ai-watermark-guard

What it does

AI-written text arrives with punctuation people rarely type: em dashes, curly quotes, an ellipsis character. Pasted text brings worse - zero-width spaces, a stray byte order mark, mojibake left by something that read UTF-8 as Latin-1. None of it is visible in a diff, and all of it outlives the edit that introduced it.

ai-watermark-guard finds those characters in every tracked file, and in commit messages, and either reports them or writes the plain equivalent. It is written in Rust and ships as a single prebuilt binary: no toolchain, no compile step, no install script.

What this is not

There is no hidden watermark in machine-written text, and nothing here defeats a detector. What this finds are ordinary characters that a model reaches for and a person usually does not, alongside invisible characters and mojibake that nobody meant to commit. It is a text-hygiene guard. Judge it as one.

Install

npx ai-watermark-guard --all          # one-off
npm i -D ai-watermark-guard           # pinned, for hooks and CI
cargo install ai-watermark-guard      # if you would rather build it

Prebuilt binaries ship as optional dependencies, one per platform, selected by os, cpu and libc. Your machine fetches one package of a couple of megabytes and runs it - nothing compiles, and no install script runs, so it survives an install with --ignore-scripts.

Fast enough to run on every commit

A guard that runs on every commit has to be quicker than the pause it adds. Over this soroush.tech/core site's own repository - 1,605 tracked text files, a full sweep takes about 100ms, where the Vanilla Node.js script it replaced took 230ms.

Most of what remains is git and file I/O rather than scanning, which is the honest reason to write it in Rust: not that the character matching is hard, but that a binary starts instantly. A hook that costs a tenth of a second is one nobody turns off, and --staged looks only at the files in the commit, so the usual run is quicker still.

What it looks for

Three tiers, all on by default. --rules narrows them.

invisible and mojibake are wrong anywhere. punctuation is a house rule, so turn it off if your prose wants typographic quotes:

aiwg --all --rules invisible,mojibake

What it never touches

U+200C, U+200D, U+200E and U+200F are invisible too, and this tool leaves them alone whatever the rules say. Persian needs the first between the parts of a word, an emoji sequence needs the second to hold itself together, and the last two set the direction of mixed right-to-left text. Removing them corrupts text rather than cleaning it.

Letters are never touched, in any script. This is not an ASCII-only rule.

Use it in your hooks

# .husky/pre-commit
npx aiwg --staged --fix

# .husky/commit-msg
npx aiwg --message "$1" --fix

Prose is fixed and re-staged. Code is reported and the commit stops - a curly apostrophe inside a single-quoted string becomes a straight quote that ends the string, and the file no longer parses, so fixing code needs a parser rather than a replacement.

A commit message is fixed in place before the commit object exists. That is the only moment it costs nothing: amending afterwards changes every sha from there on and voids the signatures over them.

Choosing what to scan

Binary files are skipped by the same NUL-byte test git uses. Files that are not valid UTF-8 are reported and skipped rather than scanned through a lossy decode, which would blame the decoder instead of the file.

Exit codes: 0 clean, 1 findings, 2 the run itself failed.