A tool for protecting human writing
Writermark is a free, open, nonprofit protocol that helps to prove that text was written by a human. It provides cryptographically secure human verification for a writing session.
AI detectors analyze finished text and guess. Writermark records the writing process and signs the result.
Wintertext is the desktop writing app that integrates Writermark — the way to actually use it. Learn more →
Try it
Type something below. Certification runs every 30 seconds as you write — no button to press.
Note: Scores generally start low and increase the more you type. Stream of consciousness and nonsensical random typing have the highest chances of being flagged as non-human. The system works best as you craft a longer text entry.
Writermark’s Nonprofit Mission
We are currently in the process of establishing this project as a nonprofit initiative. Our objective at Writermark is to do the following two things:
- Assist in protecting human writing by providing writers the tools to help prove that their work is human-authored.
- Making sure that the tools we provide are free for any individual writer to use, forever.
How it works
- Write. A small client SDK runs alongside your editor. It records keystroke timing, pauses, revisions, cursor movement, and paste events. No text content is ever stored on the server.
- Certify continuously. Every 30 seconds, the SDK sends behavioral telemetry to the server. The server scores it, updates running aggregates, and returns a signed checkpoint.
- Chain. Each checkpoint references the previous one, forming a tamper-evident chain. No server-side storage — the client holds the chain.
- Verify. The latest checkpoint is a portable JWT. Anyone with the public key can verify it. No account, no API call, no database lookup.
≡≡≡ WRITERMARK CERTIFICATE ≡≡≡ Status: VERIFIED HUMAN Confidence: 74% Date: 2026-02-22 Text hash: 3a7f2b91c0d4e8f1... ——— Token: eyJhbGciOiJFZERTQSIs... ——— Verify at: writermark.org/verify ≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡
Ongoing work
ML-trained scoring
Replacing the heuristic model with a machine-learned classifier trained on real writing sessions. Better accuracy, harder to fool.
Hardening
Cryptographic content binding, external timestamp anchoring, revision topology analysis, and deeper behavioral signals to raise the cost of faking.
Wider adoption
More editor integrations, platform partnerships, and working toward formal standards for process-based authorship verification.
Offline
Local signing via hardware security (TPM / Secure Enclave) so certification can continue when the writer is offline.
What is collected
Recorded
Keystroke timing · Pause patterns · Revision behavior · Cursor movement · Paste events · Scroll · Focus/blur · Session duration
Never recorded
Your text · Key values · Screenshots · Personal information · IP addresses · Browser fingerprints
Properties
Portable
A certificate is a ~400-character string. Embed it in HTML, PDF metadata, markdown, or paste it in a tweet.
Self-verifying
Verification needs only the text, the token, and the public key. No network request, no account, no database.
Tamper-evident
The certificate contains a SHA-256 hash of the exact text. Change one character and verification fails.
Free and open
The protocol is free. Verification is free. Anyone can integrate into their own apps to better assist in identifying genuine human writing.
What a certificate means
What it provides
Evidence that text was composed through a realistic human writing process — with natural pauses, revisions, corrections, and non-linear editing. Writermark helps prevent AI dishonesty, makes faking human-generated content meaningfully more difficult, and gives writers a way to back up their claim of authorship.
What it does not prove
A certificate is not absolute proof. A sufficiently sophisticated attacker could potentially beat the system. Writermark does not prove that the writer originated the ideas, or that no one else contributed. Process verification is strong evidence of how text was produced — not a guarantee.
Built into
For developers: npm install @writermark/sdk — documentation
Live stats
—Documents
—Certifications
—Events processed
- How is this different from AI detection?
- AI detectors analyze finished text and guess a probability. Writermark records the physical process of writing and produces a cryptographic certificate. It doesn't detect AI — it certifies humans.
- Can the telemetry be faked?
- Telemetry is streamed in real time with server-side timestamps. To fake a 30-minute writing session, you'd need to actually spend 30 minutes sending plausible keystroke patterns. This is possible in one of two primary ways: (1) An advanced text entry bot designed to beat all the behavioral signals, or (2) A human transcribing AI generated text in real time from an AI generator. As our algorithm becomes more advanced, these attack vectors will become more difficult.
- What about copy and paste?
- Writermark tracks an authorship map — every character is tagged as typed, pasted from a certified source, or pasted from outside. Paste certified text from another Writermark document and it keeps its provenance. Paste from outside and it gets flagged. Delete pasted text and the score recovers immediately.
- What about multiple writing sessions?
- Each checkpoint contains the full running state. Write 400 words on Monday, close the tab, come back Friday, resume from the last checkpoint. No server-side session storage needed.
- Does verification cost anything?
- No. Verification is free and will stay free. A certificate that can't be freely checked isn't useful.
- Can I integrate this into my own editor?
- Yes. The SDK works with any text editor — TipTap, ProseMirror, plain contenteditable, textarea. Grab the SDK or reach out.
- Who maintains this?
- Writermark is maintained by the team behind Wintertext.
Read more: Working To Verify and Protect Human Writing — a technical overview of the protocol, scoring, and architecture.