The EU AI Act's training rules are now enforceable. What musicians can actually do with them

On 2 August 2026 the European Commission's AI Office started enforcing the AI Act against providers of general-purpose AI models. The duties themselves are a year old. What changed is that breaking them now costs money. For musicians, two of those duties matter: every such provider must keep a copyright policy, and must publish a "sufficiently detailed summary" of the content it trained on.

This is a practical guide, not legal advice. If you are weighing a claim, talk to a lawyer or your collecting society.

What the law now requires

Article 53 of the Act sets two obligations that touch music directly. Point (c) requires a policy to comply with EU copyright law, including identifying and respecting a "reservation of rights" made under Article 4(3) of the 2019 Copyright Directive. That is the EU's opt-out from the text-and-data-mining exception. Point (d) requires providers to publish a summary of their training content using a template from the AI Office. Neither duty goes away for open-source models: the open-source exemption in Article 53(2) covers only points (a) and (b).

Article 101 lets the Commission fine a general-purpose AI provider up to 3% of worldwide annual turnover or EUR 15 million, whichever is higher, where it acted intentionally or negligently. Most of the general-purpose AI chapter has applied since 2 August 2025, but Article 113 carves Article 101 out of that date. It falls under the Act's general start date of 2 August 2026. That is why 2 August 2026 was the real start.

There is one big exception. Under Article 111(3), models placed on the market before 2 August 2025 have until 2 August 2027 to comply. Any model released before that date gets the longer runway.

What you can actually do

1. Read the summaries as they appear. The Act says the summary must be publicly available, and Music Business Worldwide notes the duty covers any model offered in the EU, even one trained elsewhere. The Article 53 text we read does not name a central registry. Our suggestion: look in the provider's own model documentation, and keep a dated copy, because summaries can change.

2. Make your reservation machine-readable. The Act points back to the Article 4(3) opt-out. The general-purpose AI Code of Practice's copyright chapter, which signatories use to show compliance, commits them to crawlers that follow robots.txt (IETF RFC 9309) and other appropriate machine-readable protocols. It also says a reservation can be made "in any appropriate manner, such as machine-readable means" for content online. If you run your own site, that means robots.txt rules plus a clear written reservation.

3. Check the crawler list. Code signatories commit to publish which web crawlers they use and how those crawlers handle reservations, and to offer a way for rightsholders to be notified when that information changes. Subscribe where you can.

4. Use the complaint routes. Signatories commit to a point of contact and an electronic complaint mechanism that rightsholders and their representatives, explicitly including collective management organisations, can use. Separately, the Commission says anyone, person or organisation, can report alleged breaches by providers the AI Office supervises through its new Complaint Tool.

5. Ask who has reserved on your behalf. Most of your music doesn't live on your own website. According to MBW, Sony Music and Warner Music Group have already withdrawn consent for training on their recordings and lyrics. If you are independent, ask your distributor and your collecting society in writing what reservation they have made, and where.

What it doesn't do

The summary is a summary. It is not a list of works, and it is not a payment. The music industry coalition of IFPI, CISAC, ICMP and IMPALA argued last year that the template would not deliver "sufficient" transparency, MBW reports. Compliance with the AI Act also doesn't settle copyright. On 31 July the Munich Regional Court ruled against Suno in GEMA's case (case 42 O 763/25). DLA Piper's analysis says the court rejected Suno's argument that meeting its Article 53 duties, including the training summary, made a licence unnecessary. The AI Act complements copyright law and doesn't authorise use of protected works by itself. The ruling is first-instance and not final, and MBW reports Suno is considering an appeal.

Artists first

There is a practical inequality here. A major label can withdraw consent for its whole catalogue in one statement. An independent artist has a robots.txt file on a site that may host none of their audio. Your tracks sit on streaming services and video platforms whose crawler rules you don't control. The Munich court found Suno's training data was obtained by stream-ripping that got around YouTube's technical protections, per DLA Piper. The Code's signatories separately commit not to circumvent such protections. The rule exists, but the evidence that it was broken came from a lawsuit, not from a summary.

So the realistic aim for most working musicians is not to police AI companies alone. It is to make sure someone with standing and resources is doing it for them, and to keep the paper trail (release dates, where the music was published, what reservation was in place) that a later claim would need.

What we think

Transparency rules only help if someone reads the summaries. Individual artists won't read hundreds of documents, and they shouldn't have to. Collecting societies should. GEMA used recognisable reproductions in outputs as evidence of memorisation and persuaded the same Munich chamber twice, first against OpenAI over lyrics in November 2025 and now against Suno. The summaries add a new evidence source. Each society in Europe should commit to checking them on members' behalf and lodging complaints where they fall short.

Our second point is less comfortable. The Article 53 duties attach to providers of general-purpose AI models. In our reading, it is an open question whether a music-only generator counts as "general-purpose", and a provider has every incentive to argue it doesn't. If it doesn't count, the summary and the fine may never reach the tools that compete most directly with musicians. In that world, copyright litigation like the GEMA cases, not the AI Act, does the real work. The Act's best use may be as a pressure tool: a public record that courts and regulators can hold providers to. Until the AI Office brings its first case, treat that as a hope, not a guarantee.