At the time of writing, there is no AI-specific IP legislation, regulations or official guidance governing AI-generated outputs, AI-assisted inventions, training data, text and data mining, or model transparency in Malaysia.

The Government issued the National Guidelines on Artificial Intelligence and Ethics (AIGE) in 2024. Recently, the Ministry of Digital has also issued a public consultation paper on the proposal for an Artificial Intelligence Governance Bill (“AI Governance Bill”) to be considered by Parliament. However, both the above initiatives do not seem to deal with the IP-related aspects of AI.

Malaysia is a party to most of the major international IP treaties and conventions and to various free trade agreements which could contain provisions on how IP is recognised and protected bilaterally or multilaterally. However, we do not know of any specific AI-IP related provisions in any of the treaties/agreements to which Malaysia is a party.

Malaysian courts do regularly acknowledge and take cognisance of foreign judgments, especially from common-law/Commonwealth countries, though generally such judgments are not binding. Hence, since it appears that no AI-IP related cases have yet been heard in the Malaysian courts, it is very likely that foreign judgments on such cases, if relevant by having similar facts, will be considered by the Malaysian courts in making their decisions.

Foreign right-holders and foreign AI developers would generally receive equal treatment under the law as with national right-holders/AI developers and there are no special rules applicable to foreign right-holders/AI developers.

There are presently no statutory definitions of AI-related terminologies in Malaysia.

That said, the AIGE and the public consultation paper for the AI Governance Bill do provide guidance on the definitions of some AI-related terminologies such as AI, AI system and generative AI as provided below.

Definition of AI:

• In the public consultation paper for the AI Governance Bill, AI is defined as a functional capability or a set of methods that enables a system to perform functions that resemble human cognitive functions which include learning, reasoning, inference, prediction, perception, language processing, recommendation, content generation or decision support.
• In the AIGE, AI is defined as the field of computer science that seeks to create intelligent machines that can replicate or exceed human intelligence.

Definition of AI system:

• In the public consultation paper for the AI Governance Bill, an AI system is defined as the actual operative system, whether stand-alone, embedded, integrated, modular, remote or service-based, through which AI functionality is manifested in practice. It is a system that incorporates, applies or is enabled by AI capability.

Definition of generative AI:

• In the AIGE, generative AI is described as models of algorithms that create new outputs, such as text, video, photo, code, data and statistics, from huge amounts of data on which those models are trained.

Although the above definitions are rather general, we are of the view that when legislation is actually passed concerning AI and IP, there will be a need to clearly define the various terms related to AI for purposes of the legislative provisions. In this regard, there will probably be some form of standardisation of the terminology.

In Malaysia, the Ministry of Digital has been given the general responsibility to oversee the use and development of AI by the country at large. In fact, Malaysia has a National AI Office which is under this Ministry. However, it should be noted that this Ministry does not handle national IP issues, which are technically under the Ministry of Domestic Trade and Cost of Living (KPDN).

The main regulator for IP administration in the country is the Intellectual Property Corporation of Malaysia (MyIPO), which is an agency under KPDN. In terms of AI, MyIPO has recognised its intersection with IP in its current efforts to update the Copyright Act 1987 (“CA 1987”). MyIPO has requested for public consultations to consider what amendments may be required for the CA 1987 including, among others, how to deal with issues concerning AI.

Currently, the Controller of Copyright (who is under MyIPO) requires an applicant for voluntary notification of copyright to declare that the work that is the subject matter of the application was not created with the assistance of AI, failing which the application will not be approved.

In terms of the court system, a specialist IP High Court has been established within the Commercial Division of the High Court at Kuala Lumpur to deal with IP disputes, although it should be noted that this IP High Court does not have exclusive jurisdiction over IP matters and IP litigants may choose to have their case heard in any relevant High Court within the country.

We believe that most elements of an AI system would, to a certain extent, be protectable by the relevant IP or related rights, subject to the relevant criteria and conditions as stipulated by the relevant IP legislation or laws, such as the laws relating to copyright, patents, industrial designs, confidentiality and trade secrets.

As an AI system may be seen to comprise mainly of a network of computer programs, the CA 1987 would be the most applicable IP legislation which protects the AI system.

A computer program is defined in the CA 1987 as an expression, in any language, code or notation, of a set of instructions (whether with or without related information) intended to cause a device with information processing capability to perform a particular function, either directly or after either or both of the following:

• conversion to another language, code or notation;
• reproduction in a different material form.

A computer program is protected as a literary work under the CA 1987, provided that the requirements of originality (with sufficient human effort) and reduction into material form are satisfied.

We believe that most elements of an AI system would qualify as a computer program, as defined above. For those elements which may not qualify as a computer program, there would be a need to consider what form of copyright protection could be accorded to those elements. For example, a prompt per se, which is obviously not a computer program, would arguably also not be protected as another form of literary work. However, a prompt library would arguably satisfy another category of protected literary work known as “compilation”. Another example would be that of outputs of intermediate occupational processes within the AI system, as such processes may have the issue of not being in “material form” insofar as they may be of a transient nature.

Finally, it should be noted that the CA 1987 expressly excludes from copyright protection “ideas, procedures, methods of operation or mathematical concepts as such”, and thus these elements in the AI system would not be protected under copyright.

