchers and technical writers
Claude output in a structured document often trips detectors for reasons unrelated to AI, because procedural and repetitive text scores badly on every classifier.
Students are the group with the most at stake and the least control, because the detector their work is scored against is one they cannot run and cannot see the output of. That is a governance problem more than a writing one, and it has its own page in the AI humanizer for students. Teams publishing at volume have a different problem, consistency across writers and a review workflow that survives it, covered on the AI humanizer for content teams page.
Before you pay for anything
Three claims we will not make about Claude
We do not claim to strip Anthropic's watermark. It went live this month, no public reader for it exists, and any vendor already advertising watermark removal is describing a test nobody can currently run. Our tool rewrites prose. That is a different operation from defeating a cryptographic provenance signal, and we would rather say so than sell you a claim that cannot be checked.
We do not promise a detector score. Nobody can. Turnitin sells only to institutions and exposes no public API, so every "guaranteed to pass Turnitin" figure in this market is an estimate dressed as a measurement. Turnitin itself says it trains against leading rewriting tools and refuses to name them, specifically so that no vendor can know whether it is on the list. What we can describe honestly is the mechanism: GPTZero states that its classifier is not trained to identify AI text after heavy modification, and that is why editing changes scores.
We cannot make a Claude draft knowledgeable. A rewrite improves rhythm and register. It does not add evidence, and it does not repair an argument that was thin when Claude produced it. It also does not fact check: if the draft contains an invented citation, the rewritten version contains the same invented citation in better prose, which is worse, because it now reads convincingly. That problem is covered in whether humanized text can still be detected.
What it costs
Sizing a plan against what you actually publish
Our plans are $9 a month for 15,000 words, $19 for 80,000 and $39 for 250,000, all billed annually, or $12, $24 and $48 month to month, and they are planned rather than on sale. The free account gives you 3 runs a day and 9,000 words a month while you decide. Annual billing carries the discount across this entire category, which is why almost every low headline price you see quoted assumes a twelve month commitment, including ours.
Size the tier from output rather than ambition. A 1,200 word article consumes roughly 1,200 words of allowance, so 15,000 words is about a dozen pieces a month. If you draft long in Claude and cut hard afterwards, count the draft, not the published length. Undetectable.ai starts lower than we do at $5.00 a month billed annually for 10,000 words and includes API access on every paid tier, which we do not, and if programmatic access on a small budget is your requirement that is the honest recommendation. The full ladder is on Undetectable.ai pricing, and the paid options are compared on the best AI humanizer page.
Frequently asked
Claude AI detector questions, answered
Can AI detectors detect Claude?
They can flag Claude output as AI-like, and they do so imperfectly. What none of them can do is tell you Claude specifically wrote it. Every commercial detector returns a probability that text resembles machine writing in general. GPTZero says it covers Claude among other models, which means Claude text is in scope, not that the tool identifies the model.
Why doesn't Claude show up in AI detectors?
Usually it does show up, and the premise is shakier than it looks. Claude text tends to be longer and more varied in construction than a short ChatGPT reply, which can lower a perplexity-style score, but no published, checkable test establishes a per-model detection rate. The confident percentages you see traded online trace back to vendor blogs, not to research.
Can third-party detectors identify Claude generated text now?
No, not as of August 2026. Anthropic began embedding an invisible watermark in Claude text this month, and that mark is the only Claude-specific signal that exists. Anthropic has not released the tool that reads it, saying details will come in forthcoming technical documentation, so no third-party detector can check for it yet.
Does Claude Sonnet get caught in AI detectors?
At roughly the same rate as other current models, which is to say inconsistently. Detector accuracy varies far more with text length, subject and how much you edited than with which model produced the first draft. Short passages score unreliably on every classifier, and GPTZero states plainly that its accuracy rises as more text is submitted.
Can AI detectors catch Claude humanized text?
Sometimes, and less often than unedited output. GPTZero says directly that its classifier "is not trained to identify AI-generated text after it has been heavily modified after generation". That is the honest mechanism behind every rewriting tool in this category, and it is a statement about training data, not a guarantee about your document.
Is there a free Claude AI detector?
There are free checkers marketed as Claude detectors, and they are general AI-likeness classifiers with Claude in the page title. QuillBot and Grammarly both run free ones. None of them read Anthropic's watermark, because the reader for it has not been published, so the Claude branding tells you nothing about what is happening under it.
Does Turnitin detect Claude?
Turnitin scores prose for AI writing without naming any model, so it neither detects nor fails to detect Claude as such. It only scores qualifying text, meaning prose sentences in paragraphs, needs at least 300 words, and shows the indicator to instructors and administrators rather than to the student.
Can Claude bypass AI detectors?
No model reliably bypasses detection, and choosing a drafting model is a weak lever compared with editing. As of August 2026 the argument runs the other way too: Claude text from new models carries a watermark that other vendors do not currently apply to text, so model choice now cuts in more than one direction.
Is Claude AI detectable?
In two different senses, with two different answers. As AI-like prose, yes, at roughly the accuracy any classifier manages on any model, which independent testing puts well short of the marketing figures. As Claude in particular, only through Anthropic's own watermark, which no public tool can currently read.
Does Claude have an AI detector?
Anthropic does not sell a detector. It has committed to helping users and third parties detect its own marks under the EU AI Act transparency code, which is a provenance check on Claude output rather than a general AI detector, and the technical documentation for it has not been published yet.
Keep reading
The questions behind this one
How do AI detectors work?
The two signals every classifier reads, and what a percentage score actually means.
Can AI detectors detect Claude?
What testing has actually shown on Claude output, and where the confident numbers come from.
Is Claude less detectable than ChatGPT?
Probably a little, and nobody has published a test you can check for yourself.
Does ChatGPT watermark its text?
What OpenAI built, what it shelved, and how that now differs from Anthropic.
Which AI is least detectable?
How Claude, ChatGPT and Gemini differ on the properties detectors measure.
ChatGPT humanizer
The same operation on ChatGPT output, which leaves a different set of fingerprints.
Publish Claude drafts that read like you wrote them
Paste in a paragraph and watch which parts change. On Claude output, the hedges usually go first.
Plans planned from $9/month billed yearly · Your runs saved to your history, delete any time