> ## Documentation Index
> Fetch the complete documentation index at: https://www.latentcast.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# AI transparency
> How LatentCast marks AI-generated video, and how anyone can check a file for those markings.
Every video LatentCast produces is **marked as AI-generated** — visibly, and in two machine-readable
layers that travel with the file. Anyone can check a video against those markings, free and without
an account, using our [detection tool](https://app.videostoryagent.com/verify).
This page explains what we mark, how to verify it, and — importantly — how to read a result without
drawing the wrong conclusion from it.
This supports our obligations under **Article 50(2) of the EU AI Act** and the accompanying
**Code of Practice on Transparent Generative AI Content**, which we have signed.
## How our content is marked
Three independent layers, because each survives different things.
### 1. A visible disclosure badge
Videos carry the **official EU disclosure icon** on-screen, drawn from the Code of Practice's own
icon set. It is part of the video image itself, so it survives any re-encoding, and it is visible to
a viewer without any tooling at all.
Customers choose the badge's position, size and style to suit their composition — but the icon
artwork is the standard one, so it reads the same across every provider that adopts it.
### 2. An imperceptible watermark
Each video carries a **256-bit watermark** embedded across the video frames using
[Meta Video Seal](https://github.com/facebookresearch/videoseal). It is invisible and inaudible, and
it is designed to **survive transcoding** — so a copy re-encoded for streaming, downloaded, or passed
through a social platform generally still carries it.
This is the layer that usually answers the question *"did this come from LatentCast?"* for a file
that has been through the internet.
### 3. Signed C2PA metadata
The delivered master carries a **C2PA manifest** — the open Content Credentials standard — signed
with our certificate and RFC 3161 timestamped. It records that the video was AI-generated
(`digitalSourceType: compositeWithTrainedAlgorithmicMedia`), names the producing system, and
includes a **bill of materials**: one entry per source asset that went into the composition.
The manifest is cryptographically bound to the exact bytes of the file. Change so much as a single
byte and the binding breaks — which is precisely what makes it useful for detecting tampering.
C2PA metadata is an open standard, so you do not have to take our word for it. Any C2PA-capable
inspector — for example
[verify.contentauthenticity.org](https://verify.contentauthenticity.org) — can read the manifest of
a downloaded LatentCast video.
Our signing certificate is not yet listed on the C2PA trust list (an application is in progress), so
inspectors will currently report the signature as **valid but untrusted**. That is a statement about
certificate listing, not about the integrity of the file.
## Checking a video
Open the **[detection tool](https://app.videostoryagent.com/verify)** and drop a video in. It reads
both machine-readable layers and tells you what it found.
MP4, MOV, WebM or MKV, up to 200 MB. Drag it onto the page or use the file picker.
The check reads the C2PA manifest and samples frames for the watermark. It usually takes well
under a minute.
You get a plain-language verdict, which layers matched, and a confidence level. See
[Reading the result](#reading-the-result) below.
### No account, and no tracking
The tool is **anonymous by design**:
* **No account, no login, no sign-up.**
* **No tracking.** We use no cookies and no device fingerprinting. We keep no record of who checked
what.
* **Zero retention.** Your upload is deleted as soon as the check finishes. We never store the file.
We consider this a feature rather than an omission. Someone verifying a sensitive video has a
legitimate interest in not being identified for doing so.
To keep the service available and to protect it from automated abuse, there is a **small daily
allowance per network address**, and the page runs a brief proof-of-work check in the background.
That check costs your device about half a second of computation and requires no interaction — we
deliberately chose it over a CAPTCHA, because CAPTCHAs routinely block people using VPNs or Tor, who
are often exactly the people who need this tool.
### Free unlimited access for journalists and researchers
If you are a **journalist, fact-checker, researcher, civil-society organisation or public
authority**, the daily allowance should not stand in your way.
Write to **[verify@latentcast.com](mailto:verify@latentcast.com)** and we will issue you an access
key that lifts the limit, **free of charge**. Paste it into the "Have an access key?" field on the
verification page.
An access key is a durable identifier: checks made with it are attributable to that key, unlike the
anonymous path. We count requests only — never what was checked — and the key is optional, so you
can always fall back to anonymous use.
## Reading the result
Our marking was found and is intact. The video was produced by LatentCast.
The result shows which layers matched. **Watermark + signed C2PA metadata** is the strongest
case: two independent mechanisms agreeing.
Perfectly normal — **not** a sign of tampering.
Streaming platforms re-encode video, and re-encoding strips C2PA metadata by design. That is the
whole reason the watermark exists: it is the layer built to survive the journey. A file captured
from a stream will legitimately show the watermark alone.
The watermark identifies the video as ours, **and** a C2PA manifest is present — but it no longer
matches the file's bytes. Something changed after we signed it.
Note this is different from the case above: here a signature exists and fails, rather than being
absent.
We found neither layer. **This does not mean the video is authentic.**
It means only that the file was not produced by LatentCast — or that our marking did not survive
what was done to it. Heavy compression, aggressive downscaling, cropping or re-recording a screen
can all destroy a watermark.
A null result is the absence of our evidence, not the presence of proof.
No detection system can prove a video is genuine. Ours answers one narrow question — *is this ours?*
— and it should be read as one input among several, never as a verdict on authenticity.
### Confidence
Results carry a bucketed confidence — **high**, **medium** or **low** — rather than a precise score.
That is deliberate: publishing an exact score would help an attacker measure which manipulations
weaken the watermark, and iterate towards removing it.
## Reporting a security issue
If you find a vulnerability in how our content is marked or verified — a way to forge a marking, to
strip one without detection, or to make the verification tool report something untrue — please tell
us at **[verify@latentcast.com](mailto:verify@latentcast.com)**.
We will acknowledge your report and keep you informed while we work on it. Please give us a
reasonable opportunity to fix the issue before disclosing it publicly. We will not pursue legal
action against anyone acting in good faith under this policy.
## Removing markings is not permitted
Our terms prohibit customers from stripping, obscuring or otherwise defeating any of these markings
on content produced with LatentCast. If you find a video that appears to have had its markings
removed, we would like to hear about it at
[verify@latentcast.com](mailto:verify@latentcast.com).
## Interoperable detection
The Code of Practice anticipates a shared way for anyone to route an unknown file to whichever
provider can identify it, rather than checking each provider's tool in turn. We intend to
participate in that mechanism as it becomes available, by **2 February 2027**. Until then, the tool
above and any C2PA inspector are the routes to us.