AI & Intellectual Property 2026 Comparisons

Japan has no IP legislation specific to AI. Although the Act on Promotion of Research and Development and Utilisation of Artificial Intelligence-related Technology (the “AI Promotion Act”), a framework statute on AI, took effect in 2025, it does not contain rules on intellectual property.

Japanese law contains no comprehensive, flexible exception comparable to US fair use. However, Article 30-4 of the Copyright Act establishes a text and data mining exception (the “TDM Exception”). Under the TDM Exception, a work may be exploited “in any way” without the copyright owner’s authorisation, provided that the exploitation is not aimed at enjoying or causing another person to enjoy, the thoughts or sentiments expressed in the work. The phrase “in any way” is significant. The exception can cover not only reproduction but also adaptation (such as annotation) and even the provision of works to third parties. Unlike the EU’s Directive on Copyright in the Digital Single Market (Directive (EU) 2019/790; the “DSM Directive”), the exception applies to commercial exploitation, provides no opt-out (rights reservation) mechanism for right holders and does not require lawful access to the work. However, the TDM Exception does not apply where the exploitation would unreasonably prejudice the interests of the copyright owner in light of the nature or purpose of the work or the circumstances of its exploitation (the proviso to Article 30-4). On the relationship between AI and copyright, the Agency for Cultural Affairs (the government agency responsible for copyright) has published a “General Understanding on AI and Copyright in Japan”, which sets out interpretive guidance.

In addition, a non-binding soft-law instrument, the “Principles-Code for Protection of Intellectual Property and Transparency for the Appropriate Use of Generative AI“ (the ”Principles-Code“), was published in August 2026. The Code is intended to protect IP and secure transparency in the development and provision of generative AI (see 10.1 Pending Legislation, Regulation and Policy).

The treatment of AI-assisted inventions under patent law is addressed in 2.3 Patent Protection for AI Technologies and 6.5 Patent Inventorship and Entitlement and developments concerning model transparency are addressed in 10.1 Pending Legislation, Regulation and Policy.

Foreign rights-holders receive the same protection as Japanese nationals. This follows from copyright-related treaties such as the Berne Convention and the TRIPS Agreement, which establish the principle of national treatment and provide the basic framework under Japanese copyright law. In the field of AI governance, the Hiroshima AI Process (HAIP), an international soft-law initiative launched under Japan’s G7 presidency, is significant. Currently, no free trade agreements or foreign judgments have a notable direct influence on the relationship between AI and IP in Japan. In any event, the current IP framework, including Article 30-4 of the Copyright Act, does not distinguish between domestic and foreign rights-holders and imposes no special rules on foreign rights-holders or foreign AI developers.

As Japan has no IP legislation specific to AI, none of the concepts referred to above is defined by statute for IP purposes. As AI-related soft law, the “AI Guidelines for Business” (jointly formulated by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry) contain certain definitions, but these have no direct bearing on the interpretation of IP law. The same applies to autonomous or agentic AI systems (ie, systems that plan tasks, select tools or make API calls with limited human direction): they are not specially defined or treated differently, under IP law.

In Japan, the bodies responsible for AI and IP issues are the same as those that handle IP law generally. The Japan Patent Office (JPO) oversees industrial property rights, including patents for AI-related inventions. The Agency for Cultural Affairs (ACA) is responsible for copyright issues relating to AI training and AI-generated material. Courts adjudicate legal disputes (see 8.1 Venue for details on jurisdiction). In addition, the Ministry of Internal Affairs and Communications (MIC), the Ministry of Economy, Trade and Industry (METI) and the Cabinet Office drive Japan’s overarching national AI policy.

As regards specialist procedures and guidance, these bodies have issued frameworks within the existing system rather than creating entirely separate legal tracks:

• JPO: the JPO has published examination guidelines and case examples for AI-related inventions, clarifying patentability with a focus on inventive step and description requirements for software-related technologies;
• ACA: for copyright, the ACA has published the “General Understanding on AI and Copyright in Japan”, which clarifies how the limitation under Article 30-4 of the Copyright Act applies to AI training and outlines the criteria for determining infringement; and
• MIC/METI: the “AI Guidelines for Business”, jointly formulated by MIC and METI, provide operational and compliance standards for AI developers, providers and users.

Each element of an AI system may be protected under several regimes: copyright, patents and trade secrets or shared data with limited access under the Unfair Competition Prevention Act (UCPA).

Copyright

As there are no AI-specific rules, protection turns on whether each element has creativity, outlined below.

• Source code and object code: source code may be protected as a work of computer programming; object code, being a reproduction of the source code, is protected where the source code is protected.
• Prompts: protected where they have creativity, as may be the case with lengthy prompts.
• Training datasets: protected by copyright where the data themselves have creativity, as with a dataset of images. In addition, under Japanese copyright law, a database is protected as a “database work” where it has creativity by reason of the selection or systematic construction of information (Article 12-2(1) of the Copyright Act). A training dataset may likewise be protected as a database work where creativity is found in the selection or systematic construction involved in its collection, classification or annotation.
• Model weights/parameters and evaluation benchmarks: these ordinarily lack creativity and so do not attract copyright.

Patents

Under the Patent Act, an “invention” is the highly advanced creation of technical ideas utilising the laws of nature (Article 2(1)). An “invention of a product” includes a “computer program, etc” (Article 2(3)(i)) and a “computer program” is defined as a set of instructions given to a computer, combined to produce a specific result (Article 2(4)). Although not direct instructions to a computer, items that specify computer processing in a similar manner (such as data structures) are treated as “equivalent to a computer program” and can also be protected.

