Every image and video a Veizik build writes is marked, in the file and in its pixels, as generated by an AI system. The builds offered today write text, which carries no mark.
An image or video build writes two marks into each output: standard AI-generation metadata in the file (the IPTC digital source type trainedAlgorithmicMedia, with the engine and model family), and an in-pixel provenance signal that survives ordinary re-encoding, resizing and cropping. Both state that an AI system produced the media. Neither is visible on the picture, and nothing is drawn onto it.
Every build that writes images or video — none is offered without the mark. The builds offered today generate text, images, and video; text is not a container format and carries no mark, while image and video builds carry the provenance mark. Each release’s notes say what that build writes.
Yes. On an image or video build it is always on and cannot be switched off. It is not a paid feature. It is required of us by law (EU AI Act Art. 50(2); Korean AI Framework Act Art. 31(2)) — see Terms §05.
The current builds do not attach signed C2PA Content Credentials. When a build adds them, its release notes say so and this page explains how to verify them.
No prompt text, no personally identifying data, and nothing that was not already part of the render configuration.
Veizik follows the AI transparency rules of the EU AI Act and Korea’s AI Framework Act wherever it is used. This is how each duty is carried, and whose duty it is.
Korean AI Framework Act Art. 31(1). The product says up front that it is a generative AI system — on this site, in Terms §05 and before you install it.
EU AI Act Art. 50(2); Korean Act Art. 31(2). Every image or video build writes a machine-readable mark into each output: AI-generation metadata (IPTC trainedAlgorithmicMedia) and an in-pixel signal that survives re-encoding. It cannot be switched off, and no such build is offered without it.
EU AI Act Art. 50(4); Korean Act Art. 31(3). When you publish AI content that could pass as real, the visible disclosure to your audience is yours to make, as the deployer. Terms §05 sets this out.
Language models write text, which is not a container for a mark. The duty to say the system is AI applies to them, and is met as above.
The metadata in a file written by an image or video build can be read with any XMP or IPTC reader, e.g. exiftool -XMP-iptcExt:DigitalSourceType <file>, which prints trainedAlgorithmicMedia. Metadata is easy to strip; the in-pixel signal is the layer that stays with the picture after re-encoding.
Removing, stripping or weakening the provenance mark is not permitted under Terms §05 and may breach the AI transparency law that applies where you publish. It does not unlock any capability and is not needed for any supported workflow. Where you publish AI-generated content that could pass as real, the visible disclosure to your audience is yours to make.
The mark is machine-readable and invisible. How AI-generated content is labelled for its audience is governed by the laws of each country or region where it is used or published, and by the policies of the employer, client, platform or service provider it passes through. Meeting those requirements — including any visible label, disclosure or watermark they call for — is the responsibility of whoever publishes the content. See Terms §05, Compliance with laws and policies.
Related: Terms §05 · Model Licences · Privacy