Opt-Outs, Presumptions and Payment: What the EU's New Consultation on Copyright and AI Puts on the Table

What the Commission launched, and what it is not

On 29 September 2026 the European Commission opened a targeted consultation on the challenges and the way forward for copyright in a period of technological and market change. The questionnaire, published on the Commission's EUSurvey platform under the title "Targeted consultation on a potential initiative for a better copyright environment for European creativity and innovation", covers four subjects: the use of copyright-protected content in artificial intelligence, online piracy of time-sensitive content such as live events, the single equitable remuneration right for music performers and producers as it applies to recordings by nationals of non-EU countries, and copyright in scientific research. It closes on 3 November 2026 at the end of the day.

The consultation's place in the process is often blurred. A call for evidence of 18 May 2026 covered both the review of the 2019 Directive on copyright in the Digital Single Market (the DSM Directive), which Article 30 requires no sooner than 7 June 2026, and "a potential targeted legislative initiative on copyright". The new questionnaire builds on that call and says its outcome "will feed into the preparation of a targeted initiative". On text and data mining, it says the challenges of the exception and its opt-out "are being assessed in the context of the ongoing DSM Directive review process". So this is neither the review nor a legislative proposal; it is the stage at which the Commission tests specific options before deciding whether, and how, to legislate.

The Commission also says what the options are not. They reflect approaches raised in the public debate, in the European Parliament's report on copyright and generative AI (2025/2058(INI)) and in discussions with Member States, and their inclusion does not mean the Commission favours any of them or that the list is exhaustive. What follows is a description of a menu, not a forecast of the meal.

Still, the menu is unusually specific. Stakeholders are asked to assess an EU-level registry for rights reservations, a duty on crawlers and AI agents to disclose who they are, a rebuttable presumption of use where an AI provider has not met its AI Act transparency duties, confidential disclosure as an alternative to the published list of crawled domains, and a remuneration right for certain generative-AI uses. A threshold question asks whether the right approach is the status quo, soft law, non-legislative support or "new binding measures under EU copyright law". A separate section, on AI-generated imitations of performers, asks whether EU rules on digital likeness are needed and whether copyright is the right vehicle at all; this article concentrates on training, inference and the licensing and enforcement options.

The legal baseline: Articles 3 and 4 of the DSM Directive

Every option in the consultation is built on two provisions of the DSM Directive.

Article 3 requires Member States to provide an exception for reproductions and extractions made by research organisations and cultural heritage institutions in order to carry out text and data mining, for the purposes of scientific research, on works to which they have lawful access. Rights holders cannot opt out of it, and under Article 7(1) any contractual provision contrary to it is unenforceable.

Article 4 is the general exception. It covers reproductions and extractions of lawfully accessible works for the purposes of text and data mining by anyone, including commercial actors, and allows the copies to be kept for as long as necessary for that purpose. But under Article 4(3) it applies only on condition that the use "has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online". Article 7(1) does not protect Article 4 against contractual override. In other words, for commercial training the default is permission, and the rights holder's tool is the reservation, the so-called opt-out.

Three gaps explain the consultation. The Directive does not say which "machine-readable means" count. It is silent on remuneration: without a reservation the use is free, with one the user needs a licence, and nothing makes that licence market work. And an opt-out is only as strong as the ability to verify it, for which the Directive provides no mechanism. The AI Act was meant to fill that last gap.

The AI Act layer: Article 53 and the training-content template

Article 53(1) of the AI Act imposes two copyright-related duties on every provider of a general-purpose AI model. Under point (c), the provider must put in place a policy to comply with Union law on copyright and related rights, "and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3)" of the DSM Directive. Under point (d), it must draw up and publish "a sufficiently detailed summary about the content used for training" of the model, according to a template provided by the AI Office. The open-source carve-out in Article 53(2) covers only points (a) and (b); points (c) and (d) apply also to models released under a free and open-source licence.