• Source code/object code: the code itself is a matter for copyright law. However, if the processing procedure it describes constitutes a technical idea satisfying novelty, inventive step and other requirements, it can be protected as an invention of a program, a process or a product.
• Model architecture: a network structure as such is abstract, but if it is configured as a concrete means of solving a specific technical problem, it can be protected as a product or process invention. The Japan Patent Office (JPO) has published a hypothetical examination case (Case 55 of the “Case Examples pertinent to AI-related Technologies”) indicating that, where an invention is titled “trained model” or the like and it is not clear from the claims and specification that this means a “program” under the Patent Act, the category of the invention is unclear and the clarity requirement is violated.
• Model weights/parameters/embeddings: under a JPO hypothetical examination case (Case 2-14, “trained model for analysing the reputation of accommodation facilities”), a trained model constitutes an “invention” if it is described as causing a computer to perform specific operations; it does not if it is described merely as a parameter set without any such description.
• Prompts: natural-language instructions as such are mere artificial arrangements and, in principle, do not constitute an “invention”. However, if configured as a product or process that produces a technical effect (such as improved processing accuracy) through a specific prompt structure, they can be protected.
• Fine-tuning materials/training datasets: a collection of data as such is, in principle, not protectable. The Examination Guidelines likewise state that something that merely defines the content or order of data elements, remaining a mere artificial arrangement, is not an “invention”. However, novel technical methods of generating or pre-processing data and computers executing them, can be protected as process or product inventions.
• Evaluation benchmarks: as with training datasets, benchmark data as such are not protected. If configured as a novel evaluation algorithm or measurement method or a computer executing them, they can be protected as process or product inventions.
• Documentation: documentation falls outside the Patent Act but may be protected under copyright law.
• APIs and interfaces (including agent workflows, tool configurations, memory systems and orchestration layers): these are assessed under the same criteria as software-related inventions generally. If configured as a processing procedure or data-exchange method solving a specific technical problem or a computer executing them, they are eligible for protection as process or apparatus inventions. They are not protected where they amount to a mere business workflow or an artificial allocation of roles that simply replaces tasks performed by humans.

Trade Secrets and Shared Data with Limited Access

Both regimes arise under the UCPA. A trade secret is information that is:

• managed as secret (secrecy management);
• useful for business activities (usefulness); and
• not publicly known (non-public nature).

Typical examples are manufacturing know-how, customer lists and R&D data. Unauthorised acquisition, use or disclosure of a trade secret constitutes “unfair competition” and may give rise to injunctions, damages and, in some cases, criminal penalties.

Shared data with limited access is a regime protecting technical or business data that, while not a trade secret, is:

• provided to specific persons on a regular basis (limited provision);
• accumulated to a significant extent by electronic or magnetic means (substantial accumulation); and
• managed by electronic or magnetic means (electronic or magnetic management).

Typical examples are databases and training datasets provided under contract. Where the requirements are met, model weights, parameters and datasets may be protected as shared data with limited access; unauthorised acquisition and similar acts by persons without access rights are prohibited and may give rise to injunctions and damages.

As discussed in 2.1 Protectable Elements of AI Systems, under Japanese copyright law, protection turns on whether each item of material has creativity. Descriptions of model architecture, system instructions and prompt libraries may be protected as works to the extent that their expression has creativity. However, ideas, methods, algorithms and functionality as such are not protected by copyright; only their concrete expression is (the idea/expression dichotomy). Model weights and the outputs of intermediate computational processes (such as embeddings) are ordinarily difficult to regard as creative human expression and are unlikely to attract copyright. To date, little case law directly addresses the copyrightability of these materials in the AI context.

In Japan, AI-related inventions made by model developers are assessed within the framework for computer software-related inventions. There is no special treatment for AI as such, but the JPO has published a collection of hypothetical examination cases for AI-related inventions.

Patent-Eligible Subject Matter

Article 2(1) of the Patent Act defines an “invention” as the highly advanced creation of technical ideas utilising the laws of nature. For software-related inventions to be patent-eligible, the information processing performed by the software must be concretely realised using hardware resources to solve a technical problem; this is sometimes expressed as a requirement of co-operation between software and hardware resources. A “trained model” structured as a computer program or a data structure can constitute a patentable “invention”. By contrast, a model claimed merely as a raw parameter set does not constitute a patentable “invention” because it is a mere presentation of information.

Enablement and Support Requirements

The specification must be written so that a person skilled in the art can carry out the invention (enablement, Article 36(4)(i)) and the claimed invention must be stated in the specification (support, Article 36(6)(i)). It is not necessary to disclose raw datasets to satisfy these requirements. However, the correlation between the input data and the output data must either be stated in the specification or be inferable from the common general knowledge as at the filing date. Where the correlation cannot be inferred, the applicant must present performance evaluation results (such as experimental results) for the AI model. Further, for product inventions whose specific function is estimated by AI (eg, chemical compounds or adhesive compositions), actual working examples or verified prediction accuracy are required.

Claim Drafting and Clarity

Claims must make clear that the invention is a program (a product) and that a computer executes it; otherwise, the clarity requirement is violated. Claims may also be rejected for lack of clarity where the co-operation between software and hardware resources is not apparent; in that case, the configuration of the neural network must be specified in detail in the claims. It is therefore advisable to describe the detailed configuration of the neural network in the specification at the time of filing, so that it can be recited in the claims where necessary. An expression such as “trained model” may be acceptable if it is clearly identified as a program that causes a computer to function; if this is not clear even by reference to the claims and the specification, the clarity requirement may not be satisfied.

Inventive Step

The assessment of inventive step (including its relationship with the prior art) is addressed in 6.6 Inventive Step, Prior Art and AI-Assisted R&D.

Summary

There is no special regime specific to AI-related inventions: they are assessed within the same framework as computer software-related inventions generally. What is distinctive is that operational points reflecting the characteristics of machine learning, such as whether the correlation between input and output data must be disclosed, have been set out through the JPO’s collection of examination cases.

As discussed in 2.1 Protectable Elements of AI Systems, to be protected as a trade secret, information must satisfy three requirements:

• secrecy management;
• usefulness; and
• non-public nature.

AI models, weights, datasets, prompts, system instructions and the like may be protected as trade secrets where these requirements are met. The requirement that most often causes difficulty in practice is secrecy management: it must be objectively recognisable that the information is managed as secret, through measures such as access restrictions, contractual confidentiality obligations and internal rules. Even where information is not protected as a trade secret, it may be protected as shared data with limited access if it satisfies the requirements of being provided to specific persons on a regular basis, accumulated to a significant extent and managed by electronic or magnetic means.

