EU consultation on copyright and AI: what it proposes

The European Commission opened a targeted consultation on technology and copyright on 29 September 2026. The deadline is 3 November. It is neither a legal reform nor a draft regulation: it serves to gather evidence before deciding whether the European framework requires additional measures. Its result will therefore depend on who responds, with what data, and from what position.

The consultation covers four distinct issues: the use of protected content in artificial intelligence systems; online piracy of live events; the application of the right to a single equitable remuneration for performers and music producers; and copyright in scientific research. For a visual artist, the first block is the most immediate, yet all four share a single question: whether current rules allow for the authorisation, remuneration, and pursuit of digital uses that no longer resemble those that existed when the rules were adopted.

The consultation is addressed to holders of copyright and related rights, providers of generative AI and other actors within their value chain, intermediaries, collective management organisations, organisers of live events, research organisations, national authorities, consumer organisations and non-governmental organisations. An artist or author who retains rights over their work may respond as a rightholder.

The Commission presents the European framework as 'robust' and asks whether it should be strengthened. That starting point is significant. It is insufficient to merely assert that AI harms creation. Anyone advocating for change must describe the specific failure, explain why current tools are inadequate, and propose a measure that can be implemented without shifting the entire burden of monitoring and proof onto each individual rightholder.

A review broader than artificial intelligence

The process began in May 2026 with a call for evidence on the review of Directive (EU) 2019/790 and a possible legislative initiative limited to specific problems. The Commission published the call on 18 May and kept the deadline open until 25 June. The September consultation is the next step and is also supported by previous studies and evaluations.

Artificial intelligence occupies the most visible space, yet the Commission is simultaneously reviewing problems involving different actors and markets. Live broadcast piracy raises the question of whether enforcement arrives only after the economic value of the event has already been exhausted. The remuneration of performers and music producers affects the distribution of revenue in the exploitation of phonograms. The scientific block necessitates a review of the balance between access to publications, research exceptions, and knowledge dissemination models.

There is no solution yet assumed by the Commission. The consultation is gathering evidence to decide whether it is advisable to amend existing law, improve its application, facilitate licensing, or combine several measures. Confusing questions with closed proposals would inflate the scope of the process.

The European rule and its practical problem

Directive (EU) 2019/790 distinguishes between two exceptions for text and data mining. Article 3 covers specific uses for the purposes of scientific research by research organisations and cultural heritage institutions. Article 4 establishes a broader exception for reproductions and extractions made through mining, provided that access to the content is lawful and the rightholder has not expressly reserved their rights. When content is available online, the reservation must be made in an appropriate manner, for example by machine-readable means.

On paper, this solution retains a right to opt out. In practice, it compels the rightholder to anticipate technical uses, publish a recognisable signal, and subsequently demonstrate that this reservation existed and should have been respected. Fundamental questions persist: which protocols are sufficient, how is the reservation expressed for images distributed by third parties, what happens to copies obtained before its publication, how is a work identified within massive datasets, and what information allows one to verify if it was used.

The AI Act adds obligations for providers of general-purpose AI models placed on the Union market. They must establish a copyright compliance policy, including the identification and observance of reservations expressed in accordance with the Directive, and publish a sufficiently detailed summary of the content used to train the model.

That summary improves transparency, but it is not a work-by-work inventory. The European template requests information on datasets, content modalities, crawling practices, collection periods, and a list of the main domains used. It may assist in locating sources and practices, but it does not guarantee that an individual will find their photograph, book, or a specific recording there. A gap remains between the general summary and individual proof.

The decisive issue is not only whether the development of a model technically contains a reproduction. It also matters whether access was lawful, which exception is invoked, whether a valid reservation existed, where the use occurred, who must provide the information, and what remedy is available. The consultation may influence precisely that chain of authorisation, transparency, licensing, and enforcement.

What the European institutions already demand

The European Parliament approved a non-legislative resolution on copyright and generative AI on 10 March 2026. It called for transparency regarding the use of protected content, effective respect for reservations, a stronger negotiating position for rightholders, and measures to facilitate voluntary licences. It also demanded that fair and proportionate remuneration be studied, including for past uses made when a licensing market did not yet exist, and considered sectoral collective agreements as a possible path.

