TikTok can identify songs. It can use filters to contour your face. It can serve up weirdly addictive videos about things you never knew you cared about, like trimming cow hooves.
What it can’t do is tell you where that video you love came from originally.
And that’s the problem. We have sophisticated platforms with billions of engaged users uploading tens of thousands of videos every minute, yet tracing a clip back to its source can become remarkably difficult once it’s been clipped, filtered, compressed, and reposted.
Short-form video is a particularly tough environment for reliable watermarking. The difficulty isn’t just technical. The format itself is built around the very transformations that can make watermarks harder to detect.
Would you like some compression with your video clip?
Most video platforms reencode videos on upload. There’s nothing malicious here. They need to do it to optimize for users’ different devices, resolutions, and bandwidths.
That said, the way platforms process video means a 15-second clip can pass through multiple rounds of transcoding, starting with the creator’s original file and potentially being transformed again each time it’s downloaded and reuploaded elsewhere. With every lossy pass, detail can degrade as pixel values shift and compression artifacts accumulate.
The result is a particularly demanding environment for forensic watermarking. A watermark embedded in those pixels has to survive compression, cropping, scaling, and other processing common to online sharing platforms. Recent research shows that these transformations can make watermark detection more difficult.
When every frame counts
Like a door with multiple locks, a watermark can be more resilient when there’s redundancy. The more video frames you have to spread a signal across, the more opportunity a detector has to distinguish it from noise.
Short-form video gives you much less room to work with.
A 10–30 second clip is a narrow window for embedding a signal that needs to remain detectable before the content ends. And that window can shrink even further when creators loop, trim, or clip the video before reuploading it.
This is where the embedding design matters. Some watermarking approaches spread a signal across multiple frames to make it easier to recover. But that creates a synchronization problem. The camera and source video may run at different frame rates, while dropped or repeated frames can throw the signal out of alignment. For very short clips, there’s simply less information for the detector to work with.
One alternative is to embed the watermark in each frame individually. That way, recovery doesn’t depend on a sequence of frames staying intact or in order.
This approach can be particularly useful for short clips. If each frame carries its own watermark, a mark could potentially be recovered even after a video has been cut down, looped, or heavily edited. That removes the need for multiple frames to work together, a useful advantage when short-form content is constantly being edited and recut.
And that’s before anyone starts changing the video itself.
Heavy editing is the default, not the exception
Long-form video watermarking generally has to survive re-encoding and maybe some cropping. Short-form video has to survive a creative culture built around changing the original. Each edit can introduce another transformation that a watermark needs to withstand.
And a watermark that survives a straight reupload might not survive multiple rounds of editing.
That’s not just a theoretical problem. Researchers have tested how watermarks hold up against the kinds of processing that happen on social platforms, including transcoding, cropping, frame dropping, recompression, and combinations of these attacks. Research from Sichuan University, for example, specifically examines whether video watermarks remain detectable after social-platform transcoding and hybrid attacks.
For short-form video, that means surviving the initial upload isn’t enough. A forensic watermark has to endure what happens to the video after that, as it gets clipped, edited, compressed, and passed from one creator to the next.
Recapture happens outside your control
Let’s say you’re a studio with a trailer or a clip from an upcoming movie. It’s a clean, high-bitrate master, the best it’s ever going to look. After that, control starts to slip.
First your video shows up on press sites, still looking good! But then fan accounts start clipping and resharing it. Next thing you know, reaction channels are recording picture-in-picture versions and republishing those recordings as new content.
Pirated copies can follow the same path. Each step can add another round of compression and other transformations, exactly the kind of journey a forensic watermark needs to survive if it’s going to trace a copy back to its source.
And then there’s screen recording, where the problem gets even harder. Instead of downloading the file, someone records a clip playing on a monitor or phone. The video now has to survive a display, a camera, and another round of compression before it becomes a digital file again.
That’s a very different challenge from a clean digital-to-digital transfer, since the watermark has to remain detectable even after the video passes through a screen and camera before being recaptured.
Forensic watermark and metadata are not the same thing
A lot of confusion in this space comes from treating “watermark” and “metadata” as interchangeable. The truth is, they solve different problems.
Metadata is information attached to a file, such as C2PA Content Credentials, EXIF data, or tags identifying the creator or capture device. Think of it as information about the content, rather than part of the content itself.
That distinction is becoming more important as AI transparency rules take effect. Regulations, including the EU AI Act, increasingly require AI-generated or manipulated content to be identifiable, with machine-readable information playing an important role in how that transparency is implemented.
But metadata lives in the file, which means it can be lost when content is converted, edited, exported, or uploaded to a platform that doesn’t preserve it. A C2PA manifest can be perfectly valid when content is created and still be missing after the content has passed through several processing steps.
Forensic watermarking on the other hand embeds an imperceptible signal directly into the pixels or audio of the content. The mark travels with the media itself, rather than depending on the original file structure staying intact.
That doesn’t make forensic watermarking foolproof. The mark still has to survive the things that happen to content in the real world, from compression and cropping to editing and format changes. It may also need to withstand deliberate attempts to disrupt detection. Researchers have developed different ways to make watermarks more resilient, but keeping a mark detectable through all of these changes is still a not-inconsiderable problem.
And that’s the trade-off. A forensic watermark has to be invisible to viewers while remaining detectable after the content has been changed. It also has to work fast enough to embed and detect markers at scale. Short-form video makes all of that harder by combining short clips with heavy processing and constant re-editing.
Why this matters
As AI-generated video and rapid content reuse accelerate, knowing where a clip came from becomes more important, not less. Metadata can provide valuable provenance, but it doesn’t always make the journey with the content.
Forensic watermarking offers a different layer of protection: a signal embedded in the content itself. The challenge is making that signal survive the messy, highly edited journey short-form video is likely to take.
That’s the problem Steg.AI is working to solve: making content traceable even after it leaves its original environment and starts living the life of a short-form video.