The AI Act applies to providers placing general-purpose AI models on the EU market whether or not they are established in the Union (Article 2(1)(a)), and providers established in third countries must appoint an authorised representative in the Union before doing so (Article 54(1)), unless the model is released under a free and open-source licence and does not present systemic risk (Article 54(6)). Recital 106 adds the point most relevant to non-EU developers: any provider placing a model on the Union market should comply with the copyright-policy obligation "regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place". The consultation repeats this in its own words: the obligations apply regardless of whether models were trained inside or outside the EU. Fines for providers of general-purpose AI models can reach 3% of annual worldwide turnover or EUR 15 million, whichever is higher (Article 101(1)).

The Commission published the explanatory notice and template for the public summary of training content on 24 July 2025. For scraped web data, section 2.3 asks for the crawlers used, their purpose and behaviour, the collection period, a description of the content and websites scraped, and a list of the most relevant domain names: the top 10% of all domains by size of content scraped, or, for SMEs, the top 5% or 1,000 domains, whichever is lower. The summary must be "generally comprehensive" but not "technically detailed", and need not identify specific works; for unlisted domains, providers are invited to answer rights holders' requests voluntarily. Summaries have been due since 2 August 2025 for new models and by 2 August 2027 for older ones, and the AI Office's enforcement powers started on 2 August 2026.

The consultation adds that the Commission "has already deployed its investigatory powers in this respect", without saying against whom.

The enforcement menu: opt-out standards, a registry, crawler identity and audits

Question 6 of the AI section is the core of the enforcement part. Respondents are asked to rank, in order of appropriateness, the following mechanisms (paraphrased closely from the questionnaire):

operationalising the existing opt-out under the TDM exception through a common vocabulary, EU standards for opt-out protocols and an EU-level registry for opt-outs;
clarifying how EU copyright rules apply, including their territorial application, to the use of protected content for AI training and other AI uses;
an obligation on web crawlers and AI agents used to crawl content online to disclose their identity and purpose;
an obligation on AI providers, dataset providers and data collectors to keep records about their use of protected content (for example clearance, licensing and verification of opt-outs) for a defined period;
an obligation of independent audit or certification on the use of protected content;
an obligation on AI providers to give information under a non-disclosure agreement, on a substantiated request from collective management organisations or rights holders, as an alternative to publishing the top 10% of crawled domain names under the AI Act template;
the possibility for courts or administrative authorities to order disclosure, under certain conditions, in legal proceedings, mediation or arbitration;
a rebuttable presumption of use of protected content by AI providers, subject to specific conditions, for example where providers do not fulfil their AI Act transparency obligations;
no need for additional mechanisms.

Question 7 then asks respondents to rate, on a five-point scale from strongly negative to strongly positive, how three of these options would affect their business or activity: operationalising the opt-out, record-keeping, and the presumption of use. Both questions are followed by free-text boxes, and the questionnaire allows supporting documents with data on the impact of the options to be uploaded.

The registry option did not come from nowhere. In its follow-up to the Parliament's resolution of 10 March 2026, the Commission said it was finalising a study on a registry of rights reservations using content-based digital fingerprinting as a complementary means of reserving rights. The same document records that signatories to the General-Purpose AI Code of Practice have committed to crawlers that respect robots.txt and other appropriate machine-readable protocols, and that the AI Office supervises Article 53 compliance for all providers, signatories or not. A fingerprint registry would change the logic of the opt-out: the reservation would travel with the work rather than with the page on which it happens to appear, answering the commonest complaint of photographers and illustrators whose works are copied to sites they do not control. It would also raise hard questions of cost, accuracy and governance, and the Commission itself stresses that such a registry must not circumvent the Union's obligations under the Berne Convention, which prohibits making the exercise of copyright subject to formalities.

The crawler-identity option addresses a different gap: robots.txt instructions only work if the crawler identifies itself. A duty extended expressly to "AI agents used to crawl content" would make reservations addressable and covert collection unlawful in itself. Record-keeping and audit duties would turn the Article 53(1)(c) copyright policy from a document into an evidential trail, and their wording reaches "dataset providers / data collectors" as well as model providers, which matters for the LAION case discussed below.

