AI Disclosure
Which AI models we use, how output is labelled, and what AI music can and cannot guarantee.
Last updated 2026-10-28
1. Everything is AI-generated
All songs, stems and lyrics suggestions on KeepYourSong are generated by artificial intelligence. No song is performed or written by a human artist.
2. Models we use
- Song generation (HD): Google Lyria 3 Pro, accessed through Google Cloud Vertex AI under Google's commercial terms.
- Prompt Enhancer (style tags and lyrics): Google Gemini models on Vertex AI.
- Stem separation: Demucs, an open-source model from Meta Research, run on Replicate or fal.ai.
- We may add or replace models; this page is updated when we do. Models we launch stay available, so songs do not change retroactively.
3. Labelling and watermarks
Audio generated by Google Lyria carries SynthID, an inaudible watermark that identifies it as AI-generated. We also write AI-generation information into file metadata where the format allows. Please do not remove these markers; some countries (including the EU under the AI Act) require AI-generated media to be identifiable.
4. Limits of AI music
- Output may occasionally contain mistakes, unexpected lyrics or artefacts. Review it before publishing.
- AI output may not be eligible for copyright protection in some countries, including the United States, where purely AI-generated works are not registrable.
- Streaming and distribution platforms set their own rules for AI-generated music. Check their policies before uploading.
- We prevent requests that imitate real artists; the models are not trained to reproduce specific performers' voices.