The evidence report for one document

Upload a thesis, article or blog post. See which passages read as machine-generated, why they were flagged, and get a signed report anyone can verify. No subscription.

Opening soon

What a report looks like

Illustration, not a real customer document.

Part of this document shows signs of AI generation

One passage of about 420 words scores in the range typical of machine-generated text. The rest shows no evidence of it.

strong signalsignalweak signalwavy = the model's most predictable wording

We started the survey in March, mostly because the funding came through late and half the volunteers had already left for the summer. In today's rapidly evolving landscape, it is important to note that community engagement plays a pivotal role in shaping sustainable outcomes. Furthermore, by leveraging a multifaceted approach, stakeholders can foster meaningful collaboration and drive lasting change. Anyway, the numbers from the second village were a mess, and I'm still not sure what to make of them.

How it works

1. Free preview

Word count, language, price, and whether the text is long enough for a reliable result. Before you pay anything.

2. GPU check

Two 7-billion-parameter language models score every token (the Binoculars method). Scored on demand, deleted after 24 hours.

3. Passage-level report

A heatmap by sentence, the exact phrases that triggered it, and a writing coach that points at generic prose without rewriting it.

4. Signed and verifiable

Every report is signed (Ed25519) and tied to a hash of the text. Anyone can check that a report is genuine and matches the document.

No subscription. Pay for the documents you check.

Free

CHF 0

5 checks every month, up to 2,500 words each. No card needed.

Short

CHF 1.90

Up to 2,500 words: an essay or blog post.

Article or chapter

CHF 3.90

Up to 10,000 words.

Thesis

CHF 9.90

Up to 100,000 words: your whole thesis, once.

Top up a wallet with CHF 10, 20 (get 21) or 50 (get 55). Credits stay valid for 24 months, nothing renews. Previews and report verification are always free. Institutions get shared wallets at 20% off, with no per-seat fees.

What it can't do

A report is evidence, not proof. TextWitness never says a text is "human-written": the best result is "no evidence of AI generation found".

It says "inconclusive", and why, when a text is under 200 words, not in English, highly repetitive, or scores between the published thresholds. It doesn't guess which AI model wrote something, and it doesn't rewrite text to avoid detection.

English only for now. German, French and Italian are planned.