Presumption of use, court disclosure and the trade on the domain list

Two options in Question 6 would change litigation more than compliance.

The first is a rebuttable presumption of use. Rights holders everywhere face the same problem: they cannot prove a work was in a training set because the evidence sits with the defendant. A presumption triggered where the provider has not met its AI Act transparency duties would reverse that burden selectively. It would not make use unlawful, since coverage by Article 3 or 4 or by a licence would still have to be decided, but it would tie a copyright consequence to the regulatory failure and give providers a strong private-law incentive to publish complete summaries.

The second is court-ordered disclosure. The template's explanatory notice already points to Article 8 of the IP Enforcement Directive as an available remedy; the consultation asks whether a specific power is needed, usable in court and also in mediation or arbitration. For developers this may matter more than any registry, because it would operate once a dispute exists and could reach well beyond a public summary.

The most interesting option is a trade: information under non-disclosure terms to collective management organisations or rights holders on a substantiated request, "as an alternative to the public disclosure of the top 10% of the top domain names of the websites crawled". Providers say the domain list exposes sensitive sourcing choices; rights holders say it is too coarse. A targeted, confidential right of inquiry could satisfy both, or neither, and would be a notable change of policy only fourteen months after the template was adopted.

Licensing and remuneration: three stages of the AI lifecycle

The licensing part of the questionnaire is built around a distinction that EU copyright law has not previously drawn in legislation. Question 8 asks which mechanisms would best support licensing and access to protected content separately for three stages: training of generative models (where rights holders have opted out of the TDM exception); inference-time use, including retrieval-augmented generation, grounding and other uses of content by a deployed system in responding to users; and AI-generated output. The options are:

supporting a marketplace for licensing protected content for AI uses, for example through technical solutions for automated licensing at scale;
model contractual clauses for licences covering AI uses;
an obligation on rights holders and AI providers to negotiate in good faith;
a structured mediation or arbitration mechanism run by an independent body that could, under certain conditions, require the parties to share information, negotiate and assist in determining remuneration;
promoting collective licensing, in particular extended collective licensing, where appropriate;
a fair compensation obligation for uses that fall under the Article 4 exception because the rights holder did not reserve;
a remuneration right for certain uses of protected content by generative AI providers or other relevant actors.

Two of these options would alter the Article 4 bargain itself. A fair-compensation obligation for unreserved uses would mean that silence no longer equals free use: the exception would remain but carry a payment, as private-copying exceptions do. A remuneration right could, depending on design, apply whether or not an opt-out was made, and would most likely be administered collectively. If the legislature followed the territorial logic of recital 106, either would reach providers offering models in the EU wherever they trained; the separate option to "clarify" territorial application is where that question will be fought.

The inference category is the sleeper issue. Retrieval-augmented generation copies content at query time, often from news sources and often in a way that substitutes for a visit to the source. The Commission's follow-up says it will evaluate whether current law can mitigate the "adverse substitutive effects" of techniques such as RAG. Separating inference from training suggests that a future instrument may treat them differently, and that an exception designed for analysis may not be the frame for content used live.

How thin are the published summaries?

The transparency options in the consultation assume that the current regime is not delivering. There is now some evidence on that point, with important caveats.

On 25 September 2026 the French specialist outlet ActuIA published an opinion piece (updated on 7 October) comparing 24 public training-content summaries from seven providers that were accessible on that date: eleven from Ant Group, four from ByteDance, three from DeepSeek, two each from OpenAI and MiniMax, and one each from Mistral AI and Z.AI. According to the review, 21 of the 24 summaries contain the section 2.3 field for scraped domain names, and none of them names a specific website: the answers describe categories of source, content types or top-level extensions such as .com. DeepSeek's three summaries omit the web-scraping section and its domain field altogether. In other words, the headline figure understates the point: on the review's account, none of the 24 summaries lists an actual domain in that field.

