A Council document dated November 18 shows fourteen European Union countries asking the European Commission to propose legislation to strengthen copyright in the age of AI. The note is on the agenda for the culture session of the November 19 and 20 Education, Youth, Culture and Sport Council meeting. It calls for rights reservations to be easier to enforce, more transparency about training data and adequate remuneration for creators. In the meantime it asks the Commission to enforce existing duties on providers of general-purpose AI models and the AI Office to examine the public training summary template. The published note is an intervention by the listed countries. Introducing new rules would require a subsequent legislative process. Revised Council note 15839/1/26 · Culture Council agenda
The proposal brings a practical question into focus. A creator can reasonably ask what happened to a work between the place it was posted and the model she suspects was trained on it. A model provider needs to know which material it can use, whose conditions apply and how to prove that a data supplier respected those conditions. A more detailed summary can help both. Whether a right is meaningful also depends on action after a publication, such as who receives a dispute, what information can be inspected and which party can change what happens next. Those are the steps on which a new law or better enforcement would be judged.
The revision changes who is making the request#
The first public registry version listed eleven countries. The revised note names Austria, Bulgaria, Croatia, Cyprus, Denmark, Estonia, Germany, Hungary, the Netherlands, Poland, Portugal, Slovakia, Slovenia and Spain. Bulgaria, Hungary and the Netherlands are among the additions to the earlier list. Early coverage described Germany and ten other states. That description matched the earlier document. The revised list is evidence that the call attracted more governments before the culture session. It is not a vote count for a law. A group of governments can agree on asking for action while leaving the mechanism to later discussion. Austrian Parliament registry and revision record · Euronews coverage of the earlier note
The signatories describe two objectives. They want to support the continued production of original human work and creators’ livelihoods, and also make lawful access to protected material possible for developers. They ask for comparable terms for European and foreign providers regardless of where a model was developed or trained. That is a policy position in the note. It leaves open how a future measure would connect the use of material, the provider and the market. As a practical matter, training and distribution can happen in several countries. A proposal would need methods for identifying the party responsible, getting evidence and allowing it to contest a claim. Revised Council note
There is more than one institutional track. A group of governments discussed the copyright framework at the Council in May. In March the European Parliament adopted a resolution on copyright and generative AI. The Commission’s separate consultation on a copyright initiative ran through November 10 after an extension. This note follows those steps, supports the Commission’s initiative and adds a direct call for a legislative proposal. The Parliament’s resolution and the governments’ note indicate what their authors want. A consultation collects material that can shape a proposal. The steps can point in the same direction while leaving specific requirements unsettled. May Council note · Commission consultation and extended deadline
The present rules answer different questions#
The EU’s Copyright in the Digital Single Market Directive addresses text and data mining. Article 4 provides for an exception or limitation for certain reproductions and extractions of lawfully accessible works used for mining. The provision makes the exception conditional on the relevant rights having not been expressly reserved in an appropriate manner. For content available publicly online, it refers to machine-readable means as an example. Article 3 separately addresses mining for scientific research by qualifying organizations and cultural heritage institutions with lawful access. The details of a particular case depend on the activity, applicable law and the facts, so the scope of a future change cannot be inferred simply from the two article numbers. EU Copyright Directive, Articles 3 and 4
The AI Act has a different job. Article 53 requires general-purpose AI model providers to put a policy in place to comply with EU copyright law, including identifying and complying with rights reserved under Article 4(3) of the copyright directive. It also requires a public summary about the material used for training, following the AI Office template. Those duties concern a provider’s practices and public information. Deciding whether an individual work was copied, covered by a licence or within an exception requires facts about that use. A provider can improve its public explanation and still face questions about specific inputs. The signatories’ request for enforcement of existing obligations is therefore separate from their call for new legislation. AI Act Article 53
The Commission describes its general-purpose AI Code of Practice as a voluntary way to demonstrate compliance with the AI Act. The underlying obligations rest in the legislation. The code’s copyright chapter sets out practical measures for providers. A commitment in a code and a finding about a work serve different purposes. To examine how a provider acts, a regulator might need to see its actual collection, processing and complaint records. A creator trying to resolve a particular problem needs to find the party who can look up its records. The public rule, voluntary compliance route and procedure for a dispute each need to fit together. Commission overview of the Code of Practice
Several issues are easy to combine under the word training. A company can collect a page and temporarily store a copy. It can license an archive, buy a dataset or update a model built by a different developer. A later application can retrieve a document at the time of a request. A model can produce an output that resembles part of a work. The legal and factual questions can differ at each stage. The note asks for stronger protection in relation to AI and singles out training transparency. Its immediate worth is to ask where rights can be observed along the path. The eventual measure will be clearer if it indicates which stage each duty covers.
