No Distinctive AI-Specific IP Regime

There is no AI-specific intellectual property (IP) legislative regime in the European Union (EU). AI-generated outputs and assisted inventions, training data, and text and data mining (TDM) are all governed by the EU’s existing IP framework:

• Copyright and related rights: InfoSoc Directive (2001/29/EC), DSM Directive (2019/790/EU) and Software Directive (2009/24/EC);
• Sui generis database right:Database Directive (96/9/EC);
• Patents:European Patent Convention (EPC) and Unitary Patent framework;
• Trade marks:Trade Mark Regulation (2017/1001/EU) (EUTMR) and Trade Mark Directive (2015/2436/EU);
• Designs:European Union designs Regulation (2026/715/EU) and Design Directive (2024/2823/EU); and
• Trade secrets:Trade Secrets Directive (2016/943/EU).

In the absence of AI-specific IP legislation, specialist bodies such as the European Patent Office (EPO) have issued soft-law guidance on the application of existing IP rules to AI. For example, the EPO’s Guidelines for Examination contain a dedicated subsection on AI and machine learning and clarify the sufficiency-of-disclosure requirements for AI and machine-learning inventions.

The AI Act and Related Copyright Obligations

What makes the EU unusual is the fact that it has an AI-specific statute, the AI Act, which defines an “AI system” and a “general-purpose AI model” (GPAI model). These definitions are not, however, specific to IP. The AI Act also adds obligations relevant to IP, but it does not create new IP rights or an AI-specific infringement action. Under Article 53(1)(c) and (d) of the AI Act, providers of GPAI models and transparency, while a July 2025 European Parliament study called for clearer rules on training, outputs, opt-outs and licensing.

The European Parliament’s 10 March 2026 resolution subsequently called for a supplementary framework, including greater transparency and licensing. The European Commission has since launched a review of EU copyright rules, including the DSM Directive, focusing in part on generative AI, licensing and enforcement.

EU AI and IP law is shaped by several international IP treaties, regional patent instruments, free trade agreements and soft-law initiatives. The principal multilateral IP treaties are:

• the Berne Convention;
• the Paris Convention;
• the Rome Convention;
• the European Patent Convention;
• the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS);
• the World Intellectual Property Organization (WIPO) Copyright Treaty (WCT);
• the WIPO Performances and Phonograms Treaty (WPPT); and
• the Comprehensive and Progressive Agreement for Trans-Pacific Partnership (CPTPP).

EU trade agreements, including CETA and the EU–Japan EPA, also contain IP commitments.

WIPO initiatives on AI and IP, including its work on frontier technologies and training data, also inform policy development, but do not create binding EU AI-IP rules.

Foreign Rightsholders and Developers

Foreign rightsholders generally receive the same protection as EU rightsholders where the relevant international or EU rules apply. Foreign AI developers are likewise subject to EU law where the relevant territorial or market-access conditions are met. In particular, the AI Act applies to certain third-country providers and deployers where the specific territorial conditions in Article 2 are met, ie, irrespective of where the model was developed or trained. EU private international law then determines jurisdiction and applicable law in cross-border disputes.

EU IP legislation does not contain a distinct set of AI-specific definitions. Existing IP concepts relevant to AI, such as “text and data mining” under the DSM Directive, continue to apply.

The AI Act separately defines an “AI system”, a “general-purpose AI model” and “training data”, but only for the purposes of that Act. The European Commission has also issued guidelines on the scope and application of the definitions of “AI system” and “general-purpose AI model”. Neither the statutory definitions nor the Commission’s guidelines have express bearing on the application of EU IP law.

The application of those IP concepts to AI is, however, still developing. In Like Company v Google Ireland (C-250/25), pending before the CJEU, the Court is considering whether LLM training constitutes reproduction under Article 2 of the InfoSoc Directive and whether the Article 4 TDM exception under the DSM Directive applies.

There is no dedicated EU body for AI-related IP disputes. The ordinary IP institutions remain responsible for IP disputes, while the AI Act introduces a separate regulatory layer for AI systems and models.

• EUIPO and EU trade mark/design courts: Any AI-related issue concerning an EU trade mark or registered EU design continues to be assessed under the ordinary EU IP framework. EUIPO handles administrative proceedings, while infringement disputes remain before the designated national courts. Questions of EU law may reach the CJEU through preliminary references, and EUIPO Board of Appeal decisions may be reviewed by the General Court and, on points of law, the CJEU.
• EPO and EPO Boards of Appeal: Any AI-related patent issue continues to be assessed under the ordinary EPC framework. The EPO remains responsible for patent grant and certain validity proceedings, with appeals from its examining and opposition divisions handled by the EPO Boards of Appeal. The EPO does not determine patent infringement.
• UPC and national patent courts: AI-related patent infringement and validity disputes are assessed by the Unified Patent Court (UPC) or competent national courts, as applicable, under the ordinary patent framework.
• National courts: These remain the principal forums for copyright, trade-secret and other IP disputes. Questions of EU law may be referred to the CJEU, which therefore may become involved through the ordinary EU judicial mechanisms.

Separately, the AI Office and national competent authorities enforce the AI Act, including the GPAI obligations relating to copyright compliance and training-content transparency. They do not determine whether IP rights subsist or have been infringed.

There is no sui generis IP right in an “AI system” as such. Protection is assessed component by component under the ordinary EU IP framework.

Source and Object Code

Source and object code are protected as computer programs under the Software Directive where they are original in the sense of being the author’s own intellectual creation. Copyright protects the code as expression, not the underlying functionality, programming languages or data formats (SAS Institute, C-406/10).

