AI detector

Free AI text detector with watermark verification

Most detectors give you one number from one model. This one runs a watermark check, a metadata check and a classifier over the same text, and shows you which of them actually answered.

Updated 2026-09-17

Checks output from

ChatGPT
Claude
Google Gemini
Meta Llama
Mistral AI
DeepSeek
xAI Grok
Microsoft Copilot
Perplexity
Cohere
Hugging Face
Ollama

Where a provider publishes a watermark detector, we call it and report exactly what it answers. Where none exists — which today is nearly everywhere on this list — the classifier layer works on the text itself, using models trained on large-language-model output. Every report names which of the two answered, and which layers could not run at all. Current status per generator is on the accuracy page.

Example report108 words

Strong AI signalsConfidence: MediumAI signal 0.87

Strong AI signals across most sentences.

Layers

CheckedAnthropic watermark
No Claude watermark found in this text
No detectorSynthID-Text (Gemini)
Gemini text is watermarked, but Google offers no third-party detector
UnavailableMetadata
Pasted text carries no manifest. Upload the original file to check one.
FoundBase classifierAI signal 0.87
Base classifier, document and sentence scores.

Sentence view

Shading follows the per-sentence score. Pale means no signal.

Our new platform represents a significant leap forward in the way distributed teams collaborate. By leveraging cutting-edge technology, we empower organisations to unlock unprecedented levels of productivity. The intuitive interface ensures that users of every skill level can get started in minutes. Moreover, our robust analytics suite surfaces actionable insights that drive measurable results. We remain committed to delivering an experience that is both seamless and secure. In the months ahead we will continue to innovate alongside our customers. The three of us argued about the onboarding flow for a week in a Lyon flat with no heating, and the version that shipped is the one nobody liked.

Read this alongside the report

Metadata is absent for pasted text. Upload the original file if you need a C2PA or XMP check.
Detectors flag writing by non-native English speakers more often than writing by native speakers.
Edited or partly AI-assisted text is the hardest case for any classifier. Read the sentence view, not only the headline.

Start here

What is an AI detector, and should you use one?

Three things worth knowing before you trust any number, ours included.

AI writing has tells

Language models leave regularities behind. Sentence lengths cluster more tightly than a person's do, the same hedges recur, and word choice sits closer to the statistical middle than human prose usually manages — the property researchers call low burstiness. A classifier trained on enough examples learns those patterns.
That is all detection is: statistics over style. It is not a machine reading intent, and it is why detection works reasonably well on a long, unedited article and badly on a short paragraph or something a person has rewritten.
Assay Layer looks for them, and for more

We do not train a classifier of our own. We buy commercial ones, trained on large paired corpora of human and machine writing drawn from models including ChatGPT, Gemini, Claude, Llama, Grok and Mistral; the strongest of them have been put through independent evaluation by the University of Chicago economists behind NBER working paper 34223, August 2025. On a deep assay we run two and report where they disagree.
Then we do the part a single-model detector never does: query first-party watermark detectors and read C2PA metadata, so that when origin can be established rather than inferred, it is. Where every threshold sits, and what the independent literature says about how far any of this can be trusted, is set out on the accuracy page.
Content transparency is now a rule

Article 50 of the EU AI Act has required providers to mark synthetic output, and certain deployers to label AI-generated text, since 2 August 2026. No certification scheme for detection tooling exists yet, so nobody can be certified against it.
What a publisher or an employer can do is show their working: what was checked, when, and what came back. That is what an export carrying the input hash, the pipeline version and a UTC timestamp is for.

Try it free

Ten checks a day without an account, thirty signed in. No card, no trial clock.

Reading the output

How to read the report

The first line of every report is a signal, not a verdict. It answers the question “what kind of evidence do we have here”, and there are five possible answers.

Signal	What it means
Provenance established	A first-party watermark or a valid C2PA manifest was found. The origin is established rather than inferred, and this outranks anything the classifier says.
No AI signals	Every layer that could run, ran, and none of them found anything.
Mixed signals	Signals are weak, only part of the text is flagged, or two classifiers disagree. Most edited text lands here.
Strong AI signals	The classifier layer scored the text high across most of its sentences.
Not enough text	Under fifty words, or unreadable. No number is reported.

Next to the signal sits a confidence — low, medium or high — and, when a classifier ran, a labelled number such as “AI signal 0.87”. You will never see a bare percentage on its own, because a percentage with no label invites people to read it as a probability of guilt. The exact thresholds behind each signal and each confidence level are written out on the accuracy page.

