AI labelling & transparency

How SovrGPT labels generated images, videos and audio under Art. 50 of the EU AI Act.

All images, videos and synthetic audio outputs produced with SovrGPT are labelled as artificially generated. This implements the transparency obligations of Art. 50 of Regulation (EU) 2024/1689 (the "AI Regulation" / EU AI Act, applicable from 2 August 2026).

The labelling is independent of your plan: it is a compliance property of the platform and cannot be disabled or removed in any subscription. It is separate from the optional SovrGPT brand overlay on videos, which is dropped in higher tiers.

Two levels of labelling

1. Machine-readable (Art. 50(2))

Every generated artefact carries embedded metadata with the IPTC provenance value digitalSourceType = trainedAlgorithmicMedia — the interoperable vocabulary that C2PA / Content Credentials also build on — plus a plain-text note Generated by AI - SovrGPT (<model>).

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 overlay
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`)

You can check the labelling with, for example, exiftool <file> or ffprobe -show_format <file> — the AI note and the digitalSourceType appear in the metadata output.

Robust invisible watermarks

In addition to the metadata (which an attacker could strip), the outputs carry an invisible watermark in the signal itself:

Images: a DWT-DCT-SVD watermark in the pixel data — survives PNG re-encoding, JPEG recompression (q80) and H.264 video compression.
Videos: the same watermark on regularly distributed frames (about 2 marked frames per second) — survives the H.264 encode.
Voice (CosyVoice, voice cloning): a neural Perth watermark in the audio signal — inaudible, robust against re-encoding.

Remaining gaps (documented, on the roadmap): OGG/Opus outputs currently carry no metadata marking; the Supertonic default voices carry metadata but not yet a signal watermark; an end-to-end C2PA signature (cryptographic provenance chain) is planned.

2. Visible in the product (Art. 50(4))

Generated images and videos carry an "AI-generated" badge directly on the medium and in the footer; the read-aloud function identifies the voice as AI-generated. On download, the machine-readable labelling stays in the file.

What this means for you as a user

Redistribution: if you publish AI-generated content, you may be subject to your own disclosure obligations under Art. 50(4) of the EU AI Act as an operator — particularly for deep fakes (deceptively realistic depictions of real people). Do not remove the labelling.
Voice cloning: cloning voices through the API requires the consent of the person concerned. The resulting clips are machine-readably labelled as well.
API outputs: speech output generated through the developer API (/v1/audio/speech) carries the embedded labelling too.