Model Architecture, Weights and Parameters

An abstract architecture or concept is not protected by copyright as such, although its particular expression in code, documentation or other expressive material may be protected. The architecture may also form part of a patentable technical invention.

Copyright protection of weights and parameters is uncertain and will depend on whether the material reflects human creative choices. They are not patentable merely as abstract mathematical matter, but may contribute to a patentable technical invention where they form part of a technical solution to a technical problem. In practice, weights and parameters may therefore be protected principally through trade secrets and contract, where the relevant requirements are satisfied.

Embeddings, Prompts and Datasets

Individual embeddings are unlikely to qualify as copyright works in the absence of sufficient human creative choices. Structured embedding stores may, however, qualify for database copyright where the selection or arrangement is original, or for the sui generis database right where there has been substantial investment in obtaining, verifying or presenting the contents (British Horseracing Board v William Hill, C-203/02). Prompts and system instructions are protected only where they satisfy the ordinary originality test. Short, functional prompts are unlikely to do so, while more elaborate prompts or curated prompt libraries may qualify. Training datasets, documentation and benchmarks may attract copyright or database protection, while non-public materials may also qualify as trade secrets subject to the applicable requirements.

APIs and Interfaces

Application programming interface (API) code may attract copyright protection as a computer program under the Software Directive, but functionality as such is not protected (SAS Institute, C-406/10). A graphical user interface (GUI) is not protected as a computer program under the Software Directive, but may independently attract copyright under the InfoSoc Directive where it is the author’s own intellectual creation (BSA, C-393/09). API names may also qualify for trade mark protection under the EUTMR.

As set out in 2.1 Protectable Elements of AI Systems, AI software and related materials are protected, where applicable, under the ordinary EU copyright framework. For computer programs, the Software Directive also protects qualifying preparatory design material. Copyright protects expression, not ideas, methods or functionality: in SAS Institute (C-406/10), the CJEU excluded functionality, programming languages and data-file formats from protection, while BSA (C-393/09) confirmed that a GUI is not protected as a computer program, although it may qualify for copyright independently.

The Software Directive also permits lawful users, subject to conditions, to observe, study and test software and to decompile it where necessary for interoperability. As noted in 2.1 Protectable Elements of AI Systems, copyright in model weights and other intermediate computational artefacts remains uncertain; there is currently no CJEU authority establishing copyright in such material as such.

The same originality test applies to prompt libraries and other collections. Individual contents may be protected where sufficiently original, while a collection may separately attract copyright where its selection or arrangement reflects the author’s own intellectual creation (Football Dataco, C-604/10). By contrast, copyright in model weights is doubtful: they are ordinarily produced through computational optimisation rather than authored through human free and creative choices, and EU law has no equivalent of the UK’s computer-generated-works regime.

AI-related inventions may be protected through national patent systems or, more commonly across Europe, through European patents granted under the EPC, including European patents with unitary effect in participating EU member states (“Member States”). The discussion below focuses on the EPC/EPO framework, while national patent laws and practice may also apply.

Eligible Subject Matter and Technical Effect

The EPO Guidelines for Examination 2026 (G-II, 3.3.1) treat AI/ML models and algorithms as mathematical methods under Article 52(2) and (3) of the EPC. They may nevertheless be patentable where they contribute to a technical solution to a technical problem, including applications such as medical devices and image, signal, audio or speech processing. An abstract AI application without a technical purpose is not patentable merely because it uses sophisticated AI. Under the COMVIK approach (T 641/00), only features contributing to the technical solution are taken into account for inventive step.

Sufficiency and Disclosure

According to Article 83 of the EPC and the EPO Guidelines for Examination 2026 (F-III, 3) on sufficiency of disclosure, an AI-related invention must be disclosed sufficiently to enable the skilled person to reproduce the claimed technical effect without undue burden. Where that effect depends on the training data or other model characteristics, those features must be disclosed to the extent necessary for reproducibility. The specific training dataset generally need not be identified or disclosed, provided the skilled person can reproduce the claimed effect without it. A mere assertion of improved AI performance is not sufficient.

Does This Differ From Usual Protection

Substantively, this does not differ from usual protection. AI inventions are subject to the same EPC patentability requirements as other inventions. The practical difference lies in their application. AI/ML features must be tied to a technical contribution, and sufficiency can be particularly challenging where the claimed effect depends on training or model characteristics that are not fully reproducible from the application.

Under the Trade Secrets Directive, information qualifies as a trade secret where it is secret, has commercial value because it is secret, and has been subject to reasonable steps, in the circumstances, to keep it secret. The Directive does not prescribe an exhaustive list of protective measures. Model weights, non-public datasets, prompts, evaluation data and deployment know-how may qualify where these conditions are satisfied.

Maintaining Secrecy

What constitutes reasonable steps is fact-specific. Measures such as access restrictions, confidentiality arrangements or technical security may be relevant. Public disclosure will generally prevent the disclosed information from retaining trade-secret status.

Secrecy and Disclosure

Trade-secret protection does not prevent all disclosure. Article 9 of the Directive requires safeguards for confidential information in legal proceedings, including measures limiting access to such information.

Under Article 53(1)(d) of the AI Act, GPAI providers must publicly disclose a summary of training content, although the prescribed template seeks to protect trade secrets and permits limited, justified redactions.

Information submitted to EU institutions may also be subject to access requests under Regulation (EC) No 1049/2001 on public access to EU institutions’ documents, although access must generally be refused where disclosure would undermine commercial interests, including IP, unless an overriding public interest requires disclosure.

Does This Differ From Usual Protection?

There is no AI-specific trade-secret right. The regime is particularly relevant to AI because weights, datasets and development know-how may depend on confidentiality where other IP protection is limited.

