*This post was created by Epidemic Sound and published as a sponsored article. It provides general information and is not legal advice.
The EU now requires AI-generated audio to carry a machine-readable mark. While that mark identifies the output as synthetic, it doesn’t provide a track-by-track record of what the model learned from. For brands publishing the output, that leaves a separate question of training-data provenance.
In 2025, French streaming service Deezer commissioned market-research firm Ipsos to run a blind test with 9,000 people across eight countries. Each heard two tracks made entirely by AI and one made by people. Ninety-seven percent couldn’t tell which was which.
From August 2, 2026, Article 50 of the EU AI Act requires providers of systems generating synthetic audio to mark their output in a machine-readable format. Systems already on the market before that date have until December 2, 2026, to comply. California has taken a similar, though not identical, approach. Its AI Transparency Act, operative since August 2, 2026, requires large generative-AI providers to embed provenance information in generated audio, images, and video.
However, these marks say nothing about what the models learned from, or whether any of the learning material was cleared. For a listener, not knowing the origin of an AI track is a curiosity. For a brand, it’s a legal question with potentially devastating consequences.
If a brand can’t independently verify what an AI model was trained on, it also can’t independently verify whether all of that material was licensed or otherwise lawfully used. A brand that publishes the track anyway takes on a risk that may be difficult to quantify if the provider can’t document the model’s training provenance. And the brand takes it on in its own name, since the track goes out on the brand’s channels.
Why can’t the model itself show what it was trained on?
A trained AI model doesn’t contain a readable inventory of its training set. During training, examples are used to adjust large numbers of numerical parameters, but the resulting weights don’t preserve a one-to-one record showing which song changed which parameter. Looking at those weights afterwards therefore can’t reconstruct a complete list of the music used to train the model.
Some training material can nevertheless be memorized. In 2021, researchers prompted a public language model into reproducing exact passages from its training data. That proves those particular passages were in the training set but says nothing about what else was, and what a brand needs to know is that nothing unlicensed was.
A test can show limited presence; it cannot show absence.
What are the risks of using a track from a model with unknown training data?
Using a track from an AI model with unknown training data exposes a brand to three main risks: (1) a copyright claim it might not be able to defend with evidence, (2) a claim from a real person whose voice the model imitated, and (3) an audience that skips music it knows was made by AI.
- When a rights holder argues that a track resembles a protected work, the usual response is to show how the track was made: the session, the writer, the reference material. If the vendor can’t document the model’s training provenance, the brand can’t obtain that evidence from the model itself. The dispute then turns on what the two tracks sound like, and the brand argues that question with the weaker evidence.
- In Tennessee, voice imitation can create a separate legal issue from copying the underlying music. Tennessee’s ELVIS Act, in force since July 1, 2024, added voice to the attributes a person can protect, and defined voice to include a simulation, which is where AI imitation falls. If the provider can’t document the training data, a brand can’t independently establish whether the model learned from a particular singer. Additional regulatory authorities are poised to mimic this legislative framework.
- Forty percent of music-streaming users in the Deezer/Ipsos survey said they would skip a fully AI-generated track without listening to it. As machine-readable marking makes synthetic origin easier for platforms and software to detect, brands should increasingly assume that the AI-generated nature of the music will be visible rather than hidden.
How can brands use AI for music safely?
One way to reduce training-data uncertainty is to use AI on music whose provenance and rights are already documented. The distinction isn’t simply between AI and no AI, but between generating new music from a training set whose contents may be difficult to verify and using AI to discover, place, or adapt rights-cleared music from known sources.
Epidemic Sound takes a different approach to AI. Its published music is created by human artists, while its AI-powered tools are used to find, place, and adapt rights-cleared music rather than generate finished tracks from scratch. Adapt, for example, is trained exclusively on Epidemic Sound music and other licensed music.
Epidemic Sound owns the economic rights to more than 55,000 tracks, while its artists receive upfront payments, a 50/50 share of streaming royalties, and additional bonus payments. That gives the company a documented rights basis for the music its customers use, rather than asking a brand to infer provenance from the finished model itself.
Epidemic Sound’s AI features at a glance:
- Assistant finds tracks and sound effects through conversational prompts, so a user can describe a project, mood, scene, or creative brief in natural language instead of browsing manually.
- Adapt reshapes a chosen artist track from a prompt, changing length, structure, or instrumentation, so one track can be fitted to several formats.
- Studio analyzes a video and places music and sound effects against it, producing a first-draft soundtrack from tracks the license already covers.
Epidemic Sound isn’t the only way to meet that standard. Any provider that can document where its source material came from, and that it holds the rights required for the way AI is being used, can provide the same kind of provenance. A brand commissioning music directly from a composer can create a similar record of its own.
What Epidemic Sound adds is scale and an established rights model. The company has operated this way since 2009, and its tracks are heard in videos with over 3.5 billion views a day across TikTok and YouTube.
The mark on a track tells a listener that a machine made it. The agreement behind a track tells a brand whether it can be used. Only the second one answers the question that matters.
Related Categories