Why we don't remove AI watermarks
Remuxe drops the AI metadata tags in a file but leaves invisible AI watermarks alone. The legal picture in the EU and the US, why watermarks are hard to remove, and why it would take data-center GPUs.
Every copy Remuxe exports is a brand-new file, so the AI metadata tags that editing apps leave in your source don't come along. Why Remuxe removes AI metadata tags Invisible AI watermarks are a different thing: they live inside the picture and the sound. We don't remove them, and we won't build a tool that does.
There are three reasons: the law, the technology and the hardware.
Tags and watermarks are not the same
A metadata tag is a note in the file's container: “made with this app,” “created by an AI model.” Write a new file without it and it's gone.
An invisible watermark is a pattern woven into the pixels or the sound samples themselves: spread across the whole frame, repeated frame after frame, and kept below what eyes and ears notice. Google's SynthID, which marks content from Google's AI models, is designed to stay detectable after cropping, filters, color changes, frame-rate changes and lossy compression — exactly the kinds of edits repurposing makes.
So the two behave differently. Tags disappear when a file is rebuilt. Watermarks are built to survive.
The law in the EU
Remuxe operates in the EU, under the AI Act. Its transparency rules — Article 50 — have applied since August 2, 2026:
- providers of AI systems that generate images, video, audio or text must mark what they generate in a machine-readable way, so it can be detected as AI-generated;
- businesses and professionals who use AI to make deepfakes must disclose that the content was generated or manipulated.
In June 2026, the European Commission published a Code of Practice on marking and labelling AI-generated content. It's voluntary, and it shows how providers can comply: watermarks, metadata and visible labels working together. Breaking the transparency rules can cost up to €15 million or 3% of worldwide annual turnover.
The whole system only works if the marks stay on. A tool built to strip them would exist to undo what the law asks for, and we won't make one.
The law in the US
The US has no federal law yet that requires AI watermarks or forbids removing them as such. Congress has debated provenance and digital-replica bills, such as the COPIED Act and the NO FAKES Act, but neither had become law at the time of writing.
The rules come from elsewhere:
- States. California's AI Transparency Act (SB 942, amended by AB 853) became operative on August 2, 2026. Large generative-AI providers must embed hidden disclosures in the images, video and audio they generate and offer a free tool to detect them. Later phases extend duties to large online platforms and to makers of cameras and phones.
- Copyright law. The DMCA forbids removing or altering “copyright management information” — such as names and identifying details attached to a work — when you know, or have reason to know, that it will help hide an infringement. Provenance data can carry exactly that kind of information.
- Platform rules. TikTok, YouTube and Instagram ask creators to label realistic AI-generated content, and they use watermarks and metadata to apply labels themselves.
The technical problem
You can't delete what you can't see. A watermark isn't a field in the file; it's spread across millions of pixel values, and only the provider's detector — with its secret key or model — can read it.
What has worked in research isn't deletion but redrawing. Researchers add noise to an image and let a diffusion model paint it again; studies have shown this “regeneration” can break several watermarking schemes. But it also redraws the picture: faces, text and fine detail come back different.
Video makes it harder still. Every frame has to be redrawn consistently with its neighbors, or the result flickers, and the sound needs the same treatment. And it's an arms race: watermark makers keep improving their schemes, so whatever works today may not work tomorrow.
The hardware problem
Remuxe processes your video on your own computer, in a browser tab. That's why your videos never leave it.
Redrawing a video with a diffusion model needs data-center GPUs. A 60-second clip at 30 frames per second is 1,800 frames, and each one would pass through a model with billions of parameters, plus the extra work of keeping frames consistent. By our estimate, cleaning one video that way costs roughly ten times more than generating it with an AI model in the first place.
A browser tab on a laptop can't do that. Doing it on our servers would mean uploading your videos — the opposite of how Remuxe works.
What Remuxe does instead
- Writes every copy as a new file, so tags from editing apps don't follow your real footage and get it mislabeled.
- Leaves watermarks inside the picture and the sound alone.
- Asks you to label AI content. If your video was made with AI, turn on the platform's AI label when you post.
Questions
Can Remuxe's changes remove a watermark by accident?
Remuxe's changes are tiny by design — 0.5–1% per frame — and robust watermarks are built to survive much stronger edits, like compression, cropping and filters. We don't test or tune Remuxe against watermark detectors.
Will platforms still label my AI video?
They can: from watermarks, from their own detection and from the label you choose when you post. Choose it yourself.
So why remove metadata tags at all?
Remuxe doesn't single them out: it writes a new file and copies nothing from the old one. That fixes real footage wrongly flagged as AI. If your video was made with AI, label it — and in places where AI labels are required, like China, keep them.