Datasets, corpora, annotations, labels, embeddings and synthetic datasets may be protected under the ordinary EU IP framework. There is no AI-specific database right.

The Database Directive provides two distinct forms of protection:

• Database copyright: This protects the selection or arrangement of contents where it reflects the author’s own intellectual creation. Labour, skill or investment alone is insufficient (Football Dataco, C-604/10).
• Sui generis database right: This protects substantial investment in obtaining, verifying or presenting the contents. Investment in creating the underlying data does not qualify (British Horseracing Board v William Hill, C-203/02; Fixtures Marketing, C-338/02).

This distinction between creating data and obtaining, verifying or presenting it may be particularly relevant to synthetic data and model-generated labels.

Other Protection

Individual materials within a dataset may separately attract copyright protection. Non-public datasets and related materials may qualify as trade secrets, while contracts may govern access and permitted uses.

Competition law may also be relevant because data can be an important source of market power. The European Commission has examined the use of non-public third-party seller data under Article 102 of the TFEU in Case AT.40462 – Amazon Marketplace. The DMA also addresses access to strategically important data, including the Commission’s 2026 Alphabet proceedings on access to Google Search data, which concern sharing anonymised Google Search data with eligible search engines, including AI chatbots with search functionality. In addition, some Member States recognise abuse of economic dependence under national law, which may be relevant where a business is commercially dependent on access to data or a data-enabled service that is considered essential or strategically important.

Using copyright works to train or fine-tune AI may engage the reproduction right under Article 2 of the InfoSoc Directive. Where reproduction occurs, permission is required unless an exception applies. The principal exceptions relevant to AI training are the TDM exceptions in Articles 3 and 4 of the DSM Directive.

TDM Exceptions

Article 3 of the DSM Directive permits TDM by research organisations and cultural heritage institutions for scientific research, subject to lawful access. Article 4 is broader and may apply to commercial AI development, provided the works are lawfully accessible and the rightsholder has not expressly reserved their use. For works made publicly available online, the reservation must be made by appropriate machine-readable means.

Case Law and Pending Clarification

National courts have begun addressing these issues. In Kneschke v LAION (OLG Hamburg, 5 U 104/24, 10 December 2025), the Court considered the application of the TDM exceptions to copying undertaken in creating an image-text dataset and the effectiveness of rights reservations. In GEMA v OpenAI (LG München I, 11 November 2025), the Court held that protected lyrics memorised in model parameters constituted reproductions. These decisions are national and do not establish an EU-wide rule. EU-wide clarification may shortly come from Like Company v Google Ireland (C-250/25), in which the CJEU has been asked directly whether using protected material to train an LLM constitutes reproduction and, if so, whether the exception in Article 4 of the DSM Directive applies. The Advocate General’s Opinion is expected on 3 September 2026.

Application to AI Training and Areas of Uncertainty

There is no separate AI-training exception. Scraping, downloading and assembling copyright works into training or fine-tuning corpora will commonly involve reproduction. Whether later stages, such as tokenisation, embedding or training itself, involve further reproduction depends on whether protected expression is reproduced at those stages. The principal uncertainty is whether and to what extent the Article 4 TDM exception applies to generative-AI training. A 2025 European Parliament study identified a potential mismatch between generative-AI training and the existing TDM framework, including uncertainty around rights reservations, transparency and licensing. The EUIPO’s 2025 study likewise identified unresolved practical issues in applying existing copyright rules to generative AI. As a result, the European Commission has launched a review of the DSM Directive and is considering targeted measures addressing generative AI, including licensing and enforcement. It has also examined the feasibility of an EU-level registry for Article 4 TDM reservations, given their importance for AI training.

As discussed in 3.1 Use of Copyright Works for Training, the DSM Directive contains two specific TDM exceptions: Article 3, for scientific research by specified institutions; and Article 4, which may also apply to commercial AI development.

• Scientific-research TDM: Article 3 permits qualifying research organisations and cultural heritage institutions to reproduce and extract lawfully accessible works and other protected subject matter for scientific research. The exception is mandatory and cannot be overridden by a rightsholder’s reservation of rights. Copies may be retained securely for scientific research, including verification of research results. It may also apply to qualifying public-private research partnerships, subject to the statutory conditions.
• General TDM, including commercial use: Article 4 is available more broadly, including potentially to commercial AI developers, for TDM of lawfully accessible works and other protected subject matter. Unlike Article 3, it is subject to the rightsholder’s express reservation of rights. For works made publicly available online, the Directive specifically contemplates reservations made by appropriate machine-readable means. Copies may be retained only for as long as necessary for TDM.

Lawful Access

Lawful access is required for both exceptions. It may arise through a licence, subscription, institutional access or legitimate public availability. Lawful access and the Article 4 reservation mechanism are separate: material may be lawfully accessible but still fall outside Article 4 because the rightsholder has validly reserved TDM rights.

Temporary Copying and Other Exceptions

Article 5(1) of the InfoSoc Directive permits certain temporary reproductions that are transient or incidental, integral to a technological process and necessary for a lawful use, provided they have no independent economic significance. The CJEU applies these conditions strictly (Infopaq, C-5/08; Filmspeler, C-527/15). It may cover ephemeral technical copies generated during AI processing, but is unlikely to cover persistent training datasets, embeddings, checkpoints or model weights.

Areas of Uncertainty

For the reasons outlined in 3.1 Use of Copyright Works for Training, the topic of the TDM exceptions for the purpose of AI model development remains subject to significant uncertainty. Further clarification may come from Like Company v Google Ireland (C-250/25), which asks the CJEU whether LLM training involves reproduction and, if so, whether Article 4 applies.

