I Built an Open-Source Toolkit for AI Watermark & Provenance Forensics

AI-generated content is becoming increasingly common.

But when you receive an AI-generated image, document, or other digital artifact, an interesting question remains:

What can we actually learn from the file itself?

There may be metadata, provenance information, invisible Unicode characters, statistical signals, or other artifacts that aren't visible from the surface.

I wanted a developer-friendly way to investigate those layers locally.

So I built ClaudeMark.

🔗 GitHub: https://github.com/karthikrshet/ClaudeMark

What is ClaudeMark?

ClaudeMark is an open-source, local-first toolkit for AI watermark and provenance forensics.

The goal isn't to make an overconfident binary claim like:

"This file was definitely generated by AI."

Instead, ClaudeMark analyzes multiple types of signals and presents them as forensic evidence.

That distinction matters.

A statistical signal isn't automatically proof of authorship, and the absence of a signal doesn't necessarily prove that AI wasn't involved.

What can it analyze?

ClaudeMark currently brings several different analysis layers together.

🔎 Statistical AI signals

ClaudeMark can analyze supported content for statistical patterns that may be associated with AI-generated material.

These results are treated as signals rather than definitive proof.

🧬 Invisible Unicode analysis

Text can contain characters that aren't visually obvious.

ClaudeMark can inspect Unicode characteristics and help visualize suspicious or invisible characters that may otherwise be overlooked.

This is particularly useful when investigating text that appears completely normal to a human reader.

📋 Metadata & provenance

Digital files can contain metadata and provenance information such as:

• EXIF
• XMP
• C2PA
• Embedded metadata
• Other file-level artifacts

ClaudeMark provides tooling to inspect these layers and, where supported, sanitize them.

🛡️ Security auditing

Forensics isn't only about provenance.

Files can also contain potentially dangerous or unexpected structures.

ClaudeMark includes defensive security-oriented checks so that analysis isn't limited to simply asking whether content is AI-generated.

🧹 Sanitization

For supported formats, ClaudeMark can remove selected metadata and provenance artifacts.

The important part is that the workflow is designed around inspection first.

You can understand what is present before deciding what to clean.

Why local-first?

This was one of the most important design decisions.

A forensic tool may be used with sensitive:

• Research data
• Private documents
• Proprietary images
• Internal company files
• Unreleased content

Sending those files to an external service isn't always desirable.

ClaudeMark therefore follows a zero-egress/local-first approach.

The goal is simple:

Your files stay on your machine.

Architecture

The project is designed as a collection of focused analysis and processing components rather than one giant detector.

At a high level:

                 ┌─────────────────┐
                 │     Input       │
                 │ Files / Content │
                 └────────┬────────┘
                          │
                          ▼
                ┌───────────────────┐
                │ Detection Pipeline│
                └─────────┬─────────┘
                          │
        ┌─────────────────┼─────────────────┐
        ▼                 ▼                 ▼
   AI Signals       Unicode Analysis   Metadata /
                                      Provenance
        │                 │                 │
        └─────────────────┼─────────────────┘
                          ▼
                ┌───────────────────┐
                │ Forensic Results   │
                └─────────┬─────────┘
                          │
                ┌─────────┴─────────┐
                ▼                   ▼
             Audit              Sanitize

The project also provides developer-facing CLI/API capabilities and is designed to be extended with additional detectors and integrations.

Why open source?

I wanted developers to be able to inspect the implementation instead of treating AI-content analysis as a black box.

Open source also makes it possible to:

• Audit detection logic
• Reproduce results
• Add new detectors
• Report edge cases
• Contribute format support
• Build integrations
• Experiment with new provenance signals

What ClaudeMark is NOT

This is an important distinction.

ClaudeMark does not claim that every AI-generated file can be reliably identified.

AI detection is inherently difficult, and different signals have different levels of reliability.

ClaudeMark is intended as a forensics and analysis toolkit, not an infallible AI-authorship oracle.

What's next?

I'm interested in expanding ClaudeMark around:

• More media formats
• Additional provenance standards
• More watermark research adapters
• Stronger forensic visualization
• Additional security checks
• Developer integrations
• AI-agent workflows
• More reproducible benchmarks

The project is open source, so contributions and feedback are welcome.

Try ClaudeMark

GitHub:

[https://github.com/karthikrshet/ClaudeMark]

``
(https://github.com/karthikrshet/ClaudeMark)

If you're working on AI-generated content, provenance, digital forensics, privacy, or developer security tooling, I'd love to hear what problems you're running into.

What should ClaudeMark investigate next?