Skip to main content
MakeYourMusicMakeYourMusic

AI Training & Transparency

Last updated: July 8, 2026

This page describes how MakeYourMusic uses AI models, what data those models were trained on, and how rightsholders can exercise opt-out rights under the EU AI Act (Article 53), the EU Copyright Directive (TDM exception, § 44b UrhG and equivalents), and analogous laws.

1. Models we use

MakeYourMusic does not train its own foundation models. All AI generation is performed by third-party providers under their own licensing terms.

  • Music generation and lyrics — MiniMax (e.g. music-2.x /music-3.0 and reasoning chat models for lyric craft). Provider terms apply to training-data provenance.
  • Cover art, preview video, and long-form music video — Grok Imagine (xAI) by default (grok-imagine-image-quality, grok-imagine-video-1.5). MiniMax image/Hailuo video remain optional fallbacks when explicitly enabled. Provider terms apply.
  • Transcription — OpenAI Whisper (gpt-4o-mini-transcribe).
  • Stem separation — Demucs via Replicate when configured (open-source MIT-licensed model trained on MUSDB18).

Every track generated on MakeYourMusic has a corresponding AI Credit metadata record attached at the time of publication, which captures the provider name, model version, and a hash of the user-supplied prompt. This metadata is surfaced on the track API and accompanies the track when it is distributed to streaming services (DDEX AI disclosure tags, Apple Music Transparency Tags).

2. User-uploaded audio & opt-in

Audio you upload to MakeYourMusic (reference audio, voice samples, original tracks) is not used for training any AI model — ours or our providers' — unless you explicitly opt in via Settings → Privacy → AI Training. The opt-in toggle is off by default and can be revoked at any time, with the change taking effect for all future training runs.

3. Rightsholder opt-out (TDM)

Under § 44b UrhG (Germany), Article 4 of the EU Copyright Directive, and analogous opt-out mechanisms in other jurisdictions, rightsholders may reserve their works against text-and-data mining for AI training purposes.

If you are a rightsholder and wish to opt your catalog out of any AI training or fingerprint matching on this platform, you can either:

  • Send a request to legal@makeyourmusic.com with the URL(s) or ISRC(s) of the works in question, proof of authority, and the desired scope (training, output filtering, or both), or
  • File the opt-out programmatically by POSTing to /api/legal/tdm-opt-out with { url|isrc|upc|fingerprintSig, rightsholderName, rightsholderEmail, scope }. We email a verification link to the address provided; the opt-out becomes enforceable on click.

Opt-outs are honoured within 30 days of verification. We confirm receipt within 5 business days.

4. Output watermarking & provenance

Per Article 50 of the EU AI Act (application from August 2, 2026), AI-generated content should be marked in a machine-readable way. MakeYourMusic already stamps AI Credit metadata on every published track (provider, model, prompt hash) and includes DDEX / Apple transparency tags on distribution payloads.

Current watermark status: a machine-readable audio waveform watermark (e.g. SynthID Audio, Audible Magic, or an in-house marker) is only applied when an operator configures WATERMARK_PROVIDER in production. When that provider is not configured, tracks are still labeled with AI Credit metadata and transparency tags, but the audio itself is not steganographically watermarked. We do not claim universal SynthID coverage unless that provider is live. Health and admin surfaces report whether watermarking is configured.

Cover-art automated NSFW classification is similarly opt-in via MODERATION_PROVIDER; without it, policy review is manual / best-effort. Audio copyright fingerprint matching against known catalogs requires a fingerprint provider (e.g. AcoustID tooling) to be installed and configured.

5. Voice impersonation

Prompts that request the cloning, imitation, or impersonation of an identifiable artist's voice are rejected at the input stage. See our Voice Rights Policy for the policy in full and the forthcoming licensed voice marketplace.

6. Training-content summaries

We do not train our own foundation models, so we do not publish a training-content summary under Article 53(1)(d) of the AI Act. The providers we use are responsible for their own AI Act compliance and publish their own training-content summaries:

7. Changes to this page

We update this page when our AI provider set, models, or training policies change. Material changes are announced in-product and via email to verified account holders.

8. Contact

Questions about AI training, opt-outs, or provenance: legal@makeyourmusic.com.

AI Training & Transparency | MakeYourMusic