Complying with transparency and disclosure demands (such as those under the Principles-Code discussed in 10.1 Pending Legislation, Regulation and Policy) may, depending on the content and scope of the disclosure, undermine the non-public nature or secrecy management of a trade secret and result in the loss of protection. In practice, disclosures should be limited to summaries or reasonable explanations and designed so that core secrets, such as model weights, do not become publicly known.

Neither the trade secret regime nor the shared data with limited access regime is specific to AI and the protection available does not differ from that for confidential information and data generally.

Japanese law has no independent right equivalent to the EU’s sui generis database right. Protection for datasets and databases is provided mainly through the following frameworks; this does not differ from the protection of databases generally and there is no regime specific to AI training datasets:

• First, where the data themselves have creativity, they may be protected as works.
• Second, a database that has creativity by reason of the selection or systematic construction of information may be protected as a “database work” (Article 12-2 of the Copyright Act). The relevant creativity concerns the selection or systematic construction of information and does not require creativity in the content of the individual data. Unlike the EU database right, substantial investment in creating the database is not a requirement for protection.
• Third, protection may be available as a trade secret or as shared data with limited access under the Unfair Competition Prevention Act.

There is likewise no special regime for annotations, labels or synthetic datasets: their protection is assessed under the same frameworks (creativity, status as a trade secret or shared data with limited access and contract).

Article 30-4 of the Copyright Act permits a work to be exploited, in any way and to the extent considered necessary, where the purpose is not to personally enjoy or to cause another person to enjoy, the thoughts or sentiments expressed in the work (a “non-enjoyment purpose”). This applies whether or not the use is commercial and even where it involves providing to third parties a dataset containing works. Unlike the EU’s DSM Directive, there is no opt-out mechanism for commercial use. That said, such exploitation is not permitted where it would unreasonably prejudice the interests of the copyright owner in light of the nature or purpose of the work or the circumstances of its exploitation (the proviso to Article 30-4). Whether a use “unreasonably prejudices the interests of the copyright owner” is, in part, unsettled as a matter of interpretation.

Pending litigation is expected to reduce this uncertainty: major Japanese newspaper publishers have sued Perplexity – one suit brought by the Yomiuri Shimbun group and another brought jointly by the Asahi Shimbun and the Nikkei – and copyright infringement is among the issues in dispute.

These rules apply in the same way – that is, by asking whether the purpose is non-enjoyment – to pre-training, fine-tuning, RAG, evaluation and safety testing alike (for the output stage of RAG, Article 47-5 may come into play; see 3.2 Text and Data Mining and Other Exceptions).

Japanese law does not have a comprehensive exception equivalent to US fair use, but AI development may be justified under specific exceptions. The principal one is Article 30-4 of the Copyright Act, though Article 47-5(2) is also relevant.

Article 30-4

As discussed in 3.1 Use of Copyright Works for Training, Article 30-4 is the most important limitation on rights for AI training and development. Beyond use for information analysis (including text and data mining), it applies broadly so long as the purpose is non-enjoyment and no opt-out is available even for commercial use. Nor does it distinguish between research and product development or deployment: it applies to commercial product development as well. It also contains no lawful-access requirement, so the exception may apply even where access was unlawful. However, an AI developer or AI service provider may collect training data from a website knowing that the site hosts pirated or otherwise infringing copies. That knowledge increases the risk that the developer itself will be held to be the infringer where copyright infringement arises from use of the resulting generative AI.

Because Article 30-4 applies only to non-enjoyment purposes, it does not apply where an enjoyment purpose coexists. For example, it does not apply where, in additional training of an existing trained model, works are reproduced in order intentionally to cause the model to output all or part of the creative expression of works contained in the training data.

Article 47-5

Article 47-5(1) permits minor exploitation of a work, without the copyright owner’s authorisation, incidental to providing the results of computerised information analysis and the like (for example, an internet search engine displaying a short snippet). Article 47-5(2) permits exploitation of a work to the extent necessary as “preparation” for the minor exploitation under paragraph (1) and AI development may fall within this.

Other general exceptions, such as quotation (Article 32) and reproduction for private use (Article 30(1)), may also apply in particular settings, but the two provisions above operate centrally in the context of AI training and development.

At present, there are few publicly known, concrete examples of the licensing of AI training content and no established market practice appears to have formed. That said, the interim report of the “Study Group on Intellectual Property Rights in the AI Era”, set up by the Cabinet Office, highlights the importance of contractual solutions – that is, licence agreements between right holders and AI developers. In addition, the “Stakeholder Network on AI and Copyright”, jointly set up by the Agency for Cultural Affairs and METI, provides a forum for ongoing dialogue between right holders and AI developers to build smoother licensing practices.

Statutory licensing schemes for AI training, such as extended collective licensing or compulsory licensing, do not currently exist. For works whose right holders are unknown (orphan works), the ruling (compulsory licence) system administered by the Commissioner of the Agency for Cultural Affairs (Article 67 of the Copyright Act) is available, but it is not widely used in the AI training context. No statutory regime for data trusts exists.

Copyright law provides no EU-style opt-out mechanism. A technical measure such as robots.txt does not, by itself, have the legal effect of uniformly excluding Article 30-4. However, such a measure may be considered to have been implemented to avoid competition in the market for a database work. In other words, this refers to a database created by accumulating website data and organising it into a form usable for information analysis, then sold as a copyrighted work. In such a case, collecting data by circumventing the measure in these circumstances may be assessed as unreasonably prejudicing the interests of the copyright owner, which would mean that Article 30-4 does not apply. In addition, a separate contract may be formed – for example, by agreeing to terms of use or through API-based access. In that case, a claim may lie for breach of the relevant contractual restrictions (such as an anti-scraping clause), which is an issue distinct from copyright law; where an AI agent recognises such notices but acts contrary to them, that awareness may be taken into account in the assessment described above.

