Why Remuxe removes AI metadata tags (and what it leaves alone)

You filmed it yourself. You trimmed it, fixed the color, maybe let an AI tool remove a microphone cable from one shot. You post it, and the platform stamps it “AI-generated.”

The label didn't come from your pictures. It came from a few lines of metadata riding along inside the file. This article explains where those tags come from, why they so often misfire on real footage, and what Remuxe does about it — and what it deliberately doesn't do.

Where AI labels live inside a video file

A video file is a container: a set of nested boxes. One box says what kind of file it is (ftyp). One holds the index — tracks, timing, codec settings (moov). One holds the actual frames and sound (mdat). And a few boxes can hold almost anything: udta, meta, uuid. That's where AI labels live.

There are three kinds you'll meet:

1. C2PA Content Credentials. A signed manifest — a list of the tools and actions that produced the file, such as “created with an AI model” or “edited” — stored inside the file and cryptographically tied to its exact bytes. Some cameras, phones, editing apps and AI generators add them.
2. XMP / IPTC metadata. A text packet with descriptive fields. The one platforms care about is the IPTC digital source type: trainedAlgorithmicMedia means “made by an AI model,” and compositeWithTrainedAlgorithmicMedia means “a real image with AI-generated parts” — the value tools write after inpainting or generative fill.
3. AIGC labels. Since September 1, 2025, China's labeling rules require AI services to write “implicit labels” into the metadata of what they generate: what the content is, who made it and a content ID. Apps built for that market write them, and the files travel.

Platforms read these tags. In May 2024, TikTok became the first video platform to read Content Credentials, so content made with AI elsewhere is labeled automatically when it's uploaded. Meta reads C2PA and IPTC data for its “AI info” label. More platforms are adopting the same standards.

Why real videos get flagged

Metadata describes the tools in the chain — not how much of the video is synthetic. One AI-assisted step can mark the whole file:

• removing an object with generative fill;
• AI noise reduction, upscaling or frame interpolation;
• a phone's AI eraser or “enhance” button;
• an editor's AI feature used on one clip, or a template or stock asset that carried a tag into your timeline;
• exporting through an app that tags everything it outputs.

The result is a label that's technically traceable and practically wrong.

Meta learned this in public: in 2024, its “Made with AI” label landed on real photos that had only basic edits. After photographers complained, Meta renamed it “AI info” in July 2024 — but a small edit like generative fill can still trigger it.

For a creator, a wrong label costs trust. Viewers read “AI-generated” as “not real,” and the comments turn into an argument about your footage instead of a conversation about your content.

How Remuxe writes a clean file

Remuxe doesn't scrub your file or hunt for tags. It never edits your original at all. It builds every export from scratch:

1. Decode. Your video is opened in the browser and read frame by frame, along with its sound.
2. Process. Every frame goes through the same pipeline as the preview: your edit, then the small invisible changes that make each copy unique.
3. Encode. The frames are compressed again — on your computer's hardware video encoder where it's available — as H.264, with the sound as AAC (or Opus where the browser has no AAC encoder).
4. Write a new container. Remuxe writes a brand-new MP4 (or MOV, to match a phone's format): a new file-type box, new track headers, new timing and sample tables, and a new data section with the freshly encoded frames.

There's a second reason this is the honest approach. Content Credentials are bound to the exact bytes of the file with a cryptographic hash. Once a video is re-encoded, the old manifest describes a file that no longer exists; copying it over would only produce a credential that fails validation.

What Remuxe leaves alone

• Invisible watermarks. Systems like Google's SynthID hide a signal inside the pixels or the sound themselves, not in the file's metadata. They survive a new container, and Remuxe doesn't try to remove them. Why we don't remove AI watermarks→
• Platforms' own detection. Platforms also run their own models on what they receive. A clean container doesn't change what's in the picture.
• Your content. Remuxe changes how the file is built, not what it shows.

Questions