Market Practice

Where the TDM exceptions do not apply, licensing is the principal route to lawful use of protected training content. The market remains fragmented, with agreements largely bilateral between AI providers and publishers, news organisations, music companies and other rightsholders. Terms are generally confidential and may provide for fixed, recurring, usage-based or revenue-sharing remuneration. Attribution remains principally contractual.

Collective Licensing and Remuneration

There is no EU-wide collective or compulsory licence for AI training. Collective licensing is governed in part by the Collective Rights Management Directive, including its provisions on multi-territorial licensing of rights in musical works for online use in the internal market (2014/26/EU), while Article 12 of the DSM Directive allows Member States to establish extended collective licensing for defined uses, subject to safeguards. Voluntary collective licensing is developing in sectors such as music, and a notable example is GEMA’s German AI licensing model, which provides for licensing protected music for AI training and related uses.

Remuneration

Articles 18–23 of the DSM Directive establish protections concerning appropriate and proportionate remuneration, transparency and contract adjustment where authors and performers license or transfer rights.

EU Developments

The European Parliament’s March 2026 resolution on copyright and generative AI called for a functioning licensing market, including sector-specific voluntary collective licensing, fair and proportionate remuneration, good-faith negotiations and transparency. It also called for consideration of remuneration for past uses where licensing markets had not previously existed. The European Commission has launched a review of the EU copyright framework to consider ways to enhance the licensing and enforcement of copyright and related rights in the AI context, and to improve the conditions for creators’ remuneration, while making it easier for providers of generative AI to access copyright-protected content.

Statutory Reservation Mechanism

Article 4(3) of the DSM Directive allows rightsholders to expressly reserve their rights against general TDM. For works made publicly available online, the reservation should be expressed through appropriate machine-readable means. A valid reservation prevents reliance on the Article 4 exception. Article 53(1)(c) of the AI Act separately requires GPAI providers to adopt a copyright-compliance policy and identify and comply with Article 4(3) reservations using state-of-the-art technologies.

Technical and Contractual Signals

EU law does not prescribe a single opt-out protocol. Machine-readable mechanisms such as the Robot Exclusion Protocol (robots.txt) may be used to communicate reservations, while metadata, website terms and other technical or contractual notices may also be relevant. Their legal effect depends on whether they constitute an appropriate reservation under Article 4(3). In Kneschke v LAION (OLG Hamburg, 5 U 104/24, 10 December 2025), the Hamburg Higher Regional Court considered the effectiveness of a natural-language reservation and found it insufficiently machine-readable.

Autonomous Agents

The fact that a reservation is encountered by an AI agent does not alter the reservation’s legal effect. Failure to respect a valid reservation prevents reliance on the Article 4 exception. For GPAI providers, failure to maintain and implement a policy to identify and comply with Article 4(3) reservations, including through state-of-the-art technologies, may also constitute non-compliance with Article 53(1)(c) of the AI Act. The GPAI Code of Practice addresses robots.txt and other appropriate machine-readable reservation protocols.

Developing Standards

The European Parliament’s March 2026 resolution proposed a limited number of standardised machine-readable opt-out formats potentially managed or listed by EUIPO. The European Commission’s 2026 feasibility study considers an EU-level registry as a possible complement to existing mechanisms for recording and identifying Article 4(3) reservations.

Documentation and Transparency

EU law does not generally require a complete record of all training or retrieval activity. For GPAI providers, however, Article 53 of the AI Act requires technical documentation, a copyright-compliance policy and a public summary of training content. These obligations do not require disclosure of the complete training dataset.

For agentic systems, there is no general EU requirement to retain every tool call, browsing or API event, although such records may be relevant to demonstrating compliance or establishing responsibility in a dispute.

Cross-Border Training

Where datasets, servers, developers and users are in different jurisdictions, the applicable law depends on the location of the relevant act. For copyright, the country-of-protection principle – reflected in Article 5(2) of the Berne Convention and, in the EU, in Article 8 of Rome II – generally applies the law of the country for which protection is claimed, so the location of the server or provider is not decisive and more than one country’s law may apply. The Munich Regional Court’s decision in GEMA v Suno (42 O 763/25, 31 July 2026) illustrates this: the court applied US law to the training carried out in the USA and German law to the acts occurring in Germany. The judgment is not final and is likely to be appealed.

There is no AI-specific infringement regime or horizontal EU rule governing attribution for autonomous AI acts. A claimant must establish the ordinary elements of infringement for the relevant IP right and act.

• Copyright: Training, fine-tuning, evaluation and retrieval may involve reproduction or extraction. Under Article 2 of the InfoSoc Directive, reproduction extends to reproduction in whole or in part where protected expression is reproduced (Infopaq, C-5/08). Whether a training-related reproduction is lawful may depend on the TDM exceptions addressed in 3. Training Data and Model Development. Persistent memorisation may raise a separate reproduction issue: in GEMA v OpenAI and GEMA v Suno, the Munich Regional Court treated memorised lyrics/music in model parameters as reproductions. These are national judgments and are not final.
• Datasets: For database rights, infringement depends on whether protected database contents have been unlawfully extracted or re-utilised. The sui generis right claim therefore requires the claimant to establish that a qualifying database exists and that there has been a protected extraction or re-utilisation within its scope.
• Trade marks and patents: Trade mark infringement depends on the relevant use of the sign under Article 9 of the EUTMR, including use in the course of trade and the applicable conditions. Patent infringement depends on performance of an act falling within the patent claims in the relevant territory; AI-generated or automated performance does not by itself alter that assessment.