The resolution does not amend the law nor does it bind the Commission. It does, however, highlight the issues that will likely structure any future negotiation:

What information developers must provide.
How a rightholder can express and enforce a reservation.
Whether authorisation should be negotiated individually or collectively.
How to economically value the repertoires used.
What response to provide for uses prior to the emergence of licences.
Who bears the costs of identifying a work and filing a claim.

The Commission has responded that it will examine the effectiveness of the text and data mining exceptions, licensing practices, remuneration, and enforcement. It is also studying technical solutions for machine-readable reservations and the possibility of digital fingerprint-based registries to complement signals associated with a website or file. None of these lines yet equates to a new obligation.

The courts show why clarity is lacking

European and British litigation do not offer a single answer because they adjudicate different facts, rights, and jurisdictions.

The Regional Court of Munich I upheld a significant part of GEMA's claim against OpenAI in November 2025. The court understood that the memorisation of protected lyrics in the models examined and their reproduction in responses infringed exploitation rights, without the text and data mining exceptions covering those acts. OpenAI maintained that the model did not store copies of specific works and that the responses depended on user instructions. The ruling does not allow its result to be generalised to any model, work, or output, and its finality is not confirmed in the sources consulted.

A few days earlier, the English High Court resolved the litigation between Getty Images and Stability AI. Shortly before closing submissions, Getty abandoned its claim regarding the training and development of the model, as there was no evidence that it had been carried out in the United Kingdom, as well as the claim relating to generated outputs, the prompts for which Stability had blocked. The court dismissed the secondary copyright infringement claim: a model that neither stores nor reproduces protected works is not an 'infringing copy' under sections 22 and 23 of the UK Act. Getty achieved partial success regarding trade mark infringement due to watermarks generated in certain versions of the model, with conclusions that the court itself describes as historic and of extremely limited scope. The judgment applies United Kingdom law and is not automatically applicable within the European Union.

These cases do not contradict each other in a simple manner. One analyses the reproduction of identifiable lyrics in responses and the potential for memorisation; the other was conditioned by the territorial evidence of the training and examined whether the model is an infringing copy. Together, they demonstrate what information is habitually missing: the provenance of the data, the location of the acts, the operation of the model, and the correspondence between input works and results.

The litigation between Robert Kneschke and LAION adds another issue: the scope of the scientific research exception and the efficacy of rights reservations when datasets are constructed. While the courts delineate these boundaries, European policy is attempting to determine whether it is sufficient to await a case-by-case interpretation.

In case C-250/25, currently pending before the Court of Justice of the European Union, the Commission itself notes that it shall be for the Court to interpret the exception under Article 4 and the scope of reservations in accordance with Article 4.3.

I analysed this fragmentation in greater detail in ‘Three thousand dollars for my book’.

What a visual artist can contribute

Copyright protects specific works and their original expression, not a style considered in the abstract. A result generated ‘in the manner of’ an artist may prejudice their market, induce false attributions, or substitute commissions without necessarily reproducing protected elements of an identifiable work. It is not advisable to obscure this harm, but rather to describe it with precision.

A useful response can separate several issues:

Use of works for training. Which works were available, under what conditions, whether a reservation existed, and what information would allow the use to be verified.
Lack of bargaining power. Whether an individual can offer a licence or whether only collective and sectoral mechanisms could function.
Economic substitution. Which commissions, licences, or audiences now compete with synthetic results and how that substitution can be measured.
Imitation and confusion. Whether the artist's name appears in prompts, commercial descriptions, or results presented as linked to their work.
Cost of claiming. Time, advice, technical evidence, and the difficulty of identifying the provider or the party responsible for the use.
Information asymmetry. The developer knows the data, tools, and records; the rightholder only observes a result and must reconstruct its origin.

Not all these issues are resolved by the same rule. The reproduction of a specific work pertains to copyright. False attribution may activate other remedies. The imitation of personal traits may affect image, voice, identity, or the rights of performers. The loss of commissions without demonstrable copying poses an economic and cultural policy problem that copyright may not be able to absorb on its own.

Precision strengthens the response. It is more useful to provide two documented cases, with dates, screenshots, and costs, than to state in a general sense that a model ‘has stolen a style’.

Reserving rights is not enough if no one can verify them

An individual can express their opposition via robots.txt, metadata, terms of use, and other machine-readable signals. I explained the options and their limits in ‘How to reserve your rights against data mining’.