Metadata, content credentials (provenance information), website notices and licensing registers likewise have no independent legal effect of their own. They may be taken into account in assessing the Article 30-4 proviso and may serve as evidence of the formation of a contract or of a party’s awareness, but no more.

Currently, no law consistently requires documenting training data sources or similar information. That said, the Principles-Code discussed in 10.1 Pending Legislation, Regulation and Policy calls, from the standpoint of protecting IP and securing transparency, for disclosure of matters relating to the data used for training and validation and of information about crawlers. The Code follows a “comply or explain” approach: a business that accepts the Code must explain its reasons where it cannot comply. The Code is soft law and imposes no legal obligation.

As regards agentic systems, there is likewise no statutory provision uniformly requiring logs of tool calls, browsing, retrieval, API access or autonomous publication or deployment steps to be kept.

As to the law governing copyright infringement, injunction claims and damages claims are treated separately. Broadly, an injunction claim is governed by the law of the place of exploitation, while a damages claim is governed by the law of the place where the result occurred (Article 17 of the Act on General Rules for Application of Laws). Where datasets, servers, developers, model providers and users are located in different jurisdictions, identifying the place of copying or infringement and the governing law is a significant practical issue. Cross-border enforcement is discussed further in 8.4 Monetary Remedies and Cross-Border Enforcement.

Copyright infringement is the principal issue. To establish it, the claimant must generally prove:

• that the subject matter is a work;
• that the claimant is the copyright owner of that work;
• that the defendant’s material was created in reliance on the claimant’s existing work;
• that the defendant’s material is similar to the claimant’s existing work; and
• the act of exploitation (reproduction, adaptation, transmission to the public, etc).

The defendant may then plead, as a defence, that its exploitation was for a non-enjoyment purpose under the main clause of Article 30-4 of the Copyright Act. The claimant may in turn argue, by way of reply, that Article 30-4 does not apply because the exploitation would unreasonably prejudice the interests of the copyright owner (the proviso).

This framework applies equally to training, fine-tuning, evaluation, deployment and operation and acts carried out automatically by an AI agent or tool-using system are assessed under the same framework in relation to the party to whom the acts are attributed (the user, the business operator, etc; see 4.3 Secondary, Authorisation and Intermediary Liability).

The pending Perplexity litigation (see 3.1 Use of Copyright Works for Training) is being watched as a case that may yield rulings on the characterisation of acts of exploitation at the training and provision stages and on the interpretation of the Article 30-4 proviso.

Internal model artefacts such as weights and parameters are, in principle, no more than the statistical and numerical result of processing the training data and are not usually regarded as retaining the expression of an existing work in a form perceptible to humans. The scope for treating them as works or as reproductions or adaptations of existing works, is therefore limited. That said, where a model has strongly memorised particular training data such that its creative expression can be reproduced almost verbatim at the output stage, the question whether that output constitutes reproduction or adaptation may arise (see 5.1 Copyright Infringement in Outputs). As for minor exploitation under Article 47-5, see 3.2 Text and Data Mining and Other Exceptions.

To date, Japanese courts or regulators have only made few, if any, decisions that squarely assess memorisation, substantial similarity or de minimis copying in the AI context.

Where the generation or use of AI output amounts to copyright infringement, the user, as the physical actor, is in principle liable as the infringer. However, under the so-called normative-actor doctrine – under which a party other than the physical actor may, in law, be treated as the infringer – a business that develops generative AI or provides a service using it may be assessed as the actor responsible for the infringement and held liable. Broadly, the following factors are considered:

• Where a particular generative AI frequently generates infringing material, the business is more likely to be assessed as the infringing actor.
• The same is true where the business was aware of a high probability that material similar to existing works would be generated but took no measures to deter such generation.
• Conversely, where the business has taken technical measures to deter such generation, it is less likely to be assessed as the infringing actor.
• Where infringing material is not generated with high frequency, the business is unlikely to be assessed as the infringing actor, even if a user inputs prompts intending to generate material similar to an existing work and infringing material results.

Japan has no safe harbour that comprehensively exempts model providers from liability for copyright infringement by AI output.

As this shows, liability is a fact-specific, case-by-case assessment and little case law has established settled criteria.

Where information qualifies as a trade secret or as shared data with limited access, injunctions and damages are available in respect of the following categories of act.

Trade Secrets (Article 2(1)(iv)-(x) of the UCPA)

The prohibited acts broadly fall into three patterns, as outlined below.

Improper acquisition

Acquiring a trade secret by theft, fraud, duress or other wrongful means and then using or disclosing it; and acquiring, using or disclosing a trade secret while knowing of (or being grossly negligent in not knowing of) an intervening wrongful acquisition, including where this comes to light after acquisition;

Breach of trust

A person to whom the holder lawfully disclosed the secret using or disclosing it for wrongful gain or causing damage to the holder; and downstream acquisition, use or disclosure with knowledge (or gross negligence) of an intervening wrongful disclosure; and

Infringing goods

Assigning, exporting or importing goods produced by the use of another’s technical secret.

Shared Data with Limited Access (Article 2(1)(xi)-(xvi) of the UCPA)

The protection requirements are that the data be provided to specific persons on a regular basis, accumulated to a significant extent and managed by electronic or magnetic means (trade secrets are excluded). The prohibited acts broadly correspond to the trade-secret patterns above, but are narrower in two respects: only actual knowledge is caught for downstream acquirers (gross negligence does not suffice) and where the wrongful origin becomes known only after acquisition, only disclosure is caught (disclosure within the scope of the authority acquired is excluded). The categories are as follows.

• improper acquisition: acquiring the data by theft, fraud, duress or other wrongful means and using or disclosing them and downstream acquisition, use or disclosure with knowledge of an intervening wrongful acquisition;
• breach of trust: a person to whom the holder provided the data using them (limited to use in breach of their data-management duty) or disclosing them, in either case for wrongful gain or to cause damage to the holder; and
• downstream acquirers: acquiring, using or disclosing the data with knowledge of such a wrongful disclosure.

