WriterMark is like cryptographically secure (and currently rather naive) Recaptcha for a writing session.
Try it online quickly.WriterMark is a protocol for assessing and certifying text as human-written by observing the process of writing. WriterMark analyzes keystroke timing, pauses, paste events, and more complex metrics to evaluate whether typing appears human, and — if it does — outputs a cryptographically signed certificate. The text is hashed on your local machine, and the hash and the certificate are both signed with a secure JWT, meaning provenance can not be faked, and the system can confirm any text/certificate pair at any later time. The content of your text is never shared with the server.
A lightweight client-side collector records telemetry as the writer types, capturing:
These events include typing speed, deletion, cursor jumps, selection, paste, and a large number of additional metrics that all get collected and sent to the WriterMark API.
WriterMark does not capture:
Every ~30 seconds, WriterMark passes the telemetric information up to the server, along with:
The server then verifies any previous checkpoint, merges the new events into the running aggregates, scores the aggregates, signs an updated JWT, and returns the checkpoint, the pass/fail status, and a score (if passing).
Additionally, further logic maintains score across documents within the WriterMark ecosystem (see "Connecting to WriterMark" below), so if you copy and paste from one document to another, that text is properly validated against its original provenance and scored appropriately.
The current scoring algorithm uses a weighted average of more than a twenty independent signals, each producing a 0-1 confidence weight. One example of a basic signal is pause distribution. Repeated pause lengths are a simple indicator of non-human text editor interaction. More advanced behavioral signals provide more robust insights.
The certification system makes additional checks on top of the score, like the coverage of the telemetric data, looking for any injected text or unexplained changes in body length. These additional checks further modify the final pass/fail result. The algorithm will continue to evolve over time as more individuals use the system.
One thing I'd like to note about the scoring system and the architecture in general — this verification system directly addresses the current most frequent form of AI writing falsification: simple copy and paste of AI text. Copying an entire document of text into a WriterMark-integrated system will result in an instant failure. I believe this has immediate implications for assisting in protecting against undisclosed AI content.
The actual WriterMark server itself is stateless. All running state is embedded into each signed checkpoint JWT, which the client stores locally. If a writer returns to a document days later, the client sends the last stored checkpoint (see section "Connecting to WriterMark" below), and the server picks up exactly where it left off. This means there's no session table, no user tracking, no server-side database. The server is purely a verification and certification system used to ingest a previous checkpoint in and push a new checkpoint out. The checkpoints are chained by hash, resulting in an append-only system that disallows injecting fake checkpoints.
WriterMark provides an SDK (currently private) with a built-in integration for TipTap, an open-source text editor for the web. Using this integration, we've hooked the WriterMark system up to a lightweight desktop app called Wintertext (www.wintertext.com) for writing verified text documents. The app is completely free and you can use it to test the system if you are interested. The app uses a dedicated file format .wtxt — a simple JSON wrapper that stores document content, styling, metadata, and the certificate together. Export to Markdown, JSON, HTML, and Word (.docx) is also supported, with the certificate included in each format.
Its very likely that I will (1) publish the SDK publicly in the future so anyone can integrate it, (2) release a more robust extension for TipTap that other developers could easily integrate into their own projects, and (3) work on further integrations for more text editing systems (I'm looking at you, Google Docs, sorta, not really).
writermark.org/challenge page will invite individuals to try their hand at beating the system in an environment where adversarial data can be properly collected.I touch on this more in my closing statement, but in a brief word: its necessary. There is no other possible solution barring collecting genuine biometric data (which I think may also become necessary and common in the future, e.g. a keyboard that confirms human fingers are touching it). Without a system to verify text, we are right now in the middle stages of completely losing the ability to validate whether a human or a machine wrote something. I think this has extremely serious consequences, especially for the landscape of creative writing.
If you'd like to skip this overview, and just see the system in action, you can check www.writermark.org for a simple example, or download the fully functional text editing app Wintertext (www.wintertext.com). WriterMark and Wintertext are both currently completely free to use, collect no personal info like email, and require no special verification. On a personal level (for individual writers) my hope and intention is for the system to remain free indefinitely. I may explore enterprise partnerships as a way of funding future development.
Short answer: analyzing the text content doesn't work. AI are already too smart, and can vary their writing style to beat text content tests.
I'd like to speak briefly to people's ethical and pragmatical compasses. Without a system like this, humans will lose the ability to validate whether text was human generated. I really think many people do not understand the magnitude of what that could mean. On the fiction writing side, it means essentially the death of the human author. In a world where AI and human can both publish novels, especially with the continuous improvement of intelligence models, AI speed will overrun and flood the human market, and publishers, editors, and literary agents will be endlessly overrun with AI written content submissions. Publishers will have no way to identify whether something is human written or AI written, and neither will readers.
Already, online fiction platforms and literary magazines, and even novel publishers are seeing the effects of this issue in real time. As an individual who personally enjoys creative writing as a hobby, I am very alarmed by the progression of AI writing capabilities and the current lack of ability to evaluate the veracity of a body of text.
Thank you for your time,
JHYolm