Liability for Automated and Agentic Acts

EU law contains no horizontal AI-specific rule determining attribution for acts performed autonomously by an AI system. The system is not itself a legal person. Responsibility therefore falls, where the ordinary IP rules permit, on the developer, provider, deployer or user whose conduct caused, authorised or controlled the relevant act, depending on the circumstances and the right concerned. The fact that the act was carried out automatically does not, by itself, prevent attribution to a human or legal person. For agentic systems, the allocation of responsibility will therefore depend on the system’s configuration, instructions, degree of autonomy and the respective roles of the provider and deployer.

EU law does not establish that model weights, parameters or embeddings are inherently copies of training works. The key question is whether protected expression is reproducibly embodied in the model or other artefact. The same analysis applies to caches, checkpoints and similar intermediate materials.

Recent German litigation suggests that persistent memorisation capable of reproducing protected expression may constitute reproduction. In GEMA v OpenAI and GEMA v Suno, the Munich Regional Court treated memorised lyrics or musical works in model parameters as reproductions. These judgments do not establish that all information encoded in model parameters is itself an infringing reproduction.

For databases, the issue is whether an artefact contains or reproduces a protected part of the database or whether its creation involved unlawful extraction or re-utilisation.

For copyright, similarity alone does not establish infringement. The issue is whether the output or artefact reproduces protected expression from the earlier work. Infopaq (C-5/08) confirms that infringement can arise from reproduction of a protected part of a work, even if the work as a whole is not reproduced.

There is no general, harmonised EU doctrine of secondary or contributory IP infringement. Liability for acts committed by users or downstream systems therefore depends on the relevant IP right, the provider’s own conduct and applicable national law.

For copyright, the key distinction is between the provider’s own infringing act and an infringement committed by a user. InYouTube/Cyando (Joined Cases C-682/18 and C-683/18), the CJEU held that merely providing a platform does not normally make the operator itself responsible for a user’s communication to the public; the provider’s knowledge, control and active contribution may, however, be relevant. This is relevant by analogy to AI providers, but does not establish a general AI-specific liability rule.

The Digital Services Act provides the EU framework for intermediary liability for user-supplied content, including the hosting safe harbour and prohibition on general monitoring. Article 8(3) of the InfoSoc Directive also permits injunctions against intermediaries whose services are used for copyright infringement. These provisions concern intermediary status and remedies; they do not determine whether the underlying IP act infringes.

For patents and trade marks, there is no comparable harmonised EU regime of secondary liability. Questions of authorisation, procurement or accessory liability are principally governed by applicable national law within the relevant UPC or national framework.

The CJEU’s recent judgment in WebGroup/Coyote (Joined Cases C-188/24 and C-190/24) provides further guidance on the limits of the hosting regime where a provider exercises control over stored information and its dissemination. This may be relevant by analogy to AI providers that actively control the selection, retrieval or publication of user-supplied material, but the judgment is not an IP case and does not establish an AI-specific liability rule.

For autonomous agents, automation does not itself create provider liability. The relevant factors may include the provider’s and deployer’s respective roles, instructions, permissions, technical control and contribution to the infringing act.

The Trade Secrets Directive and applicable national law govern the unauthorised acquisition, use or disclosure of confidential information. There is no AI-specific regime. Training, fine-tuning, retrieval, prompting, storage or evaluation may constitute unlawful use where the information qualifies as a trade secret and the relevant statutory or contractual requirements are met. Contractual confidentiality obligations may apply independently.

The principal AI-specific issues are control, provenance and proof of what information entered or remained in the system. A model may reproduce confidential information in an output, while the position where it merely derives or infers information without reproducing the source material is less settled. The extent to which confidential information is retained in weights, memory, retrieval stores or logs may therefore be important. Training records, prompts, access logs, tool calls and model documentation may be central to proving acquisition, use or disclosure, subject to the safeguards applicable to confidential evidence.

Procedural Defences

Depending on the claim and forum, a developer or provider may raise jurisdiction, applicable law, standing, limitation periods and evidential or disclosure objections, subject to the applicable EU and national procedural rules. EU IP enforcement rules also provide safeguards concerning evidence, interim measures and confidentiality.

Substantive Defences

The ordinary IP defences apply, including absence of a protected right or restricted act, licence or authorisation, statutory exceptions and limitations, exhaustion, and failure to establish the relevant territorial or other conditions. Independent creation may be relevant to copyright where it shows that protected expression was not copied. The EU has no general fair-use or fair-dealing defence.

• For copyright, the principal AI-specific limitation is the TDM framework in Articles 3 and 4 of the DSM Directive, addressed in 3. Training Data and Model Development. Other exceptions may apply where their conditions are met, including quotation, education, research, and parody or pastiche. The judgment in Deckmyn (C-201/13) provides the principal CJEU guidance on parody, while the one in Pelham II (C-590/23) clarifies the scope of pastiche. Mere imitation of style, without reproduction of protected expression, is not itself sufficient for copyright infringement.
• For patents and trade marks, the applicable statutory limitations and exhaustion rules apply in the ordinary way.
• For trade secrets, the Trade Secrets Directive recognises exceptions including lawful acquisition, independent discovery and reverse engineering.

Other Arguments

Competition law may be relevant to particular licensing or enforcement conduct, but it is not a general defence to IP infringement.

There is no AI-specific EU test for copyright infringement by outputs; ordinary copyright rules apply. An output may infringe where it reproduces protected expression, in whole or in part, or where its use constitutes an unauthorised communication of a protected work to the public, subject to applicable exceptions and limitations.