If a model reproduces confidential information and discloses it to third parties, the secrecy management and non-public nature (for a trade secret) or the limited-access character (for shared data with limited access) may be lost, so that protection ceases. Where a model infers confidential information from fragmentary data and outputs it, the use or disclosure of that output may likewise raise questions of prohibited acts under the UCPA or breach of contract, to be assessed case by case.

Separately, where a confidentiality agreement is in place, damages and other remedies may be sought under that contract. To date, no court decisions or pending cases squarely assess these categories in the AI context.

Against a claim of copyright infringement, an AI developer or model provider will principally invoke a limitation on rights, such as Article 30-4 of the Copyright Act (see 3.1 Use of Copyright Works for Training). Other possible defences include that the subject matter is not a work, that there was no reliance (an independent-creation argument; see 5.1 Copyright Infringement in Outputs), that there is no substantial similarity and that an implied or express licence was obtained from the right holder.

Procedurally, a damages claim under copyright law is subject to the general extinctive prescription of the Civil Code: a tort claim is time-barred three years after the injured party learns of the damage and the identity of the perpetrator and in any event twenty years after the act. For trade-secret infringement, the UCPA’s prescription and long-stop periods may also be relevant: an injunction claim against ongoing wrongful use is barred three years after the holder learns of the continuing act and in any event twenty years after the act began.

Exhaustion does not usually arise in the context of AI development or provision. There is no general public-interest defence and competition-law (Anti-Monopoly Act) arguments operate as a defence to copyright infringement only in limited situations. A defence that the right holder’s claim is an abuse of rights (Article 1(3) of the Civil Code) may be raised, but succeeds only in limited cases.

As with the conventional framework for copyright infringement, two requirements are at issue when assessing infringement by AI output: reliance and similarity. Reliance is affirmed where the AI user is found to have been aware of the existing work. When the user inputs an existing work as a prompt or reference image and causes the AI to generate output containing its creative expression, the user is usually aware of the work, so reliance is readily found.

Even where the user was unaware of the work, reliance may still be found. If the work was included in the generative AI’s training data and was learned at the development/training stage, objective access to the work is taken to have existed. Accordingly, where that AI is used to generate material similar to the work, reliance is ordinarily presumed and the user may be liable for copyright infringement. However, the user may rebut this presumption. Where it can be assessed, as a legal matter, that the creative expression of the learned work is not in a state capable of being output at the generation/use stage, the user may plead facts supporting that assessment and reliance may then be negated even though the work was learned in training.

Output that merely imitates a style, genre or technique, without similarity in concrete expression, does not in principle constitute copyright infringement, under the idea/expression dichotomy.

As noted in 5.1 Copyright Infringement in Outputs, an AI user may infringe an existing copyright unwittingly.

A claim to enjoin the use is a no-fault remedy, so intent, knowledge, prompt design and human review do not affect the outcome.

By contrast, a damages claim requires intent or negligence. Accordingly, where the user did not know of the existing work and was not negligent in failing to know of it, liability in damages may be denied. Whether there was human review, the design of the prompt (in particular, whether it gave specific instructions aimed at imitating an existing work) and reliance on the provider’s terms of use or disclaimers are among the circumstances that may be weighed in assessing negligence. The scale of use (the extent of commercial use and the manner of publication or distribution) may also affect the assessment of negligence and the quantum of damages.

Where an infringing act is carried out as a result of configuring or deploying an agentic system, the basic framework is the same and the design and supervision involved in that configuration or deployment may be taken into account in assessing negligence (see 4.3 Secondary, Authorisation and Intermediary Liability and 5.5 Agentic AI, Tool Use and Autonomous Acts).

In Japan, there is no independent common-law tort equivalent to the English and US doctrine of passing off. Protection corresponding to passing off is instead provided by the statutory provisions of the Unfair Competition Prevention Act (the “UCPA”) and “unfair competition” under Japanese law is limited to the acts exhaustively enumerated in Article 2(1) of the UCPA.

Where AI output contains a sign identical or similar to a registered trademark and the user uses it in the course of trade for goods or services identical or similar to the designated goods or services, trademark infringement can arise. Even unregistered signs may be protected: well-known or famous indications of goods or business and product configurations, are protected under the UCPA as acts causing confusion (Article 2(1)(i)), misappropriation of famous indications (Article 2(1)(ii)) and imitation of product configuration (Article 2(1)(iii)), among others.

Within this framework, the Japanese government has made clear that the existing legal principles apply equally to AI-generated outputs. The Cabinet Office has explained that trademark infringement and unfair competition involving AI-generated outputs are assessed in the same manner as conventional infringement, because:

• reliance is not a required element of trademark infringement or unfair competition; and
• AI outputs present no AI-specific considerations in determining similarity.

The absence of a reliance requirement is a fundamental distinction from copyright infringement, where similarity and reliance must both generally be established and it is significant in practice: even where an AI system happens to have learned and reproduced a third party’s trademark or other protected sign, infringement may be found if an objectively identical or similar sign is used in the course of trade, regardless of reliance or intent to copy.

Furthermore, if output is used in advertising in a way that falsely suggests an endorsement by referencing a celebrity’s name, likeness or voice, there may be liability for infringing the right of publicity, as established in case law or general tort principles. Advertising regulation may also apply, including the Premiums and Representations Act (misleading representations as to quality or terms) and the UCPA’s prohibition on misleading indications (Article 2(1)(xx)).

Regardless of whether the allegedly infringing product was generated (or the allegedly infringing process used) by an AI, the AI user’s manufacture, use or other exploitation of the product or process can constitute direct infringement. Even where the AI user did not intend to infringe, the user will be liable for direct infringement if the AI-generated product falls within the scope of the patent or design right.

Indirect infringement may also be established against an AI developer in respect of the provision of an AI tool. To take patent law as an example: where the output of an AI drug-discovery tool is indispensable to solving the problem addressed by the invention and the developer knew that the output would be used in working the invention – that is, where there actually exists a high probability that a non-exceptional segment of the tool’s users will use it to infringe the patent and the developer recognised and accepted this – the provision of the AI tool may constitute indirect infringement (Article 101(ii) of the Patent Act, among others).