A rights reservation has to reach the next person#
Imagine a small magazine that puts a rights reservation on new work published at its website. Its old columns are also available through an archive partner and a set of past issues. A provider assembling a dataset may receive a copy from the partner rather than from the main website. Even if a signal is technically readable at one location, that tells us little about whether it is recognized when a copy changes hands. To make a reservation useful, the parties need a way to associate it with the work, identify a version and show when they saw the instruction. The note asks to improve enforceability. It does not lay down a specific format or repository for those records. Council note on rights reservations
Handling older and newer copies also requires dates. A site instruction can change after a supplier collected a page. A magazine can modify terms when a columnist changes representation. A provider can receive a pool of files collected over years. A dispute about use may depend on what conditions were communicated and when, and on the law that applies. A record of the collector, collection date and source can allow both sides to ask a precise question. A durable record also helps a provider explain why it excluded or licensed a batch. A system that records only the final dataset name can lose the information that matters when someone raises an objection.
A provider that buys from a data supplier faces a related responsibility. A contract can describe what the supplier collected and what checks it performed, including the dates and the types of reservations it was able to identify. It can also describe what happens when a rights holder raises a dispute after delivery. If a newer dataset replaces the old one, the model developer can record which version went into which training run. None of this requires a creator to learn the exact technical structure of an internal dataset. It requires the companies with that knowledge to preserve enough of it for an authorized review, and to say which parts they can actually vouch for.
There is a choice about who has to do the work of making this system reliable. A large publisher can maintain metadata and legal relationships. An illustrator may have a small site, copies on other platforms and no technical staff. A model developer may deal with a chain of suppliers and have only indirect information about some earlier collections. A framework can reduce these burdens by setting recognizable ways to express rights and clear responsibilities for keeping records. It can also increase burdens if a reservation works only for rights holders who can afford to monitor many platforms. A fair test is whether a small creator and a small developer can use the same rules without paying for bespoke verification on every file.
The Parliament’s March resolution proposed ways to support transparency, remuneration, licensing and reservations, including consideration of an opt-out list. Those are positions with different implementation choices. A common list could make a reservation easier to locate. It would also need a way to handle identity, changing representation and competing claims. Signals on the original material can be nearer the source, while copies and legacy datasets can lose the signal. Standardizing one element can make the workflow more usable. The result still depends on whether a creator can have an incorrect entry fixed and whether a provider has a dependable way to identify the relevant work. European Parliament March resolution
A public summary needs to lead somewhere#
The AI Office’s public summary template is meant to provide a common baseline for what providers describe about training. A useful summary can show the main types of material, important sources and the approach to collection or processing. It can help the public identify the policy a company followed and compare summaries from two providers. A person trying to find out whether one essay appeared in a particular data collection can need different information. A broad category can contain many collections over time. A public baseline and a way to follow up a specific claim can serve complementary roles. The note asks the office to reconsider whether the current template offers enough transparency. Training content template and explanatory notice
A summary can also fail to distinguish which version of a model it describes. A provider may change some input sources between versions, add fine-tuning, or distribute a model that relies on a predecessor. A reader needs to know whether the summary concerns the original corpus, later additions or both. A revision date and a clear model identifier can help match the disclosure to the activity at issue. More detail is most valuable when it makes the evidence easier to use. A longer list of categories without dates, model identifiers or a contact route can still leave a creator unable to begin. This is the practical question underneath the request to examine the template.
Full publication of underlying records raises its own issues. A dataset can contain confidential sources, information about people and business terms. A protected process can permit a regulator or other authorized reviewer to examine more detail than the public statement. A usable system should mark which facts are disclosed, which can be checked privately and how a person learns the result of that check. That does not settle every conflict about who may see a record. It makes the trade-off concrete. An owner might need enough evidence to dispute a use, while an investigator could rely on a file history without posting the contents of the file online.
The redress process will matter regardless of the exact template. Suppose a creator can identify a particular edition of an article and the time at which it was available with stated conditions. There should be an office that can acknowledge the question, see the available record and give a usable explanation. A reply might identify the dataset or explain the limits on searching it. It might identify a licence, confirm a correction or explain why the company disagrees. A published time limit and an avenue for further review make these outcomes easier to evaluate. A system that collects complaints without describing their disposition provides a thinner form of accountability.
