
The audio task
Before using an AI-assisted tool on a client's paid project, a producer wants to know what the tool was trained on, what it is meant for, and where it is known to fail. A vendor's marketing claim of ethical AI answers none of that. A published model card, when one exists, is a structured document a producer can actually check against the intended use.
What the documents show
Meta's MusicGen model card names its training sources directly, stating the model 'was trained on licensed data using the following sources: the Meta Music Initiative Sound Collection, Shutterstock music collection and the Pond5 music collection,' states its primary intended use as research, and lists limitations including that it 'is not able to generate realistic vocals' and performs unevenly across music styles and cultures. Google DeepMind's Lyria 3.5 model card states its training dataset was 'audio data' with datasets 'annotated with text captions,' describes data processing steps including deduplication and safety filtering, points to Google's Generative AI Prohibited Use Policy for acceptable use, and states the model applies SynthID watermarking to its output.
Rights status
Neither card is a licence by itself; MusicGen's separate CC-BY-NC 4.0 weights licence and Lyria's separate terms of service govern actual use. What the cards establish is disclosure: MusicGen names specific licensed sources, while Lyria's card describes its training data in broader terms without naming individual sources. That is a real difference in specificity, not a claim that either model's data practices are improper; a less specific disclosure is a gap to weigh, not evidence of a violation.
What to check before you use it
This is an editorial method record describing what a model card should state and how to treat its absence, built from named examples that do publish one. Look for a model card naming training data sources or categories, stated intended use and out-of-scope use, known limitations, and any watermarking or provenance signal. If no model card exists for a plugin under consideration, record that plainly in the studio's own procurement notes rather than assuming one exists, and treat the absence as an open question, not as proof of wrongdoing.
- Does this tool publish a model card, and does it name training data sources or only describe them generally?
- What does the card state as the model's intended use, and does a client project fall inside or outside that scope?
- If no model card exists, has the vendor answered the same questions anywhere else in writing?
A model card will not settle a rights dispute on its own, but it gives a producer something specific to check against a client's requirements, which a slogan cannot.
Sources & reading trail
MusicGen's model card names specific licensed training sources, states intended use as research, and lists known limitations.
Source published: Not established · Retrieved: 16 September 2026
Google DeepMind's model card describes Lyria 3.5's training data in general terms, its data-processing steps, acceptable-use policy, and SynthID watermarking.
Source published: Not established · Retrieved: 16 September 2026
Documentation, licences and platform policies establish the note; the what-to-check reading is Music Tech Field Notes editorial analysis. This retrospective draft does not imply the site published on the event date.