Limits

What a score can and cannot tell you

Short text is close to noise

Below fifty words we return no number at all. Between fifty and roughly a hundred and fifty words a score exists but deserves little weight, and the report says so in its caveats. Every public evaluation we have read shows the same cliff at short lengths.

Non-native English is flagged more often

This is the best-documented bias in the field and it is not specific to any one vendor. If your workflow touches writing by people who learned English later in life, a detector score is not a safe input to a decision about them. We keep this caveat on English-language reports.

Edited and hybrid text is the hard case

A draft generated by a model and then rewritten by a person, or a human draft tidied by a model, is exactly what classifiers are worst at. These land on “mixed”, which is the correct answer, and the sentence view is more informative than the headline.

A watermark is the only strong evidence

Everything else is inference from style. A first-party watermark, or a signed C2PA manifest, is a statement by the system that produced the content. When one turns up, the report leads with it and the classifier becomes a footnote.

None of this is a reason to give up on checking. It is a reason to check with more than one method and to keep the uncertainty visible, which is the whole design of this product.

Design

Why one model is not enough

A single classifier gives you a single failure mode. When it is wrong, nothing in the output tells you that it is wrong — the number looks exactly the same as when it is right.

Running two independent classifiers changes that. Where they agree, the score is worth more than either alone. Where they disagree, the disagreement is the finding, and we surface it as an agreement score on deep assays rather than averaging it away. That is one of the reasons a deep assay costs ten credits per thousand words instead of one: it buys a second opinion from a different vendor with a different training set.

Adding non-statistical layers changes it again. A watermark check and a metadata check do not look at style at all, so they fail differently from a classifier — and when they succeed, they are far more conclusive than any stylistic score can be. A report that combines them tells you which kind of evidence you are holding, which is the thing an editor or a compliance officer actually needs to know.

How we measure accuracy Run this from your code

Comparison

Assay Layer next to a typical single-model detector

We are not naming competitors, because the point is the category, not a scoreboard. Most tools sold as AI detectors are one classifier behind a form.

Capability	Typical single-model detector	Assay Layer
First-party watermark check	No	Anthropic text watermark; SynthID-Text adapter ready
File metadata (C2PA, XMP)	No	On file uploads, in the app
Second, independent classifier	No	On deep assays, with an agreement score
Sentence-level highlights	Sometimes	On every classifier result that returns spans
Refuses to score short text	Rarely	Under fifty words, no number is returned
EU hosting	Usually hosted in the US	Frankfurt, Germany
Input retention	Often stored by default	Hash and report only, unless a workspace opts in
Auditable export	Screenshot only	JSON and PDF with input hash, pipeline version, timestamp
API and MCP access	API on higher tiers	Both, from the first paid tier

Questions

AI detection, answered plainly

Is this AI detector really free?

Yes. Ten standard assays a day without an account, thirty a day when you are signed in, with no card and no trial clock. The limits exist because each assay costs us money at the classifier providers, not as a sales tactic.

Which AI models can you detect?

The classifier layer is model-agnostic: it looks at the text, not at a list of models, so it responds to output from any recent large language model. The watermark layer is different — it can only confirm output from systems that watermark, and only where the vendor exposes a detector. Today that means Anthropic's text watermark in private preview, with a SynthID-Text adapter ready for the day a public detector exists.

Can I check a PDF or a Word file?

File uploads run in the signed-in app rather than on this page. A file gets the same layers plus the metadata layer, which reads the C2PA manifest and XMP fields that pasted text simply does not have.

Why does my report say too short?

Under fifty words the classifier layer returns too short and no score. That is deliberate. A number computed from two sentences looks authoritative and is close to noise, so we would rather show you nothing than something misleading.

Will this flag my writing if English is not my first language?

It may, and so will every other detector. A 2023 Stanford study found seven detectors misclassified 61 per cent of TOEFL essays by non-native writers as AI-generated, against under 5 per cent for essays by native writers. We keep that caveat attached to reports and we do not recommend using any detector as evidence against a person.

Can I use this to accuse a student or a freelancer?

Please do not. A report gives you grounds to ask a question, compare against earlier work, or request a draft history. It is not proof of anything about a person, and we deliberately do not print verdict language that would let you pretend otherwise.