Text provenance
Is this text AI? Get an answer you can defend.
Three surfaces over one pipeline: a free checker on this page, an API and MCP server for your own systems, and a compliance log for EU AI Act Article 50.
Updated
- Anthropic watermark Checked
- SynthID-Text (Gemini) No detector
- Metadata Unavailable
- Base classifier Found
Strong AI signals across most sentences.
Pipeline 2026.09.1 · Input hash b5bef6159c21e2e0…
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.
-
14.8%
of English-language student essays now read as 80% or more AI-written
Turnitin, February 2026
-
60+
universities have switched AI detection off, citing false positives
Reported institutional policy changes, 2023 to 2026
-
2 Aug 2026
AI Act Article 50 transparency obligations have applied since this date
Regulation (EU) 2024/1689
-
< 50 words
the length below which every detector studied becomes unreliable
NBER working paper 34223, August 2025
Other people’s numbers, with the date and the source on each one. How we read them, and the thresholds they inform, are on the accuracy page.
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.
Ten checks a day without an account, thirty signed in. No card, no trial clock.
How it works
Four layers, one report
Each layer answers a different question, and the report shows all four answers — including the ones that could not be given.
- 01
Watermark
First-party detectors, where the provider exposes one.
- 02
Metadata
C2PA manifest and XMP fields on an uploaded file.
- 03
Classifier
The base classifier on every assay, the deep classifier as well on a deep one.
- 04
Report
One shape on screen, in the PDF, over the API and MCP.
We never print the sentence “this is AI” or “this is human”, because no detector on the market can support it. What a report gives you is a defensible reason to ask a question. If you want the detail, read how to read a report or the evidence behind our thresholds on accuracy.
Who uses it
Five jobs, one pipeline
One report, reached three ways: the checker on this page, the API, or an MCP tool inside an agent. Pick the row that looks like your week.
Publishers and editors
A freelancer files 1,400 words an hour before deadline and something in the third paragraph reads wrong. You need grounds to ask a question, not a verdict, and you need it before the piece goes out.
Surface Free checker, then history and PDF export once you sign in
See how it worksTwo credits. A reason to ask, or a reason to stop asking.
Platforms and marketplaces
Listings, reviews and seller applications arrive faster than anyone can read them. You want a provenance signal attached to each one at ingest, with your own reference on it so you can find it again.
Surface API, one POST per item, with your reference echoed back
See how it worksHiring and HR
Cover letters and written assessments. This is the highest-stakes use on this page and the one where a false positive does the most damage — which is why the report refuses to score short answers and keeps the non-native-writing caveat attached.
Surface Checker or API, always with a human reading the result
See how it works- Under 50 words: no number is returned at all.
- Detectors flag non-native English writing more often.
- A score is not evidence about a person.
Compliance teams
Article 50 has applied since 2 August 2026. You need to show what you checked, when, and what came back — in a form that still means something at an audit next year.
Surface JSON and PDF exports carrying the input hash and a UTC timestamp
See how it works- input hash
- b5bef6159c21e2e0…
- pipeline
- 2026.09.1
- checked at
- 2026-09-01T09:12:45Z
Developers and agents
An agent receives a draft, or writes one, and needs to check it before passing it on. Same pipeline, same report, reachable as an MCP tool rather than an HTTP call you have to wire up.
Surface MCP server at mcp.assaylayer.com/mcp, same bearer key as the API
See how it worksassay_text({
text: draft,
kind: "deep",
reference: "post-8412"
})
→ signal: "mixed"
confidence: "medium"
caveats: [ … ]Where your text goes
Built for the EU, on purpose
Built for the EU
- EU AI Act Article 50 ready
- Hosted in Frankfurt, Germany
- GDPR: inputs not stored
No official EU certification scheme for Article 50 tooling exists yet, so nobody can be certified against it — us included. Ready here means the pipeline produces the evidence the obligation asks for: a layered check, and an export carrying the input hash, the pipeline version and a UTC timestamp. It is not an approval, an endorsement, or a certificate.
Hosted in Frankfurt
The application, the database and the queue run in Frankfurt, Germany. The classifier layers run on third-party services; what they receive, and for how long, is set out on the accuracy page and in our privacy notice.
Inputs are not stored
We normalise your text, hash it with SHA-256, run the layers, and keep the hash and the report. The text is not written to the database. A workspace can switch retention on if its own rules require it, and that switch is recorded in the audit log.
Exports you can file
Every report exports as JSON and as a one-page PDF carrying the input hash, the pipeline version and a UTC timestamp. That is what makes a check auditable six months later, when nobody remembers the text.
Article 50 ready
The obligations that started applying on 2 August 2026 are about marking, labelling and keeping records. A layered check plus a hashed, timestamped export is the shape of evidence those obligations ask for.
Pricing
Free to try, one unit to reason about
Everything is priced in credits. A standard assay costs one credit per started 1,000 words.
| Operation | Cost | Notes |
|---|---|---|
| Standard assay | 1 credit | Per started 1,000 words. Base classifier, watermark and metadata layers. |
| Deep assay | 10 credits | Per started 1,000 words. Adds the deep classifier and an agreement score. |
| File upload | +1 credit | Added once per file, for text extraction and the C2PA or XMP read. |
Ten free standard assays a day without an account, thirty signed in. Credit packs from 19 euro, API plans from 29 euro a month, and a compliance tier with retention controls and a signed data processing agreement, sold by conversation rather than by checkout.
Questions
Frequently asked
Does Assay Layer tell me whether a human wrote this?
No, and nothing on the market can. We report a signal — provenance, none, mixed, AI, or insufficient — with a confidence level and the per-layer evidence behind it. The editorial decision stays with you, which is also why every report carries its caveats.
Do you store the text I paste?
No. By default we keep the SHA-256 hash of the normalised text and the report itself, not the text. Retaining inputs is a workspace setting that is off unless someone deliberately turns it on.
How short is too short?
Under fifty words the classifier layer returns too short and no number at all. Short passages do not carry enough signal, and published evaluations consistently show detectors degrading sharply below that length.
What does a check cost?
The checker on this page is free: ten standard assays a day without an account, thirty when you are signed in. Beyond that, one credit covers each started 1,000 words of a standard assay, and credit packs start at 19 euro.