For reproduction, the question is whether elements expressing the author’s own intellectual creation have been reproduced, rather than whether a “substantial part” has been taken (Infopaq, C-5/08). Even a short extract may qualify: in Infopaq (C-5/08), the CJEU held that an 11-word extract could constitute reproduction in part if it contained elements expressing the author’s own intellectual creation.

The same principles apply where protected expression from a prompt or reference image is reproduced in the output. By contrast, imitation of style, genre, voice or artistic technique is generally insufficient for copyright infringement since copyright protects original expression rather than ideas, methods or styles as such.

Where a user supplies a protected work as a prompt or reference, the user’s own acts may engage copyright. Whether reproduction in the resulting output is attributable to the user or provider remains unsettled. In Like Company v Google Ireland (C-250/25), the CJEU has specifically been asked whether reproduction of protected content in a chatbot response following a user prompt constitutes an act of reproduction by the service provider.

A user, deployer or customer may infringe copyright unwittingly where generating or using an AI output involves reproduction of protected expression. Publishing or commercialising the output may additionally engage the communication-to-the-public right under Article 3(1) of the InfoSoc Directive, for which the actor’s deliberate intervention and knowledge can be relevant. In GS Media (C-160/15), for example, the CJEU held that knowledge or reasonable knowledge of the unauthorised nature of the underlying publication was relevant, with knowledge presumed where hyperlinks were provided for profit.

Prompt design, human review and control may therefore affect attribution. Commercial purpose and scale may affect liability and remedies.

AI-generated output may infringe an EU trade mark where a protected sign is used in the course of trade and the requirements of Article 9 of the EUTMR are met, including likelihood of confusion or, for marks with a reputation, unfair advantage or detriment. In L’Oréal v Bellure (C-487/07), the CJEU confirmed that taking unfair advantage may arise even without confusion, including where a trader seeks to benefit from a mark’s reputation. In Louboutin v Amazon (C-148/21 and C-184/21), the CJEU held that an online marketplace may itself be regarded as using a third party’s trade mark where users could perceive the third-party advertisement as forming part of the marketplace’s own commercial communication. This may be relevant by analogy when determining whether a trade mark appearing in AI-generated output is used by the user, provider or deployer.

AI outputs that falsely suggest affiliation, sponsorship or endorsement may also engage EU consumer protection and advertising rules, including rules against misleading commercial practices and misleading advertising. Passing off, broader unfair-competition claims and personality rights, including protection of celebrity names, images, voices or other indicia, are principally matters of Member State law.

No AI-specific infringement regime applies to patents, designs or AI-generated products, which are all assessed under the existing rules.

Patents

For patents within the UPC’s jurisdiction, the Agreement on a Unified Patent Court (UPCA) gives the proprietor the right to prevent specified direct and indirect uses of the invention without consent. Direct use includes making, offering, placing on the market or using a patented product; using or, in specified circumstances, offering a patented process; and dealing in a product obtained directly by a patented process (Article 25 UPCA). Indirect use includes supplying or offering means relating to an essential element of the invention for putting it into effect, where the supplier knows or should know that those means are suitable and intended for that purpose (Article 26 UPCA). Member State patent law applies where the patent is outside the UPC’s jurisdiction.

Designs

For EU designs, infringement may arise where, without the holder’s consent, a third party uses a design that does not produce a different overall impression on the informed user. Relevant acts of use include making, offering, placing on the market or using a product incorporating the design, and certain acts involving creating, downloading, copying and sharing or distributing to others any medium or software which records the design.

There is no horizontal EU rule governing attribution of autonomous AI acts, and an AI agent is not itself treated as the legally responsible actor. Attribution is therefore generally determined under applicable Member State law, which tends to attribute an AI system’s actions to a natural or legal person, although approaches may differ, particularly for unintended or unpredictable actions.

Accordingly, an agent’s autonomous actions such as scraping, retrieval or copying may engage copyright or database rights; generation or publication may engage copyright or trade mark rights; and implementation of technical or design choices may engage patents or designs. Liability then depends on attribution of the relevant act under the applicable IP and national rules. Separate third-party liability may also arise where another person facilitates or contributes to the infringement.

Separately, the Product Liability Directive (2024/2853/EU) provides a product-liability regime for defective products, including software, which may be relevant where an AI system or AI-enabled product causes compensable damage.

EU copyright law does not provide for AI authorship or a separate copyright regime for outputs lacking sufficient human creative contribution. Copyright protection requires an original expression reflecting the author’s free and creative choices (Cofemel, C-683/17).

Applied to AI-assisted outputs, the key question is therefore whether, and to what extent, the final expression reflects human creative choices, for example through direction, selection, arrangement or editing (Painer, C-145/10). The greater the system’s autonomy, particularly in agentic systems, the more difficult it may be to establish the necessary human creative contribution.

As discussed in 6.1 Human Authorship and Copyright Protection, EU copyright protection requires sufficient human creative contribution. Unlike UK law, EU law has no separate regime for computer-generated or purely AI-generated works. Accordingly, where that human contribution is absent, there is no special EU rule conferring copyright.

The treatment of joint authorship, derivative works and adaptations is not fully harmonised at EU level. Joint authorship – whether several human contributors qualify as authors and how their ownership shares are divided – is principally determined by Member State law, subject to the EU originality standard.

EU law has no general adaptation or derivative-work right. However, transforming an existing work may still require permission if protected expression is reproduced under the InfoSoc Directive, with additional adaptation rights potentially arising under national law. For software and copyright-protected databases, EU legislation expressly grants rights over adaptation and alteration.

There is no EU copyright registration system and therefore no EU-level requirement to disclose AI involvement. In enforcement, however, authorship, originality and ownership may need to be established, making the extent of human contribution relevant.

