AI Act music: what the European regulation changes for you
The AI Act music primarily involves two sets of obligations. Transparency over generated content applies from August 2, 2026, and major models must comply with copyright rules.
The regulation does not state who owns a song created with AI. However, it changes what you can know, demand, and must display.
- The dated schedule of the AI Act
- Article 53: copyright and training data
- The code of good practices and its signatories
- Article 50: tagging and reporting generated content
- Four situations for musicians
- Questions to ask your tool
- What the AI Act does not address
- Checklist for your catalog
- What TuneLockr Does, and Does Not Do
- FAQ
The dated schedule of the AI Act music
The Regulation (EU) 2024/1689 dated June 13, 2024. Its obligations come into effect step by step. The table summarizes those relevant to a musician, according to the Commission pages consulted on November 4, 2026.
| Date | What applies | Source |
|---|---|---|
| August 2, 2025 | Obligations of general-purpose AI model providers (Articles 53 and following) | Commission, AI Act page |
| July 10, 2025 | Publication of the code of good practices for these models | Commission, GPAI code |
| July 24, 2025 | Official summary model of training data | Commission, news |
| June 10, 2026 | Final version of the code on transparency of generated content | Commission, transparency code |
| July 20, 2026 | Final guidelines on Article 50 | Commission, library |
| July 27, 2026 | Entry into force of the AI omnibus (postponement of 'high-risk' rules) | Commission, AI Act page |
| August 2, 2026 | Transparency (Article 50) and Commission's sanctioning powers over models | Commission |
| December 2, 2026 | Tagging (Article 50, paragraph 2) for systems marketed before August 2, 2026 | Lewis Silkin, July 9, 2026 |
| August 2, 2027 | Compliance of models marketed before August 2, 2025 | Commission, FAQ |
The 'high-risk' AI rules have been postponed to December 2, 2027, and August 2, 2028, by theAI omnibus. They have little impact on music creation.
Article 53: copyright and training data
Article 53 imposes two obligations on general-purpose AI model providers. They directly concern music rights holders.
- A Union copyright compliance policy. It must particularly respect the rights reservations provided for in Article 4, paragraph 3, of the directive 2019/790.
- A sufficiently detailed public summary of the content used for training, according to the model published by the Commission on July 24, 2025.
According to the Commission, this summary should help rights holders exercise their rights. It does not necessarily list every work: the FAQ page does not specify this.
Breaches can cost up to 3% of global turnover or 15 million euros, whichever is greater (Commission questions and answers). The Commission can impose these fines since August 2, 2026.
Are music generators concerned?
It's not certain for all. The Commission's guidelines retain an indicative criterion: more than 10^23 calculation operations for training. The model must also be able to generate language (text or audio), images, or video.
A model that only produces music is not explicitly targeted by this criterion. The issue must therefore be assessed service by service. The obligations of Article 50 apply to any system that generates synthetic audio.
The role of opt-out
The 2019 directive allows text and data mining unless the holder properly objects. On October 13, 2023, SACEM announced it was exercising this opt-out for its members' repertoire, according to Complete Music Update.
If you are an independent author, also express your reservation yourself: mentions on your site, terms of your licenses, metadata. Our page on AI training data in music details these options.
The code of good practices and its signatories
The code of good practices for general-purpose AI was published on July 10, 2025. It includes three chapters: transparency, copyright, and safety and security. Adherence is voluntary.
The consulted list includes OpenAI, Google, Microsoft, Anthropic, and Mistral AI. xAI only signed the safety and security chapter. No specialized musical service is included, but the page notes that some signatories may not yet appear.
Article 50: tagging and reporting generated content
TheArticle 50 applies from August 2, 2026. It distinguishes two actors: the tool provider and the person using it to publish, called the 'deployer'.
- The provider of a system that generates audio must tag outputs in a machine-readable format so they are detectable as generated.
- The deployer who publishes a deepfake must indicate that the content has been generated or manipulated.
- For a manifestly artistic work, it suffices to appropriately signal the content without hindering the presentation of the work.
The code on transparency of generated content was finalized on June 10, 2026. It defines deepfake as an image, audio, or video resembling existing people or events. A synthesized voice imitating a real singer may thus fall within the scope.
