AI Text Watermark can now detect a watermark while text streams

What it is

The question everyone asks about a piece of text is whether a model wrote it. This library answers it for text you generated yourself: it nudges the model's word choices in a pattern derived from a secret key, then detects that pattern later with the same key.

No extra tokens are added and quality is not degraded, which is what separates this approach from appending a visible marker. It runs in Node and in the browser.

What changed

Commit dd95fe6 added inflight analysis: an online detector built on MidStream that scores a stream as it arrives rather than waiting for the completed text. Commit a1d43ca then put that detector into the live playground so you can watch it work.

The standalone npm package moved from 0.3.0 to 0.4.0 in the same series.

Get started

Node 18 or newer:

npm install ai-text-watermark

There is a Ruflo scoped build and a Rust crate for the same algorithm:

npm install @claude-flow/watermark
cargo add ruflo-watermark

Expected result: the package installs and exposes generation and detection against a key you supply. The hosted playground lets you try it with no install at all.

No documented MCP server. The supported routes are the npm packages, the Rust crate and the browser playground.

Use it today

Practical case: a publishing pipeline that needs to prove which drafts came from its own assistant. Input is the model output plus your key. Workflow is watermark at generation, store nothing extra, detect on any later copy. Output is a detection score.

Acceptance test: watermark a passage, paraphrase a quarter of it, then detect. The score should drop but stay above the unwatermarked baseline.

Push it further

Experimental commentary. Inflight detection makes a firewall plausible: score a stream as it arrives and cut it off mid sentence when provenance fails.

Limitation: watermarking only covers text you generate. It says nothing about text from a model you do not control. Falsifiable test: feed unwatermarked text through the online detector and confirm the false positive rate on your own corpus.