For EPO patents, there is no general requirement to disclose the use of AI, but an inventor must be designated under the EPC. Where AI is material to the claimed invention, ordinary sufficiency requirements may require disclosure of relevant AI methods, training data or other information necessary for the skilled person to reproduce the claimed technical effect. The EPO Guidelines for Examination expressly address this issue for AI and machine-learning inventions.

An AI system cannot be designated as inventor under the EPC: the inventor must be a natural person (J 8/20; J 9/20). In T 0528/25 (5 February 2026), the EPO Board of Appeal confirmed this principle, while also holding that an inventor may in principle be designated for an invention developed using AI. The designation must, however, clearly and consistently identify the human inventor.

The EPC does not establish a specific minimum-human-contribution test for AI-assisted inventions. Entitlement follows the ordinary Article 60 EPC rules, including succession in title and the applicable rules for employee inventions.

Under the EPC, AI does not itself change the patentability standards. Inventive step may nevertheless be affected where AI tools were routinely available to the skilled person in the relevant field, since their capabilities may form part of the assessment of what would have been obvious at the relevant date.

For sufficiency, details of an AI system may need to be disclosed where necessary to reproduce the claimed technical effect. AI-generated material may constitute prior art if it was publicly available before the relevant date and provides an enabling disclosure; AI generation alone does not determine its prior-art status.

Designs

AI-generated designs can be protected through the ordinary EU design regime. Both registered and unregistered EU designs are governed by Regulation (EU) 2026/715. Under the regulation, the definition of “product”, whose design can be protected, covers both physical and non-physical products, expressly including GUIs, icons, graphic works and spatial arrangements. Therefore, GUIs, icons, avatars and virtual goods are registrable EU designs.

Protection of a sign, whether registered or unregistered, solely depends on novelty and the presence of an individual character, rather than creativity or authorship (Deity Shoes, C‑323/24). Thus, AI involvement does not preclude EU design protection. However, the registrability is limited by features dictated solely by technical function, as these are non-registrable (Doceram, C‑395/16).

The EU design right is vested in the designer or its successor in title and, where several persons jointly develop the design, to them jointly. If the designer is an employee and the development of the design is conducted within the employee’s duties, the right to the EU design is vested in the employer, unless otherwise agreed or provided under applicable national law.

Trade Dress

“Trade dress” is not an autonomous concept in EU law. However, it may fall within the protection of trade marks as is the case with the layout of the Apple Store and the red sole of Louboutin (Apple Store, C‑421/13; Louboutin, C‑163/16). This is, however, limited by the shape exclusions (Hauck, C‑205/13; Lego Juris, C‑48/09 P), which preclude protection of shapes which are imposed by the very nature of the goods.

Trade Mark Registration

EU trade mark law contains no authorship or human-creation requirement; how a sign was devised is irrelevant to its registrability. Under Article 4 of the EUTMR, any sign – in particular, a name, logo, slogan or sound – may be registered if it can distinguish goods or services and be represented clearly and precisely.

AI-generated signs are assessed under the ordinary regime of absolute grounds for refusal listed in Article 7 of the EUTMR. While slogans do not need to achieve a threshold of originality (Audi, C‑398/08 P), sounds may fail where they are perceived to lack a distinctive character, for example by being purely functional (Ardagh, T‑668/19).

The practical risks for AI-generated trade marks stem from earlier, existing rights as AI outputs may reproduce or approximate existing marks. Additionally, generative AI has made it easier to file signs at scale not with an intent to use them, but rather to block their use by third parties. This “bad faith” practice can, however, invalidate the trade mark (SkyKick, C‑371/18; Koton, C‑104/18 P).

Ownership

The proprietor of the trade mark is the one that registers the mark, which can be any natural or legal person (Article 5 EUTMR).

Moral Rights

Moral rights are not harmonised at EU level. Recital 19 of the InfoSoc Directive leaves them to Member States’ law, the Berne Convention, the WCT and the WPPT.

These rights presuppose a protected work and a human author. Therefore, purely AI-generated output has no author and, following on from that, no moral rights. AI-assisted works could attract them to the extent human-authored expression is present. The European Parliament’s resolution of 10 March 2026 on copyright and generative artificial intelligence would keep purely AI-generated content outside copyright (paragraph 25). While this is non-binding, it indicates the possible direction of the evolution of EU copyright law.

Attribution and Integrity

A creator’s style as such is not protected: imitation infringes only where protected expression is reproduced, subject to exceptions. Two relevant copyright exceptions are the pastiche exception, ie, the deliberate reproduction of a creator’s style, and the parody exception (Article 5(3)(k) InfoSoc Directive, Pelham II (C‑590/23) and Deckmyn (C‑201/13)).

There is no harmonised false-attribution regime. Works falsely presented as a named person’s creation can be addressed through national moral-rights and personality laws, the Unfair Commercial Practices Directive (misleading claims) (2005/29/EC) and claims of misleading consumers through trade mark law (PMJC, C‑168/24).

Finally, Article 50 of the AI Act mandates the machine-readable marking of AI-generated content and deepfake disclosure, so that false attribution to human creators of AI-generated works is precluded.

There are no unitary EU publicity or personality rights; protection is a patchwork of privacy and IP-adjacent rights.

A person’s name, image and voice are personal data once identifiable, so generating or deploying a replica requires a lawful basis under the GDPR and is subject to erasure and objection rights. Where such data reveal racial, ethnic or belief characteristics, the special-category regime of Article 9 of the GDPR may also apply. Additionally, the fundamental right to privacy (Articles 7 and 8 EU Charter and Article 8 ECHR) protect image and voice more generally (ECtHR Reklos and Davourlis v Greece and ECtHR P.G. and J.H. v UK).