The normative-actor doctrine discussed in 4.3 Secondary, Authorisation and Intermediary Liability is the central framework for assessing liability for an AI agent’s autonomous acts. Even where it is the agent itself that performs the physical act (scraping, generation, publication and so on), the business that develops or provides the agent or the user that configures or deploys it, bears liability as the infringer where, based on a normative view, it is assessed as the actor responsible for that act. At present there is no case law or settled test specific to agentic AI and the matter will be decided case by case by applying the existing normative-actor framework.

Whether AI output is a work is decided case by case for each item of output. It depends not on mere effort but on an overall assessment of the extent to which contributions that can be regarded as creative have accumulated. The main factors that may be weighed are:

• the volume and content of the instructions/input (such as prompts);
• the number of generation attempts; and
• selection among multiple outputs.

If these factors indicate that human creative input is present in the output’s concrete expression, the output can be recognised as a work. On the other hand, if only a simple instruction was given and the result was used without further input, the level of creative contribution is likely to be minimal and copyright protection is likely to be denied.

Prompt chaining or decomposition and autonomous task planning in agentic systems are assessed under the same standard, namely the existence and degree of ultimate human creative involvement. No distinct treatment specific to these techniques has yet been indicated.

Under Japanese copyright law, a “work” is defined as “a creative expression of thoughts or sentiments” (Article 2(1)(i)), which requires creative expression of human thoughts or sentiments. Accordingly, material generated autonomously by AI, with no recognisable human creative contribution, is not a work and receives no copyright protection. Nor is there any special provision for computer-generated works, as in the United Kingdom. Such material is, in principle, in the public domain and freely usable by anyone.

The unresolved issue is the precise point at which human involvement is sufficient for copyright to arise. Future case law will clarify the necessary degree and manner of human input for copyright protection, as discussed in 6.1 Human Authorship and Copyright Protection.

Where several people jointly and creatively contribute to preparing prompts or to selecting and editing the output, the result may, depending on the nature and degree of their involvement, be a joint work (Article 2(1)(xii) of the Copyright Act: a work whose contributions cannot be exploited separately), with copyright shared among the contributors. The co-owners’ shares are determined by their agreement or, absent agreement, by the degree of each person’s creative contribution; where the shares are unclear, they are presumed equal (see Article 250 of the Civil Code). Exercising jointly owned copyright in principle requires the consent of all co-owners (Article 65(2) of the Copyright Act).

Where AI output can be regarded as adding new creative expression to an existing work used as material, it may be a derivative work (Article 2(1)(xi)). In that case, the copyright owner of the original work must authorise it (the adaptation right, Article 27). In addition, the author of a derivative work can obtain separate rights in that work.

Whether the human involvement in the generation process amounts to the addition of creative expression is determined case by case by reference to the standard in 6.1 Human Authorship and Copyright Protection.

Currently, Japanese IP law does not require disclosure of AI involvement.

As regards patents and design rights, as long as a natural person has led the invention (or the design), the applicant need not disclose that AI assistance was used. However, naming an AI system as an inventor or co-inventor is not permitted (the DABUS case: IP High Court, 30 January 2025).

As regards copyright, disclosure of AI involvement is likewise not required for copyright registration. However, copyright protection extends only to creative expression originating from humans.

Under Japanese law, an AI system cannot be named as an inventor or co-inventor (the DABUS case: IP High Court, 30 January 2025). Japan does not yet have guidance equivalent to the USPTO’s Revised Inventorship Guidance for AI-Assisted Inventions; however, discussions, including potential legislative amendments, are underway regarding how to determine inventorship for AI-related inventions.

In light of the case law on the determination of inventorship for inventions in general, a substantial creative contribution to the completion of the technical idea is considered necessary for a person to be recognised as the inventor. For AI-related inventions, formulating the specific technical problem, establishing correlations within the training data through data pre-processing and similar steps and evaluating and validating the AI’s raw output to ensure its practical applicability would be such a “substantial creative contribution”. Even where an AI agent proposes experiments, selects parameters or generates candidate solutions, inventorship is determined by the existence and degree of such human involvement.

The right to obtain a patent initially vests only in the human inventor. Where an employee uses AI tools, Article 35 of the Patent Act (employee inventions) applies: through work rules and the like, the right to obtain a patent can be made to vest originally in or be transferred to, the employer, provided the employee receives reasonable benefits. For contractors, independent researchers, external collaborators or AI developers, entitlement is governed by the specific terms of the relevant joint development or licence agreements.

In Japanese patent law, the rapid advancement and spread of AI are fundamentally affecting the assessment of inventive step and the notional “person skilled in the art”.

A person skilled in the art is a person who has ordinary knowledge in the technical field to which the invention pertains (see Article 29(2) of the Patent Act). The definition itself is unchanged. As the use of AI expands, however, the person skilled in the art will be taken to be someone able to use AI as a routine technical means for research and development. As the capabilities of standard AI improve rapidly, technology that can be created using standard AI will lack an inventive step and the inventive-step hurdle may become higher than before.

Publicly available AI systems and AI-generated material form part of the state of the art. Accordingly, simply applying publicly known AI technology to another field or replacing existing non-AI technology with AI, without technical difficulty, is regarded as an exercise of ordinary creative ability and is therefore likely to be found to lack an inventive step. For an inventive step to be recognised, the invention must produce an unexpected, outstanding effect beyond the state of the art; for example, by:

• setting specific technical conditions; or
• unique characteristics in the:
• training method;
• the parameters;
• the type of dataset; or
• the data pre-processing.
• Merely using different data content is insufficient.

The Patent Act does not expressly address synthetic datasets or automatically generated technical documents. However, all information that has become publicly available constitutes prior art (Article 29(1)), so if such materials were made available to the public before the filing date, they constitute prior art.

AI-generated designs and product appearance are protected through overlapping regimes under the Design Act, the Unfair Competition Prevention Act and the Copyright Act.