The caveats are real. The piece is an opinion article and says it is not a census; the sample is weighted heavily towards one group; it concerns one field of the template and acknowledges that the summaries name other sources, such as public datasets; and whether a category-level answer satisfies Article 53(1)(d) is for the AI Office to decide. We have not independently checked the 24 summaries. But the template is plain: section 2.3 asks for a list of the most relevant domain names, "top and second-level domain, e.g. 'example.com'". An answer of ".com" does not obviously meet that wording.

If the finding holds, it cuts both ways. It supports the presumption-of-use option, which would bite hardest where summaries are routinely non-specific. It also supports the confidential-disclosure alternative: if the public list is not being produced anyway, a targeted inquiry right may be the more realistic route to useful information. The AI Office's first assessments of summaries may shape the debate more than the consultation itself.

What the German courts have said

The consultation lists legal uncertainty about the TDM exception as the first candidate for the most pressing issue. German courts have produced most of the European case law so far.

Kneschke v LAION

The case concerns a dataset, not a model. LAION, a non-profit association, publishes datasets of links to public images with descriptions, usable to train image generators. In compiling one, it downloaded a photographer's image from a stock agency's website to check it against its description. The Hamburg Regional Court dismissed the photographer's claim on 27 September 2024 (310 O 227/23) under the scientific-research exception in section 60d of the German Copyright Act, which implements Article 3.

On 10 December 2025 the Hanseatic Higher Regional Court dismissed the appeal (5 U 104/24). According to the court's press release, it held, first, that LAION could rely on the general TDM exception in section 44b, a question the Regional Court had left open. The reservation published by the agency was attributable to the photographer, who used the agency to exploit his images, but the reservation present on the agency's website at the time of the download did not have the statutory form, machine-readability, required by section 44b(3), second sentence. Commentary on the full judgment reports that the court treated machine-readability as a technology-neutral concept but found it not established that a reservation in natural language could be read and interpreted by machines in 2021, when the download took place; how the question would be answered with today's technology was not decided. Commentators differ on how far that reasoning goes. Second, the court agreed that the use was also justified by the research exception: building the dataset was itself a methodical, verifiable process aimed at later knowledge, and the fact that commercial providers could use the dataset did not change that, because no private company had decisive influence over the research organisation. The court allowed a further appeal on points of law.

Press reports state that the Federal Court of Justice heard the appeal on 3 September 2026 and that its first civil senate indicated a need for clarification of EU law, raising the possibility of a reference to the Court of Justice; no decision date was announced. At the time of writing we are not aware of a decision.

The Hamburg judgment shows how much turns on what counts as "machine-readable", and when. A common vocabulary, protocol standards and a registry are all attempts to settle that by legislation rather than case by case.

GEMA v OpenAI

On 11 November 2025 the 42nd civil chamber of the Munich Regional Court I largely upheld claims by GEMA, the German music collecting society, against two OpenAI companies (42 O 14139/24). According to the court's press release, the lyrics of nine songs were reproducible from the GPT-4 and GPT-4o models (memorisation), and their fixation in the model parameters was a reproduction even though stored as probability values. The TDM exception could cover preparatory copies made in compiling training data, but storing the works themselves in the model is not text and data mining, and the court refused to apply the exception by analogy, noting that it provides no remuneration. The outputs also infringed the reproduction and making-available rights, and the defendants, not the users entering simple prompts, were responsible. A personality-rights claim was dismissed. The judgment is not final; OpenAI has appealed to the Higher Regional Court of Munich.

The decision shows that even where Article 4 covers collection, a court may treat what ends up in the model as a separate act outside the exception, and it places responsibility for outputs on the provider, the premise of the consultation's output-stage question.

GEMA v Suno, and the case everyone is waiting for

On 31 July 2026 the same Munich chamber ruled largely for GEMA against the music-generation service Suno (42 O 763/25). According to the court's press release, GEMA's announcement and published summaries of the judgment, the court assessed training carried out in the United States under United States law and found it not covered by fair use, so that a licence was required, and treated memorisation in models stored on servers in Germany as a reproduction not covered by section 44b. According to published summaries, it also rejected the argument that compliance with Article 53(1) of the AI Act made a licence unnecessary. That judgment too is at first instance and not final.

