Claude Watermark Remover & AI Provenance Cleaner

What "VERIFIED CLEAN" means

MarkMyAss shows VERIFIED CLEAN only when a supported signal was actually found before cleaning, its own re-inspection confirms it's gone, and every independent verifier that could run (ExifTool, and c2patool where applicable) agrees too. It is not a claim that the file contains no possible identifying signal whatsoever — only that the specific, supported metadata categories MarkMyAss targets are independently confirmed gone. See the AI Watermark Lab for the full, honestly-scored capability matrix.

Is MarkMyAss only for Claude?

No. MarkMyAss is not limited to Claude. It works across supported watermark and provenance signal types, including hidden Unicode and common file metadata. Its support depends on the signal type, not just the AI provider.

A real check we ran

We generated an image with Google's Gemini image model and ran it through MarkMyAss. It found Google's file-level provenance — XMP, PNG text and a C2PA manifest — and Clean removed those supported signals, so a second inspection showed 0 supported file-level signals. We then uploaded that same cleaned image to Gemini, which replied: “Based on the digital watermarking analysis, all or part of this image was created or edited using Google AI.”

After MarkMyAss removed the supported file-level provenance and a second inspection showed 0 supported file-level signals, Gemini still identified the image as created or edited using Google AI based on digital watermarking analysis. This is consistent with a remaining content-level watermark such as SynthID, which is separate from metadata — though it does not prove SynthID alone was responsible. Cleaning metadata is not the same as removing SynthID — so we don't pretend it is.

Why this exists — the EU AI Act (plain English)

From 2 August 2026, EU AI Act Article 50 asks providers of generative AI to ensure synthetic audio, image, video or text outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated.” It targets the outcome (detectability) and does not mandate SynthID — SynthID is Google's own technical approach, not an EU standard. Article 50(2) also carves out an exception where the AI performs “an assistive function for standard editing” or does not substantially alter the content or its meaning. This is background, not legal advice.

What MarkMyAss helps writers do

Compare your original draft with the AI-edited version and see which wording stayed yours, which were mechanical corrections, and which were substantial AI rewrites — so you can preserve or restore your own voice. MarkMyAss does not remove Google's SynthID-Text; it helps you distinguish human-authored wording from substantial AI rewriting. Built for authors, journalists, students, editors, copywriters and researchers.

These figures are an example layout only — never shown as real unless computed from your actual document.

Our local research: how much AI-generated text does the watermark need?

These are MarkMyAss local research results using our own SynthID keys and reference setup. They are NOT measurements of Google's private production detector, and do not apply directly to Gemini.

Replication across 5 independent SynthID keys × 5 generation seeds × ~250 / 500 / 900-token documents, using a local reference implementation (distilgpt2 as the generation vehicle) and the tuning-free Weighted-Mean detector. Contiguous AI-generated fraction → local detection rate:

• 5% AI1%
• 10%9%
• 15%19%
• 20%43%
• 25%67%
• 30%83%
• 40%93%
• 50%95%

• ≤10% AI-generated content: mostly undetected in our local setup
• ~15–30%: transition zone
• ≥40–50%: commonly / strongly detected

Our measured 50% detection crossing was about 21% AI for one contiguous span and 27% for scattered blocks. Full method & data: study 1, inverse study, replication.