AI-generated GUIs and icons may be registered as “image designs” (operation images and display images), introduced by the 2019 amendment to the Design Act. The requirements for registration are industrial applicability, novelty and creative difficulty (Article 3 of the Design Act). Designs consisting solely of shapes indispensable for securing a function are excluded from protection (Article 5(iii)). The creator of a design must be a natural person; AI cannot be recognised as a creator. For employee-created designs, Article 35 of the Patent Act applies mutatis mutandis, so the right to obtain a design registration can be made to vest in the employer through work rules and the like (Article 15(3) of the Design Act). The term of protection is 25 years from the filing date.

Avatars and virtual goods, as content, are excluded from image designs, leaving a gap in protection. An amendment to the Design Act addressing this is under consideration but has not been enacted as at the time of writing (the position should be re-checked before publication).

Unregistered appearances and trade dress may be protected as three-dimensional trademarks under the Trademark Act and under Article 2(1)(i) to (iii) of the Unfair Competition Prevention Act (acts causing confusion, misappropriation of famous indications and imitation of product configuration).

Whether a mark has been generated by AI does not affect its registrability. A trademark is not a creative work but a sign selected for use in commerce. Accordingly, it may be registered provided that it possesses distinctiveness under Article 3 of the Trademark Act and does not fall within any of the grounds for refusal under Article 4. The 2014 amendment to the Trademark Act expanded the scope of protectable subject matter to include non-traditional trademarks, such as sound marks, motion marks, hologram marks, marks consisting solely of colour and position marks (Article 2(1)). A slogan may be filed as a word mark; however, if it lacks distinctiveness, it may be refused registration under Article 3(1)(vi).

As to ownership, trademark law does not recognise the concepts of an inventor or a creator. The owner is the natural person or legal entity that actually selects the mark and files the application for registration. Under Japan’s first-to-file system (Article 8), the first applicant is entitled to registration.

An author’s moral rights (the right of attribution, the right to integrity, etc; Articles 18 to 20 of the Copyright Act) belong exclusively to the author who created the work. Where AI output is not a work, no question of moral rights therefore arises. That said, where AI output can be regarded as an adaptation of an existing work, its relationship with the moral rights of the author of the original work (such as the right to integrity) may be in issue.

Where output imitating the style of a living creator is falsely presented or advertised under that creator’s real name or as their genuine work, this may attract criminal liability for false attribution of authorship (Article 121 of the Copyright Act) and civil liability for infringement of the person’s name and reputation. It may also constitute a misleading act under the Unfair Competition Prevention Act (Article 2(1)(xx)) or a general tort.

Synthetic performances imitating the voice or performance of a real performer raise issues of performers’ moral rights (attribution and integrity), other performers’ rights and the right of publicity (see 7.4 Personality, Publicity, Performers’ and Neighbouring Rights).

A person’s name, likeness, voice, performance or digital replica is protected not by a single regime but by a combination of doctrines:

• the right of publicity established in case law – the exclusive right to exploit the economic value (the customer-drawing power) of a name, likeness and the like. It derives from personality rights but also has a proprietary character in that it protects economic interests; it is not codified as an IP right;
• personality rights, such as the right to one’s likeness and the right to one’s name – protection of the personal (privacy) interest in not having one’s likeness and the like used without cause;
• for performers, performers’ moral rights (attribution and integrity) and neighbouring rights such as the rights of sound recording and visual recording under the Copyright Act – protection codified as IP; and
• the Unfair Competition Prevention Act – protection, where a name or likeness functions as a well-known or famous indication of goods or business, against acts causing confusion and the misappropriation of famous indications.

Japanese law thus treats this protection as a composite spanning IP (neighbouring rights), personality/privacy rights, judge-made doctrine (the right of publicity) and unfair competition law. The report of the Ministry of Justice’s Study Group on Civil Liability for the Unauthorised Use of Likenesses, Voices and Similar Attributes, published in August 2026 as interpretive guidance on infringement of the right of publicity and related rights by generative AI, expressly confirms that a person’s voice falls within the scope of protection of the right of publicity and of the right not to have one’s likeness and similar attributes used without authorisation.

AI and IP disputes are brought before the ordinary civil courts and the procedure is no different from a general IP claim.

As to jurisdiction, actions concerning patents, utility models, circuit layouts and the author’s rights in works of computer programming fall within the exclusive jurisdiction of the Tokyo District Court (for eastern Japan) or the Osaka District Court (for western Japan) (patent-type exclusive jurisdiction: Article 6 of the Code of Civil Procedure). Actions concerning design rights, trademarks and copyrights in works other than computer programs may be filed with the ordinarily competent court or, additionally, with the Tokyo or Osaka District Court (Article 6-2). Appeals in the patent-type exclusive-jurisdiction cases are heard by the IP High Court; in other IP cases, appeals go to the competent high court.

Trade secret actions under the Unfair Competition Prevention Act are not subject to the exclusive-jurisdiction rule above, although in practice, technically complex cases are often heard by the specialised IP divisions.

In Japan, parties to civil litigation can obtain AI-related evidence (such as training data, source code and system logs (including electronic records such as an agent’s tool-call histories, browsing records and API logs)) through mechanisms including document production orders under the Code of Civil Procedure (CCP) and the Patent Act.

First, on a prima facie showing that the documents are necessary to prove infringement, the claimant can ask the court to order the defendant to produce the documents needed to prove infringement (and to calculate damages) under a document production order (Article 220 of the CCP) or an order to produce documents (Article 105 of the Patent Act).

Second, in patent infringement litigation, where in addition to the necessity of the evidence, the probability of infringement, subsidiarity and appropriateness are established, the claimant can ask the court to use the inspection (Sasho) system under Article 105-2 of the Patent Act. Under this system, a court-appointed expert (the inspector) may enter the defendant’s offices or premises, inspect servers and systems and prepare and submit an inspection report on the existence of source code and other material necessary to prove infringement.

In these disclosure procedures, the defendant can seek an in camera procedure (Article 223(6) of the CCP) to prevent the disclosure itself or a protective order (Article 105-4 of the Patent Act) to restrict the opposing party’s use once disclosed.