The guidelines on Article 50 were published on July 20, 2026. According to Lewis Silkin, systems marketed before August 2, 2026, must ensure this tagging by December 2, 2026.
Four situations for musicians
The beatmaker selling beats made with a generator
Example: Sarah sells instrumentals online that are partially generated. The tool she uses must tag its outputs. She generally has no labeling obligation unless it is a deepfake.
Her real issue remains contractual: do the tool's terms allow resale? The comparison of generators for commercial use responds tool by tool.
The duo publishing a clip with the voice of a deceased singer
Example: an electro duo has a chorus sung by a synthesized voice similar to a famous deceased singer. It is an audio deepfake.
Even in an artistic work, the duo must indicate the generated nature of the content, for example in the description and credits. Image and voice rights issues are additional, outside of the AI Act.
The publisher wanting to know if their catalog was used
Example: a small music publishing house suspects a model of having trained on its titles. They can read the public training summary published by the provider, if it falls within the scope of Article 53.
This summary remains general. To go further, they will need their dated elements and, in case of dispute, an intellectual property lawyer.
The studio delivering a commercial music piece to a European client
Example: a studio delivers a jingle with a generated voice. The client broadcasts the advertisement in France.
If the voice does not imitate anyone, tagging is not necessarily required as a deepfake. However, the contract should specify the generated part. Our page on the AI clause in a music contract offers phrasing.
Questions to ask your musical AI tool
No need to read the entire regulation to use it. A few specific questions to your tool provider can reveal a lot about their stance.
- Is your model a general-purpose AI model under the AI Act, and where is your training summary?
- Does your copyright policy respect machine-readable opt-outs, and how?
- How are my generated tracks marked, and since when?
- Do your terms allow me to credit AI usage without losing my rights?
Keep the answers, or lack of answers, with their date. If a provider doesn't say whether their outputs are marked, assume that a platform may someday flag your files.
What the AI Act does not address
The AI Act music does not create any copyright on a generated song. Protection of a work remains a matter of national law, in France originality as per the Intellectual Property Code.
It also does not create compensation for authors whose works were used for training. License negotiations and ongoing lawsuits are happening elsewhere. See our page on the AI music copyright.
AI Act music checklist for your catalogue
| Action | Why | Record to keep |
|---|---|---|
| Express your reservation of rights | Article 4, paragraph 3, of Directive 2019/790 | Dated capture of your mentions |
| Read the model training summary | Article 53, paragraph 1, point d) | Dated copy of the summary consulted |
| Report any voice imitating a real person | Article 50, paragraph 4 | Description, credits, captures |
| Keep tool outputs with their date | Show what comes from you and the tool | Exports, successive versions |
| Specify AI permissions in your licenses | Frame training and generation | Signed licenses |
To frame who can train a model on your tracks, see our page on the music license and artificial intelligence.
What TuneLockr Does, and Does Not Do
TuneLockr helps you document your works in response to questions posed by the AI Act. Each deposit associates a date, a SHA-256 hash, and deposit information with a version, along with a certificate.
- What TuneLockr does: the certificate helps establish a timeline and can be kept among useful elements in case of a dispute.
- Licenses: the Creator and Pro offers allow licenses with AI permissions, "AI Training" and "AI Generation."
- What TuneLockr does not do: it does not mark your files within the meaning of Article 50. It does not automatically detect generated music. A detection function is announced, not yet available.
- What it does not replace: SACEM, another collective management organization or a lawyer. It does not alone determine the definitive ownership of rights.
The first deposit is free, no credit card required.
AI Act music FAQ
Do I have to label each track made with AI?
Not necessarily. The marking obligation primarily rests on the tool provider. Your information obligation mainly targets hyper-realistic fakes, like a voice imitating a real person.
Does the AI Act apply to an American tool?
The territorial scope is found in Article 2 of the regulation. It does not solely depend on the provider's country: read it with a lawyer if the issue is significant.
Can I get the list of works used for training?
The public summary is "sufficiently detailed" but not necessarily exhaustive. The consulted pages do not state it must name each work.
Does the AI Act give me rights to a generated song?
No. It does not address ownership. That depends on national copyright law and your human contribution.
This page is general information, not individual legal advice. Facts verified on November 4, 2026.