Above all of these sits Like Company v Google Ireland (C-250/25), a reference from a Hungarian court lodged on 3 April 2025. It asks, among other things, whether a chatbot's output that partly reproduces press content is a communication to the public, whether training a large language model is a reproduction, and, if so, whether Article 4 of the DSM Directive covers it. The Grand Chamber heard the case on 10 March 2026. The Commission itself notes, in its follow-up to the Parliament, that the Court "will rule on the interpretation of the exception in Article 4". An Advocate General's opinion had been expected in early September 2026; as of early October we could find neither that opinion nor the judgment published, and press reports in September referred to delays. A ruling that training falls inside Article 4 would strengthen the case for opt-out machinery and remuneration; a ruling that it falls outside would push the debate straight to licensing.

What changes for Turkish developers and rights holders

Turkish AI developers tend to assume that EU copyright rules concern them only if they train in the EU. That is already wrong under the AI Act, and the consultation suggests it may become wrong under copyright law too.

A Turkish company that places a general-purpose model on the EU market, including through an API offered to EU customers or an open-weight release downloadable in the EU, must have a copyright policy that respects Article 4(3) reservations and publish a training-content summary on the Commission's template; unless the free and open-source exemption in Article 54(6) applies, it must also appoint an authorised representative in the Union. For the copyright-policy obligation, recital 106 makes the place of training irrelevant. New record-keeping, audit, crawler or disclosure duties would most naturally attach to the same providers, with the same reach.

On liability, the EU applies the law of the country for which protection is claimed (Article 8(1) of the Rome II Regulation), and Turkish law does the same: Article 23(1) of the Law on Private International and Procedural Law (No. 5718) subjects intellectual property rights to the law of the country in which protection is sought. A rights holder suing in Germany over outputs generated or models stored there will invoke German law, and the Suno judgment, as reported, shows a German court prepared to assess training abroad under the foreign law.

Home law offers no comfort. The Law on Intellectual and Artistic Works (FSEK, No. 5846) has no text and data mining exception. Article 22 gives the author the exclusive right to reproduce a work in any form, in whole or in part, directly or indirectly, temporarily or permanently, and the Article 38 exception is limited to personal use without profit. A developer training in Türkiye is therefore, if anything, in a weaker statutory position than one relying on Article 4 in the EU. Developers serving EU customers should build compliance on the stricter of the two systems.

For Turkish rights holders the picture is different. Turkish authors and photographers are protected in the EU on national-treatment terms under the Berne Convention, and producers, performers and broadcasters under related-rights treaties such as TRIPS and the WPPT; they can reserve their rights under Article 4(3). Whether a future EU registry would be open to non-EU rights holders on equal terms, and in which languages and formats, is a question the questionnaire leaves open. In licensing to AI developers, Article 52 FSEK requires contracts on economic rights to be in writing and to specify each right separately; the consultation's split between training, inference and output maps naturally onto that rule, and a licence that does not name each use risks a narrow reading. The questions on collective and extended collective licensing are relevant to the collecting societies (meslek birlikleri) operating under Article 42 FSEK. And where performers' voices or faces are used, biometric data is a special category under Article 6 of the Personal Data Protection Law (KVKK, No. 6698).

What changes for Korean developers and rights holders

The Korean position is structurally different from the Turkish one. Korea has no specific text and data mining exception, but its Copyright Act contains a general fair-use clause in Article 35-5, with four factors modelled on United States law. In February 2026 the Ministry of Culture, Sports and Tourism and the Korea Copyright Commission published a guide on fair use in generative-AI training, and an English edition followed on 18 May 2026. As reported, the guide does not exclude commercial purposes or automated crawling from fair use as such and leaves the outcome to the four-factor balance. It is guidance, not legislation, and the question it addresses is under litigation: the case brought in January 2025 by the broadcasters KBS, MBC and SBS against Naver over the use of news content to train its HyperCLOVA models is pending before the Seoul Central District Court, and press reports suggest a first-instance judgment may come, at the earliest, around the end of 2026 or early 2027.

