AI ACT — ART. 50

AI Transparency

Last updated: 2026-06-07

Every visual produced by MannequinIA is content generated or manipulated by artificial intelligence. This page explains how we mark them, why, and what it means in practice for you and for the people who will see these visuals.

AI-generated visuals

The service creates images from your content and instructions using artificial-intelligence models. The result is not a photograph: it is synthetic content. By default the depicted models are fully synthetic and correspond to no real person.

How we mark images

Each generated visual is marked in a machine-readable way so that its "AI" origin survives download and any re-upload — even when a visible badge would be cropped out. Specifically:

Signed C2PA Content Credential (preserved): Every generated image carries a cryptographically signed C2PA provenance credential written by the generation model (Content Credentials). Our pipeline preserves it intact: we do not rewrite the file after that signed manifest, so its signature and binding to the pixels remain valid. Optionally, MannequinIA can add its own signature, chained onto that upstream credential.
Invisible watermark (SynthID): Outputs from the model we use (Nano Banana) carry an invisible SynthID-type watermark embedded at the pixel level. Our processing does not strip it.
IPTC provenance metadata (standard): The file also declares the standard field Iptc4xmpExt:DigitalSourceType with the value trainedAlgorithmicMedia (XMP). Where a visual would not already carry this metadata, we add it defensively (without touching the signed C2PA manifest when one is present). This metadata is a durable, machine-readable layer expected by the regulation.
Supported format: The defensive XMP marking is applied to PNG files (our pipeline's output format) via an iTXt chunk inserted before the end of the file — only when no signed C2PA credential is present. For any other format, the image is returned unchanged: we never corrupt a file.

Implementation detail: lib/ai-marking.ts (the markImageAsAi function). It is designed never to throw or corrupt an image: on any issue it returns the original file.

Article 50 of the EU AI Act

Article 50 of Regulation (EU) 2024/1689 (the "AI Act") requires, from 2 August 2026, that synthetic content (image, audio, video) be marked in a machine-readable format and detectable as artificially generated or manipulated. The preserved signed C2PA credential, combined with the invisible watermark and the IPTC/XMP metadata, is intended to meet this requirement on the provider side.

The AI Act distinguishes the system's provider (which must mark outputs) from the deployer(you, when you distribute the content). Where a visual constitutes a "deep fake" — i.e. it depicts a real person, object or place of authentic appearance — you may, as a deployer, be required to clearly disclose that the content has been AI-generated or manipulated.

A fully synthetic model is, in principle, not a deep fake of a real person. However, staging a real productof authentic appearance may, under forthcoming guidelines, fall within the "objects" category. When in doubt, disclose the AI nature of the image when you distribute it.

What this means for you

Your visuals are delivered with "AI" provenance metadata and an invisible watermark: their origin remains verifiable, even after download.
You must not remove, alter or hide these markings.
When you distribute a visual depicting a real element of authentic appearance, plan to disclose its AI nature if the regulation requires it.
You remain responsible for complying with the transparency rules applicable to your sector (advertising, platforms, etc.) and for use consistent with our Terms.

Each generated image therefore carries a signed C2PA provenance credential (Content Credentials) that MannequinIA preserves throughout its pipeline, in addition to the SynthID watermark and the IPTC/XMP metadata. As regulatory requirements and guidelines evolve, this page will be updated accordingly.