Content Labelling System

1. Purpose

This document describes how Aivon Vision ("Aivon") identifies and labels AI-generated content produced through its platform. It is published in compliance with Article 50 of the EU AI Act (Regulation (EU) 2024/1689), which requires providers of AI systems capable of generating synthetic content to ensure that outputs are marked in a machine-readable format and, where technically feasible, in a manner detectable by automated means.

2. Regulatory Context

The EU AI Act (in force from August 2024, with phased application through 2026-2027) establishes obligations for providers and deployers of AI systems that generate synthetic content including images. Key obligations relevant to Aivon include:

• Ensuring AI-generated images are marked as artificially generated (Article 50(2))
• Using technically robust methods that are machine-readable and resistant to tampering where possible
• Not suppressing or removing lawfully applied AI content labels

Additionally, the EU's upcoming AI Liability Directive and the Digital Services Act (DSA) inform our approach to transparency and accountability for AI-generated content.

3. Our Labelling Approach

Every image generated by Aivon is labelled using two complementary methods:

3.1 Embedded File Metadata

At the point of saving each generated image, Aivon embeds the following metadata fields into the PNG file using the PNG tEXt chunk standard:

These fields are readable by any image metadata tool (e.g. ExifTool, ImageMagick, browser developer tools) and by automated content detection systems that inspect file metadata.

3.2 Platform-Level Labelling

Within the Aivon platform, all images in a user's library and in the community gallery are displayed with a visible "AI Generated" label. Download filenames follow the pattern aivon-[id].png to provide an additional contextual signal of origin.

4. Model-Level Provenance

Aivon currently uses FLUX.1-schnell by Black Forest Labs, accessed via the HuggingFace Inference API. Black Forest Labs has committed to C2PA (Coalition for Content Provenance and Authenticity) signing in their direct API and commercial offerings. However, when images are generated through the HuggingFace Inference API route, C2PA signing by the model provider is not guaranteed to be preserved in the returned image bytes.

For this reason, Aivon applies its own metadata labelling server-side (described in Section 3.1) as the primary compliance mechanism, independently of any model-level provenance.

5. Limitations

We acknowledge the following limitations of our current labelling system:

• Metadata can be stripped: A user who downloads an image can remove PNG metadata using freely available tools. We cannot prevent this. Our obligation is to embed the metadata at generation; we are not responsible for its removal by third parties.
• No visible watermark: We do not currently apply a visible watermark to generated images, as this significantly degrades image quality and usability. We are monitoring regulatory guidance on whether visible watermarking becomes mandatory and will implement it if required.
• No cryptographic C2PA signing: Full C2PA manifest signing with a certificate authority is not yet implemented. This is on our development roadmap for 2026-2027 as C2PA tooling matures and as regulatory requirements are clarified.

6. Background Removal Outputs

Images processed through our background removal feature (AI-assisted, using the rembg library with the U2Net model) are not considered "AI-generated content" in the sense of Article 50 of the EU AI Act, as the primary creative content originates from the user. However, the processed output file does include a metadata field identifying that background removal was performed by Aivon.

7. Roadmap

We are actively working toward the following improvements to our content labelling system:

• C2PA manifest embedding - full cryptographic signing of AI-generated images using the C2PA standard, targeted for implementation by Q4 2026
• Invisible watermarking - evaluation of invisible/steganographic watermarking techniques that do not degrade image quality
• User disclosure tools - in-app tools to help users understand their obligations when sharing AI-generated content publicly

8. User Obligations

Users of Aivon agree in our Terms of Service not to remove, alter, or misrepresent the AI-generated nature of content produced on our platform. Specifically:

• Users must not present AI-generated images as authentic photographs of real events without clear disclosure
• Users must not strip metadata labels and republish images in a way that falsely implies human authorship
• Users must comply with applicable national laws regarding AI-generated content disclosure in their jurisdiction

9. Verification

Anyone wishing to verify whether an image was generated by Aivon can inspect the file metadata using a tool such as ExifTool:

exiftool image.png

Images generated by Aivon will show Software: Aivon AI Image Generator and Comment: AI-generated content in the metadata output.

10. Updates to This Policy

This policy will be updated as our technical implementation evolves and as EU AI Act implementing acts and guidance are published. Material changes will be announced on our platform and reflected in the version number and effective date at the top of this page.

11. Contact

Questions about our content labelling practices can be directed to [email protected] or via our Contact page.

UAB AIVON
Pelesos g. 17-14, Vilnius LT-02114, Lithuania
Company code: 307619983