For Korean developers offering models in the EU, a fair-use analysis under Korean law does not answer the EU question. The AI Act duties apply as they do to Turkish providers, and on the reasoning of the Munich judgments a model trained in Korea in reliance on Article 35-5 may still be found to contain reproductions once stored and offered in an EU Member State. Korea's AI Basic Act, in force since 22 January 2026, has no equivalent of the EU training-content summary, so many Korean providers will be building that documentation for the first time.

Korean rights holders, from music and broadcasting to webtoons and games, have a strong export interest in the EU, and the registry and remuneration options would affect them directly. How such mechanisms would interact with Korean collective management and contract practice is not addressed in the questionnaire.

Objections and open questions

The first objection is competitiveness. The consultation itself lists AI providers' need for high-quality data and the risk of competitive disadvantage among the problems. New obligations would apply to non-EU providers serving the EU too, which answers the level-playing-field concern in principle; in practice, records, audits and confidential disclosure weigh hardest on smaller providers, as the questionnaire's repeated references to SMEs and start-ups acknowledge.

The second is administrability. A fingerprint registry must be accurate, cheap to query and hard to abuse; a presumption needs carefully drafted triggers; a remuneration right needs a distribution mechanism that does not simply route money to the largest repertoires. None is impossible, all are hard, and the impact questions are the place to say how hard.

The third is sequencing. Like Company may decide whether training falls within Article 4 at all. Legislating on opt-out mechanics before that answer risks building on sand; waiting risks years more uncertainty. The fourth is formalities: the Berne Convention bars making the exercise of copyright subject to formalities, so an opt-out that works only if registered could be challenged, which may explain why the Commission describes a registry as a "complementary" means of reservation.

Who can respond, how and by when

The consultation is open to anyone. The Commission invites rights holders, generative AI providers and other value-chain actors, intermediaries, collective management organisations, sports and live-event organisers, research organisations, national authorities, consumer organisations and NGOs. The questionnaire's country list includes Türkiye and South Korea, and organisations above 1,000 employees are asked whether they are headquartered inside or outside the EU. A transparency-register number is requested only if applicable.

Contributions can be made in English, French or German.
Only the identification section is mandatory; respondents may answer only the sections relevant to them.
Most AI questions are rankings or impact scales, followed by free-text boxes for explanations.
A supporting document can be uploaded; confidential documents should be marked as such.
Respondents choose whether their contribution is published with their details or anonymously.
The deadline is 3 November 2026, end of day.

A parallel workstream on an AI strategy for the cultural and creative sectors, with a call for evidence published in July 2026, is described as setting an overarching vision rather than legal rules.

Practical steps and what to watch

For developers placing models on the EU market, from Türkiye, Korea or elsewhere, some steps make sense whatever the outcome:

Check that the training-content summary lists domain names in section 2.3 in the form the template describes; the AI Office can now examine exactly that.
Identify every crawler used by or for the organisation and make sure it identifies itself and respects robots.txt and other recognised machine-readable reservations.
Keep contemporaneous records of what was collected, when, from where and under which reservation status. If a record-keeping duty or presumption arrives, they will be decisive.
Separate training, retrieval and output uses in licences and internal policies, and assess memorisation and output risk, not only collection.

For rights holders, the equivalent steps are to put reservations into a recognisably machine-readable form, document when and how they were applied, and review AI licence templates with Article 52 FSEK or the relevant national rule in mind.

To watch: the Like Company judgment; the Federal Court of Justice in LAION; the Munich appeal in GEMA v OpenAI; the AI Office's first assessments of summaries; the DSM Directive review report; and the first-instance judgment in the Korean broadcasters' case against Naver.

Frequently asked questions

Is the Commission proposing new copyright legislation?

Not yet. The consultation tests options for a possible "targeted initiative" and states expressly that listing an option does not mean the Commission supports it. One threshold question asks whether the right approach is the status quo, soft law, non-legislative measures or new binding rules.