The consultation allows problems that this guide cannot resolve to be conveyed to the Commission:

A single work circulates on the author's website, social media, catalogues, the press, museums, and repositories. A reservation published at one source does not necessarily accompany all copies.
Protocols are not uniform and crawlers do not interpret them in the same way.
A subsequent reservation does not clarify what happens to data already collected.
The rightholder may be unaware of which bot performed the download, for which provider, and at what time.
The observance of a signal requires records that are normally kept by the other party.
Technical measures alone do not solve the granting of licences or the calculation of remuneration.

Hence, the issue is not choosing between reservation and licence. An effective reservation can be the condition that allows for negotiation, but it requires identification, traceability, and a proportionate path for claiming.

Individual licences, collective licences, and remuneration

Individual licences function when the parties can meet, describe the uses, and negotiate a price. That situation is common in certain commissions or catalogues and unrealistic for millions of works obtained through mass crawling.

Collective licences can reduce transaction costs, concentrate repertoires, and facilitate remuneration. They also raise questions: which entity represents each sector, how a work is incorporated or excluded, what uses the licence covers, how data is audited, and how revenue is distributed. A solution designed for music or the press cannot be transferred without modification to photography, illustration, or the visual arts.

Remuneration for past uses adds another difficulty. If an operational market for licences did not exist, one must decide what fact generates the obligation, how the repertoire used is identified, what period it covers, who pays, and how it is distributed. The Parliament's request is politically relevant, but it does not yet constitute a right to payment.

An individual response can contribute something that institutional positions tend to omit: what amounts a person obtains today for licensing a work, what uses they normally authorise, and how much it would cost to exercise their rights against an AI provider. Without market references, the expression ‘fair remuneration’ risks remaining empty.

How to prepare a useful response

Before opening the form, gather a minimum base:

Identification of the person or entity responding.
Discipline and habitual form of exploitation of the works.
Two or three verifiable examples.
Date and place of publication of the affected works.
Reservations or terms of use in force at that time.
Screenshots and files preserved unedited.
Communications with platforms, providers, or clients.
Time and costs dedicated to investigation or lodging a claim.
Specific economic or professional consequence.
Proposed measure and the entity that should implement it.

A brief structure may suffice:

Who you are and how your work circulates. Discipline, professional sphere, publication channels, and licensing forms.
What occurred. Verified fact, date, service involved, and reference works.
What information was missing. Training data, crawler, collection date, licence, reservation record, or contact mechanism.
What damage or cost was produced. Loss of licence, replacement of a commission, confusion, time spent claiming, or inability to negotiate.
What measure you propose. It must be applicable, proportionate, and clearly assign responsibility.

Among the measures that may be advocated are a common standard for reservations, additional information upon reasoned request, record-keeping, collective licensing mechanisms, proportionate complaint procedures, or public intervention that reduces evidentiary asymmetry. It is not necessary to support all of them. It is advisable to choose the one that addresses the documented problem.

Do not include personal data of third parties or assertions that you cannot substantiate. If a suspicion is not confirmed, present it as such and explain what information would allow it to be verified.

It is not necessary to complete the entire questionnaire. You may select only the relevant section, omit individual questions, and attach additional documentation. Documents containing confidential information must be expressly identified as such.

What may happen next

When the consultation closes, the Commission will analyse the responses alongside the call for evidence, external studies, and the review of the Directive. It may conclude that it is sufficient to improve the application of the rules, propose non-legislative measures, facilitate standards or licences, or present a reform. The sources consulted do not show a firm timetable for that decision.

If it proposes legislation, the text must be negotiated with the European Parliament and the Council. The consultation does not anticipate the result nor does it suspend the current framework. Reservations, licences, claims, and litigation continue to be governed by the current rules until otherwise approved.

The upcoming deadline does not make any response decisive. It does, however, open a rare opportunity for the individual costs of identifying, negotiating, and claiming to appear in the European file before the measures are drafted. Companies and organisations will respond with specialised teams. The experience of a creator can show what aggregated data does not reveal, provided it arrives as evidence and not merely as a slogan.

About the open conversation

This text distinguishes between verified facts, interpretations, and proposals. If you have responded to the consultation, work in licensing or collective management, or can clarify any of the data, you may write to me.

The author's relationship to the cases

There is no economic, professional, or collection relationship with the institutions, companies, and entities cited.

Copyright and AI: what the European Commission is asking now