AI transparency

1. Creative Credits grid

Each release on SynthCamp ships with a Creative Credits grid that the artist signs at publish time. The grid declares, on a per-strata basis, which contributions were made by a human (lyrics, melody, vocals, arrangement, mixing, mastering) and which AI tools were used (for instance Suno, Udio, ElevenLabs). Each strata definition is shown as a non-dismissible tooltip during upload to prevent puffery: prompting an AI tool to draft the lyrics does not qualify as writing them.

The artist signs an adaptive attestation that pairs a base declaration with one appendix per human strata they ticked, plus a final global signature. The signed bundle (rendered HTML, timestamp, IP, user agent, template version) is retained for ten years for audit and fraud prevention.

2. Machine-readable metadata embedded in the audio

Article 50 § 2 of the AI Act asks for machine-readable marking of AI output, and SynthCamp marks its own files too. The MP3 preview of every track carries these ID3v2.4 user-defined text frames, written by our encoder. The full stream has them too, but it's encrypted, so the preview is the one to check:

• creative_credit: the legacy triad, shown to listeners as Mostly Human / Human + AI / AI-Crafted and derived from the grid above.
• human_contributions: comma-separated list of human strata ticked by the artist (empty string for fully AI-generated releases).
• ai_tools: comma-separated list of AI tools declared by the artist.
• platform: always synthcamp.net. This is a plain convenience label. On its own it is not proof of origin (it can be copied). Cryptographic proof of origin comes from the signed payload described below.
• attestation_signed: true when the artist signed the declaration, false otherwise.

The HLS playlist (playlist.m3u8) also carries a human-readable provenance comment of the form # SynthCamp AI provenance: ai_crafted; tools: suno; human: none, so the same data can be inspected with a plain text editor. This plain-text line carries no signature and is refreshed each time the preview is signed again.

3. Signed provenance

On top of the legacy frames described in section 2, every encoded asset carries a signed canonical payload so a third-party verifier can confirm the metadata was produced by SynthCamp and not forged by an upstream tool that just wrote a TXXX frame:

• synthcamp_provenance: base64-encoded canonical JSON with release_id, track_id, artist_id, the credit grid above, the attestation timestamp, an upstream_c2pa flag and the encoder timestamp.
• synthcamp_signature: base64-encoded Ed25519 signature over the canonical JSON bytes (RFC 8032).
• synthcamp_key_id: the identifier of the keypair to look up in our public registry below.
• upstream_c2pa: true when the source asset carried a JUMBF / C2PA box at scan time. The value is part of the signed payload so a forged TXXX frame cannot upgrade an unsigned upload to "C2PA-signed" without breaking the signature.

The matching public keys are published at /.well-known/synthcamp-keys.json. The full verifier specification is available at /docs/provenance-spec.

4. Watermark preservation

Several upstream AI generators (Suno, Udio) embed an inaudible perceptual watermark in the audio data itself. SynthCamp does not strip or attenuate that watermark: the encoding chain that produces our streaming and preview assets is intentionally transparent (no loudnorm, no normalization filter on top of the AAC or MP3 encoder). If the upstream tool watermarked the output, that signal survives the SynthCamp pipeline and remains detectable by any third-party inspection tool that supports the relevant scheme.

5. User-facing AI badge

Every release page (/r/[slug]) displays a Creative Credits panel listing the strata and tools declared by the artist. A release with no human part declared carries the AI-Crafted label. A release that recreates the voice of a real person says so in its own notice: “Includes a recreated voice of a real person, used with permission.” The same information appears in the pre-purchase popover and in any embed served from the platform, so the AI nature of a release is always disclosed before a listener starts streaming or pays.

6. AI tool commercial rights

The table below summarises the commercial-use regime for each AI tool we recognise in the Creative Credits picker. The artist attestation captured at upload is what binds the artist. This table is informational. Seven regimes apply. Snapshot at generation: a paid plan held when the track was generated grants commercial use rights for that output. Snapshot at download: the same, plus, from the date the tool's terms set, a download made on that paid plan within its allowance, whose rights stay after the plan ends. Licence at download: the licence comes with the download, on the plan held at that moment. Downloaded before a date: only files downloaded before the date the tool's terms changed may be released. Continuous subscription: commercial use depends on the artist holding an active plan and may be revoked on downgrade or cancellation. Non-commercial only: the default vendor weights are CC-BY-NC and cannot be sold here, so only outputs from a permissively-licensed fine-tune are allowed. Can't be released here: the tool's licence forbids selling the music and releasing it on music services, so a release declaring it can't be signed or published here.

Plan names follow each vendor's own naming. Links point to the most recent terms of service we reviewed; SynthCamp does not guarantee they reflect a vendor's current terms after the last reviewed date.

7. Legal references

• Regulation (EU) 2024/1689 of 13 June 2024 (AI Act), article 50, in particular paragraph 2 (machine-readable marking of AI-generated content) and paragraph 4 (user-facing information). Entry into application: 2 August 2026.
• French Consumer Code, article L. 111-7: precontractual transparency obligations for online platform operators.

8. Transparency report

As the platform grows, we will publish a transparency report on this page covering AI on SynthCamp: the share of releases per Creative Credits category, the AI tools most frequently declared, the number of attestation discrepancies detected and resolved, and any moderation action taken on AI grounds. We do this voluntarily, ahead of the company-size thresholds that make the DSA transparency reporting of articles 15 and 24 mandatory for a platform our size. The first edition will follow our first full year of public operation.

Related: moderation transparency

This page covers AI transparency (AI Act art. 50 and L. 111-7). The separate moderation transparency report required by DSA article 24(1), covering orders received from authorities, internal-initiative moderation actions, complaint volumes and automated-means usage, lives at /legal/transparency/moderation. The rules of conduct that govern those moderation actions are at /legal/community-standards.

9. Contact

Questions about AI transparency, the metadata embed, or the Creative Credits grid: write to [email protected]. Communications are accepted in French and English.