Performers separately hold neighbouring rights over fixations of their performances (Articles 2(b) and 3(2)(a) InfoSoc Directive), engaged when AI reproduces recordings (training, sampling, outputs). Wholly synthetic imitations reproducing no fixation fall outside these rights — the “digital-replica gap”.

Article 5 of Directive (EU) 2024/1385 on combating violence against women and domestic violence outlaws the production and distribution of non-consensual intimate deepfakes, and Article 16 of the Digital Services Act gives individuals a notice-and-action route to require platforms to remove infringing content once notified.

There is no separate forum or procedure for AI-related IP disputes. They are brought before the ordinary courts, tribunals and IP bodies competent for the underlying right, as described in 1.4 Courts, IP Offices and Regulators. The appropriate forum therefore depends on the IP right and the relief sought, rather than on the involvement of AI.

Enforcement is governed by Directive 2004/48/EC on the enforcement of intellectual property rights (IPRED), supplemented by the European Commission’s non-binding Guidance (COM(2017) 708 final). IPRED establishes minimum standards, while Member States may provide more favourable measures (Article 2(1)); procedural rules therefore vary nationally.

Evidence may be obtained through IPRED’s targeted mechanisms, including production of specified evidence (Article 6), preservation measures (Article 7) and the right of information (Article 8). These measures must be specific, relevant and proportionate, rather than permitting general discovery. In principle, they may reach training data, model documentation, prompts, outputs, logs and agent records where relevant to the claim. Confidential information may be protected through measures such as restricted access and redaction, alongside the protections in Article 9 of the Trade Secrets Directive.

For patents, national courts apply the relevant national procedures, while the UPC has its own evidence and preservation regime, including orders under Article 60 of the UPCA.

The IPRED framework provides interim measures (Article 9), final injunctions, including against intermediaries (Article 11), and corrective measures such as recall, removal and destruction (Article 10), subject to proportionality (Article 3). Article 8(3) of the InfoSoc Directive also permits copyright injunctions against intermediaries whose services are used to infringe.

For patents, the UPC provides its own interim and final injunctions and corrective measures (Articles 62–64 UPCA).

Applied to AI, these powers may support targeted measures to halt infringing training, deployment or distribution; remove protected material from datasets; block infringing outputs or impose filters; quarantine affected model versions; or restrict agent tools and autonomous publication functions. More far-reaching remedies, such as deletion or retraining of a model because of infringing training data, remain largely untested and would be subject to proportionality.

Under Article 13 of IPRED, an infringer who knew or had reasonable grounds to know of the infringing activity may be ordered to compensate the rightsholder for the full actual prejudice suffered. Damages may take account of lost profits, unfair profits and, where appropriate, moral prejudice, or be assessed as a lump sum based at least on the hypothetical royalty or licence fee. The European Commission Guidance confirms that the latter is an alternative method of assessing prejudice and is not necessarily limited to a single royalty; Liffers (C-99/15) also confirms that moral prejudice may be recovered alongside royalty-based damages. The objective is full compensation, not punitive damages.

Where the infringer did not know and had no reasonable grounds to know, Article 13(2) permits Member States to provide instead for recovery of profits or pre-established damages. Accordingly, lack of knowledge may affect the monetary remedy even where it does not prevent a finding of infringement.

Under Article 14, reasonable and proportionate legal costs and other expenses are generally borne by the unsuccessful party, unless equity requires otherwise. United Video Properties (C-57/15) confirms that recovery must cover a significant and appropriate part of reasonable costs incurred.

Training-content licences should define precisely what rights are licensed, for what AI uses and by whom, distinguishing copyright, database and related rights. They should expressly address training, fine-tuning, evaluation, benchmarking, RAG and other retrieval, embeddings, and permitted downstream uses and outputs. Core terms should cover permitted models and successor versions, territory, duration, exclusivity, remuneration, attribution, sublicensing, audit and reporting, security, and the handling of rights reservations and opt-outs.

For agentic systems, the licence should also define the scope of autonomous retrieval and use, including permitted sources, API and tool access, caching, logging, onward use and publication. Termination should address deletion or continued retention of datasets, embeddings, caches and indexes and, critically, the treatment of models already trained on the licensed material. Warranties and indemnities should allocate contractual responsibility for rights clearance, authorised uses and third-party claims.

As discussed in 1. Legal Framework for AI and IP, EU policy on AI and IP remains under active development. Following the EUIPO and European Parliament studies and Parliament’s March 2026 resolution, the European Commission has launched a review of the EU copyright framework addressing, among other issues, generative AI, licensing and enforcement. Any resulting reforms remain to be seen.

These issues are also likely to become more prominent as the AI Act’s GPAI copyright and transparency obligations are implemented. Article 53 requires GPAI providers to maintain a copyright-compliance policy, including respect for DSM rights reservations, and to publish a summary of training content. While these obligations do not alter substantive copyright law, the resulting transparency and compliance processes may expose practical questions around training, licensing and enforcement that could inform future reform.

The EU participates in international discussions on AI and IP, particularly through WIPO and the WTO/TRIPS framework, alongside broader international AI initiatives through bodies such as the Council of Europe, OECD and G7 and relevant standards-setting processes. AI-related patent issues are also being considered within the European patent system.

Significant divergence nevertheless remains between major jurisdictions, particularly on AI training and copyright exceptions, rights reservations and licensing, transparency and the treatment of AI-generated outputs. The EU is distinctive in combining the DSM Directive’s TDM regime with the AI Act’s GPAI copyright and transparency obligations. These differences are likely to remain important for cross-border AI development and deployment despite continuing international efforts towards greater convergence.