In Japan, AI-related IP disputes use both interim injunctions (preservative measures) under the Civil Preservation Act and final injunctions under individual IP statutes, such as the Copyright Act, the Patent Act and the Unfair Competition Prevention Act. Right holders may demand the cessation or prevention of infringement, as well as the destruction of objects constituting the infringing act and the removal of equipment used for the infringement (Article 112 of the Copyright Act and equivalents).

Where recognised as “acts necessary to prevent infringement” orders to stop training or to remove works from datasets may be available. Whether courts will go further (disabling outputs, requiring filters or guardrails, suspending agents or disabling autonomous functions) is not yet settled by case law.

As to whether a trained model may be deleted or quarantined for copyright infringement: because a trained model is not itself usually regarded as a direct reproduction of the works, such orders are ordinarily not available. However, they may be available where the model generates infringing material with high probability and can therefore be characterised as an “object constituting the infringing act”. A right holder may also be able to obtain removal of its works from datasets in order to prevent subsequent model versions from exploiting them.

As to evidence preservation, the evidence preservation procedure under the Code of Civil Procedure (Article 234) is available.

Monetary damages are generally calculated on the basis of lost profits or a reasonable royalty. Statutory provisions assist in establishing damages: the amount may be calculated from the infringer’s sales volumes and the profits gained by the infringer are presumed to be the right holder’s damages (Articles 102(1) and (2) of the Patent Act; Articles 114(1) and (2) of the Copyright Act, among others). A right holder may instead claim an amount equivalent to a reasonable royalty (Article 102(3) of the Patent Act; Article 114(3) of the Copyright Act). A claim in unjust enrichment (Articles 703 and 704 of the Civil Code) is also available as an alternative. Japan does not recognise US-style punitive damages or statutory damages.

As to non-monetary remedies, as discussed in 8.3 Interim and Final Injunctive Relief, deletion of training data may be available as part of an injunction.

As to territoriality, in cases of cross-border infringement via the internet or servers located abroad, a flexible reading of the territoriality principle has become influential (the Dwango case: the Supreme Court judgment of 3 March 2025). On this approach, even where the infringing act is not formally completed within Japan, Japanese law should apply so long as the substantial place of working of the invention is found to be Japan.

The governing law is analysed separately for damages and injunctions. A damages claim is characterised as a tort, so the law of the place where the result occurred applies in principle; an injunction claim is governed by the law of the country for which protection is sought (the country where the right was established). In practice, the protecting country (the place of exploitation) and the place where the result occurred usually coincide, so the same country’s law tends to govern both claims; where the place of exploitation of a work is Japan, the Japanese Copyright Act applies in principle.

As to the enforcement of foreign judgments and orders, Japanese courts recognise and enforce foreign judgments only where the domestic statutory requirements, including reciprocity and consistency with public policy (Article 118 of the Code of Civil Procedure; Article 24 of the Civil Execution Act), are satisfied.

In a licence for AI training content, at a minimum the following should be made clear:

• the scope of the licensed content and the purpose of use (distinguishing training, fine-tuning, evaluation, RAG and so on);
• the models and services permitted to use the content and whether the licence is exclusive;
• the consideration and how it is calculated (lump sum, usage-based fees and so on);
• whether attribution (credit) is required;
• whether there is an audit right and its scope;
• the licensor’s rights to suspend use (opt-out) and to require deletion;
• whether sublicensing is permitted;
• the rights position as to outputs and downstream uses; and
• where an agentic system is used, log retention, the scope of tool-use permissions and restrictions on autonomous access and publication.

As noted in 3.3 Licensing of Training Content, no established market practice has yet formed on these points, so the parties must design them individually through negotiation.

The Principles-Code published by the Cabinet Office in August 2026 is a notable piece of soft law addressing the relationship between generative AI and IP.

The Code applies to generative AI developers and generative AI providers. Generative AI providers include those who provide the public with a service that incorporates a generative AI system into an application, product, existing system or business process. Notably, a business offering, for example, a consumer-facing chatbot service may therefore fall within scope.

The Code adopts a “comply or explain” approach and sets out three principles:

• first, a disclosure principle, requiring transparency as to the models used, the training data (including information about crawlers) and similar matters, and disclosure of the status of measures taken to protect IP rights;
• second, a principle of responding to disclosure requests (inquiries) from right holders who are pursuing or preparing to pursue litigation, mediation, ADR or other legal proceedings, by answering whether the URLs or similar information identifying their works are contained in the data used for training and validation; and
• third, a principle of responding to inquiries from users of generative AI, by answering whether the URLs or similar information identifying content identical or similar to the user’s output are contained in the data used for training and validation.

The Code has no legally binding force and compliance with each principle is voluntary. However, businesses accepting the Code are expected to publish, on their corporate websites or equivalent, an acceptance declaration and the status of implementation of each principle (with reasons for any principle not implemented) and to submit a notification in the prescribed format to the Cabinet Office (Intellectual Property Strategy Promotion Office), which publishes a list of the businesses that have filed. In this sense, although it is non-binding soft law, it carries practical compliance pressure. The practical impact differs by stakeholder: more effective inquiry handling for right holders; disclosure and governance costs for AI developers and providers; and greater transparency about the rights position of AI output for AI users.

Japan leads the Hiroshima AI Process (HAIP) and, working with the G7 and international bodies such as the OECD, is actively engaged in developing international principles and codes of conduct on AI governance. Japan also participates in the WIPO Conversation on Intellectual Property and Frontier Technologies (formerly the Conversation on IP and AI), where AI remains a central theme. By contrast, there are at present no notable initiatives specific to AI and IP within standards organisations or bilateral frameworks.

On copyright limitations, Japan was among the first in the world to enact a broad text and data mining exception (Article 30-4 of the Copyright Act) and this breadth is a principal point of divergence from other major jurisdictions. In addition, as noted in 10.1 Pending Legislation, Regulation and Policy, Japan is seeking to regulate through the Principles-Code, a distinctively Japanese form of soft law that also differs from approaches taken elsewhere.