The route also needs to account for the evidence a creator is likely to have. Someone may start with an output that looks like a passage in a published work. The resemblance gives a question to investigate. A matching phrase can arise in more than one way, and a work used in training may never appear in a recognizable output. A rights holder will often need a record of how material entered a collection rather than relying on the model’s answer as a complete record. A developer will need the opportunity to explain whether the result came from base training, later data or content fetched during a request. These explanations are easier to substantiate when the relevant model and dataset versions can be identified.
Remuneration is a separate negotiation and design question#
The signatories ask for adequate remuneration across creators. They do not specify a tax, a payment schedule or which legal claim would give rise to payment. A voluntary licence can be arranged individually or through an organization, with terms about materials, duration and purpose. Collective arrangements can reduce transaction costs, especially where a developer wants access to a large collection. They also raise questions about who can represent each artist and how payment reaches the people whose works are covered. A procedure for use where a rights holder cannot be located would raise another set of questions. The intergovernmental note makes remuneration a priority without deciding these designs. Council note and priorities
One useful comparison is between access to an archive and access to a specific new work. The archive owner may be able to license a structured collection with contracts covering contributors. A freelancer may keep rights in individual pieces or have given particular rights to several organizations. A model provider may need to navigate those permissions before opening a dataset or may acquire use under a narrower purpose. Sharing the licence terms that apply to the relevant activity can give both parties a basis for agreement. If a proposed framework only measures what a provider pays to an aggregator, creators will still want to know how their share is calculated and whether they can challenge that calculation.
Scope and market structure also matter for competition. A design that only major platforms can comply with may encourage concentration. A design that offers cheap nominal compliance but makes it difficult for creators to be heard may preserve the costs the proposal seeks to address. Smaller providers and creators both stand to gain from predictable terms, searchable records and straightforward contact points. Their interests can diverge where those terms lead to a payment or limit on use. The European Broadcasting Union’s response to the earlier call supports targeted changes while cautioning against reopening the copyright directive itself. That is one example of a stakeholder favoring enforcement changes and keeping a legislative boundary. European Broadcasting Union consultation response
It is worth examining the distribution and timing of a payment in addition to its headline amount. A measure could yield a high total while paying only a small group with extensive catalogs. It could also produce a broad, low-cost route for creators and developers, but reach few retrievable works. Data on which rights were cleared, how much representation cost and how quickly a claim was settled would show whether a scheme works for people with little administrative capacity. The debate about a fair price cannot be separated from the evidence about who received it and how the use was documented.
What would demonstrate a change#
The next institutional milestone would be a more concrete Commission response. A proposal can set out which provider or collector a duty applies to, what conduct it covers, which documents people can request and how compliance can be tested. Work on enforcement of existing obligations can move on its own timetable, as can an evaluation of the template. The Commission’s enforcement guidance describes several authorities and powers involved in the AI Act. A reader can track the authority exercising a power, the action it takes and the period to which it applies. Presence on the agenda tells us the issue was put forward for discussion. It does not establish the result of that process. Commission guidance on AI Act enforcement
The system could be tested using a work that appears at more than one location. Establish its rights terms and dates, check how recognized reservations reach an original collector and an authorized intermediary, and see what evidence remains after the work is omitted or licensed. Separately ask whether a person could obtain a useful answer about a specific use through the available process. Such a test does not decide a legal dispute. It shows whether the process can retain the facts necessary to address one. It also makes it easier to evaluate changes to the summary template, because the test can identify which question the current version leaves unanswered.
Requiring practical evidence would also help distinguish a stronger policy from a longer statement. Enforcement bodies could report whether provider summaries were filed and whether they became easier to use. They could publish aggregate outcomes of inquiries while protecting sensitive case records. Legitimate use would have an identifiable path, such as an applicable exception or an agreement. A contested use could be investigated with dates, responsibilities and a trail linking the relevant materials. These outcomes need definitions and consistent reporting if they are to be compared. A count of pages in each disclosure would be only a small part of that picture.
The revised note demonstrates that a wider group of governments wants movement on the problem. Its value will depend on the decisions that follow about rights signals, evidence and remedies. A creator should be able to determine where to express an objection and obtain a meaningful response if she thinks it was ignored. A provider should be able to learn the conditions of lawful access and document how they were applied. Meeting both needs is what would turn the request into an enforceable working arrangement. The agenda puts the issue in view. The next evidence is whether the people holding the right and the people using the material can actually use the system.