AI labelling & transparency
AI labelling under Art. 50 of the EU AI Act – how SovrGPT marks generated images, videos and audio with metadata, watermarks and a visible notice.
SovrGPT labels generated images, videos and audio as AI-generated: machine-readable in the file and visible in the product. This serves the transparency obligation in Art. 50 of Regulation (EU) 2024/1689 (the AI Regulation or EU AI Act), applicable from 2 August 2026.
The labelling applies on every plan and cannot be switched off. It is separate from the visible SovrGPT brand mark on videos, which is dropped from the Pro plan upwards.
Machine-readable in the file (Art. 50(2))
Generated files carry metadata with the IPTC provenance value
digitalSourceType = trainedAlgorithmicMedia, which C2PA / Content
Credentials also build on, plus the plain-text note
Generated by AI - SovrGPT (<model>). Edited images, and videos made from a start image or
as a continuation, contain your own material and therefore carry
compositeWithTrainedAlgorithmicMedia instead.
| Modality | Format | Technical implementation |
|---|---|---|
| Image | PNG | tEXt chunks (Comment, Software, digitalSourceType) + XMP packet (Iptc4xmpExt:DigitalSourceType) |
| Video | MP4 | Container metadata (comment, digitalSourceType), independent of the visible brand mark |
| Audio (TTS / voice) | WAV | RIFF LIST/INFO chunk (ICMT comment + ISFT software) |
| Audio (TTS / voice) | MP3 | ID3v2.3 tag (COMM comment + TXXX digitalSourceType) |
| Audio (TTS / voice) | FLAC | Vorbis comments (COMMENT + DIGITALSOURCETYPE) |
| Audio (TTS / voice) | Opus (Ogg) | Vorbis comments in the OpusTags header (COMMENT + DIGITALSOURCETYPE) |
You can check this with, for example, exiftool <file> or
ffprobe -show_format <file>. The AI note and the digitalSourceType appear in
the output.
Invisible watermarks
Metadata can be stripped. That is why many outputs also carry an invisible watermark in the signal itself:
- Images: a DWT-DCT-SVD watermark in the pixel data. It survives PNG re-encoding, JPEG compression (q80) and H.264 video compression.
- Videos: the same watermark on about two frames per second; it survives the H.264 encode.
- Special voices (CosyVoice) and cloned voices: a neural Perth watermark in the audio signal, inaudible and robust against re-encoding.
Visible in the product (voluntary, in addition to Art. 50)
Generated images and videos carry an "AI-generated" notice in the chat, on the medium and below it. The read-aloud button identifies the voice as AI-generated. On download, the machine-readable labelling stays in the file.
Limits
- Default voice and catalogue voices: they carry metadata; we do not add a watermark to the signal there.
- Text: answers in the chat and through the API carry no embedded labelling.
- C2PA: a cryptographically signed provenance chain (C2PA signature) is planned but not yet in place.
What this means for you
- Publishing: if you publish AI-generated content, you may have your own disclosure obligations as a deployer under Art. 50(4), especially for deep fakes, meaning deceptively realistic depictions of real people. Do not remove the labelling.
- Cloning voices: if you clone a voice through the API, you need the consent of the person the voice belongs to. The resulting clips are labelled as well.
- API: speech output through the developer API (
/v1/audio/speech) carries the embedded labelling too, with the same limits as above.