Does the consultation concern companies outside the EU?

Yes, in substance. The AI Act's copyright duties already apply to any provider placing a general-purpose model on the EU market, wherever it trained, and the consultation includes an option to clarify the territorial application of EU copyright rules to AI training. Non-EU organisations can respond.

What does the "21 of 24" figure mean?

It comes from an opinion piece published by ActuIA on 25 September 2026 that compared 24 training-content summaries from seven providers. In 21 summaries the domain field was present but named no specific site; the other three, from DeepSeek, omitted the section. It is a sample, not a census, and whether the summaries comply with the AI Act is for the AI Office to decide.

Does Turkish or Korean law allow AI training on protected works?

Turkish law (FSEK) has no text and data mining exception, and its reproduction right expressly covers temporary copies. Korean law has no specific TDM exception either, but has a general fair-use clause (Article 35-5 of the Copyright Act), on which the government has issued non-binding guidance; the question is being litigated. Neither answers whether use is lawful in the EU.

Burhan Doğuş Ayparlar's View

The most revealing part of the questionnaire is not any single option but its structure. For seven years the EU's answer to AI training has been a two-word rule: opt out. The consultation is the first official admission that the rule, as designed, does not work on its own. It cannot be exercised reliably without a standard, cannot be verified without transparency and cannot be monetised without a functioning licence market. Every option in the enforcement and licensing questions is an attempt to supply one of those missing parts.

In my view the most consequential option is the least discussed: the presumption of use tied to non-compliance with the AI Act. It is elegant because it requires no new regulator and no new database. It simply converts a regulatory failure into an evidential consequence in private litigation. If the early evidence on training-content summaries is borne out, a presumption of that kind would do more to improve disclosure than any amount of guidance. Its weakness is that it depends on the AI Office being able to say, with some authority, when a summary falls short; the presumption and supervision would have to be designed together.

The confidential-disclosure alternative deserves a cautious reception. A right of inquiry may produce more useful information than a list of the top 10% of domains, but it moves transparency into the private sphere, where it favours those able to make substantiated requests: large publishers and collecting societies rather than individual creators. If that trade is made, it should supplement the public list, not replace it.

For Turkish readers, the lesson is that copyright for AI is being regulated by the market in which a model is offered, not the country in which it is trained. Turkish developers need only EU customers to be exposed to European duties, while the absence of any TDM exception in FSEK leaves training at home on less certain footing than training in the EU. That argues for addressing the question deliberately, with an exception, a reservation mechanism and a transparency duty designed together, rather than waiting for courts to fill the gap under Article 22 FSEK. Korea, which has chosen fair-use guidance and awaits its first major judgment, faces a version of the same choice. The questions in this consultation are the right ones; how well they are answered will depend on the evidence the Commission receives, including on the position of developers and rights holders outside the EU whose models and works reach EU users.

This article is for information only and does not constitute legal advice. It is based on the Commission's announcement of 29 September 2026 (updated 1 October 2026), the questionnaire as published on EUSurvey and accessed on 8 October 2026, the AI Act and DSM Directive texts, the Commission's template and explanatory notice of 24 July 2025, the Commission's follow-up to the European Parliament resolution of 10 March 2026, the press releases of the Hanseatic Higher Regional Court (10 December 2025) and Munich Regional Court I (11 November 2025), the Munich Regional Court I press release and GEMA's announcement of 31 July 2026 on the Suno judgment and published summaries of that judgment, press reports of the Federal Court of Justice hearing of 3 September 2026, and ActuIA's review of 25 September 2026 (updated 7 October 2026). The description of the Suno judgment and parts of the Hamburg reasoning rely on announcements and published summaries rather than full judgment texts. The outcome of the LAION appeal, the status of the Advocate General's opinion and the timing of the judgment in Like Company and the compliance of individual training-content summaries were not known at the time of writing. Statements about Turkish, Korean and EU law are general in nature; specific cases require individual assessment. The analysis and assessments are the author's own.