An AI-related invention may be patentable where it falls within the meaning of “invention” under the Patents Act 1983 (“PA 1983”), namely an idea of an inventor that permits, in practice, the solution of a specific problem in the field of AI technology. This may include an invention contributing to an AI system or model that achieves a specific technical function within the AI system or model, provided that the requirements of patentability are satisfied, including novelty, inventive step and industrial applicability.

However, it should be noted that patent protection does not extend to excluded subject matter, including discoveries, scientific theories and mathematical methods; certain biological subject matter and processes; schemes, rules or methods for doing business, performing purely mental acts or playing games; and methods relating to the treatment or diagnosis of the human or animal body. Since an AI system or model largely consists of a network of computer programs and algorithms, care must be taken to ensure that the invention being contemplated for patent protection does not fall within any of the above excluded subject matter, in particular relating to a “mathematical method” or a “scheme, rule or method for doing business”.

AI-related materials, including AI models, weights, datasets, prompts, system instructions, evaluation data, deployment know-how and agent-related materials such as workflows, tool permissions and memory stores, may generally be protected as trade secrets or confidential information, and the nature of protection for AI-related confidential information/trade secrets is no different from that for other types of confidential information/trade secrets. To ensure that such protection is accorded, the AI model creator needs to maintain the confidentiality and secrecy of the subject matter by having reasonable safeguarding measures, such as access controls and, where possible, entering into non-disclosure agreements, as otherwise an argument may be raised that the relevant subject matter was not confidential or secret by nature.

It should be noted that should the subject matter sought to be protected enter the public domain, with or without the consent or authorisation of the proprietor of the confidential information/trade secrets, such subject matter would no longer be protected as confidential information/trade secrets, without prejudice to the rights of the proprietor to sue the party that may have breached its duty of confidence. In certain instances, the disclosure of the confidential information/trade secrets to relevant authorities may be required by law, and occasionally, a public disclosure may be required. In such a case, the proprietor of the confidential information/trade secrets would not be able to maintain its rights of confidentiality.

There is no specific IP legislation protecting databases per se. Like in other countries, Malaysia does have a law relating to the protection of personal data, namely the Personal Data Protection Act 2010 (“PDPA 2010”). The PDPA 2010 protects databases which contain personal data by imposing statutory obligations on the person who stores and processes such personal data (ie, the data controller).

Accordingly, the CA 1987 does not protect data or databases per se but does protect certain compilations of data as a form of literary work, namely where there was sufficient effort expended to make the compilation original in character (ie, selection or arrangement of the contents). In this regard, it is arguable whether, if data is being compiled in an AI system, such compilation would meet the said test for originality under the CA 1987.

Notwithstanding the absence of any IP legislation protecting data or databases per se, Malaysian law does protect data or databases when the same are considered confidential information or trade secrets based on the principles of common law and equity. The relevant principles in this regard have already been explained in 2.4 Trade Secrets and Confidential Information.

There are currently no AI-specific legislative provisions governing scraping, ingestion, tokenisation, training, fine-tuning, evaluation or safety testing. Under the CA 1987, the key copyright issue in AI training is whether a copyright work, or a substantial part of it, is being reproduced during or as a result of the training without the copyright owner’s authorisation.

Malaysian copyright law focuses on the result of an activity rather than the technology used. Accordingly, the use, analysis or processing of information is not itself infringing unless it involves reproduction or another act restricted by copyright. Whether AI training infringes copyright therefore depends on the nature and extent of copying involved.

A potentially relevant exception is Section 13(2)(q) of the CA 1987, which permits transient and incidental electronic copies required for the utilisation of works made available on a network. However, it remains uncertain, when copies are made within the AI system, whether such copies created during AI training, fine-tuning, inference, retrieval, evaluation or safety testing satisfy the requirements of being both “transient” and “incidental” so as not to be regarded as infringing copies.

At the time of writing, there are no reported Malaysian cases or legislative proposals specifically addressing AI training, fine-tuning, retrieval-augmented generation or related activities. These issues therefore remain governed by general copyright principles and are largely untested before the Malaysian courts.

Generally, if the AI training does not involve acts controlled by copyright, or if the training falls within a recognised exception, no copyright infringement arises. The exception of a transient and incidental copy of the work is discussed in 3.1 Use of Copyright Works for Training.

Malaysia does not currently have an express text and data mining (TDM) exception as in the first place, databases per se are not protected under the CA 1987.

The other exception most likely to be relevant is fair dealing under Section 13(2)(a) of the CA 1987 for the purposes of research, private study, criticism, review and news reporting, subject to the fairness factors in Section 13(2A), namely the purpose and character of the dealing, the nature of the work, the amount and substantiality used, and the effect on the market for the work. The above provisions obviously pre-date generative AI and have not been judicially considered in the Malaysian courts in the context of AI training or model development. However, we would find it difficult to argue that AI models come under fair dealing by being restricted to or limited by the above purposes.

The use of copyright works for AI training is generally governed by copyright and contract law rather than any AI-specific regime.

In practice, content is most likely to be licensed through voluntary commercial arrangements between right-holders and AI developers, including voluntary licences, and collective licensing.

Collective licensing is regulated by the CA 1987, which gives power to MyIPO to approve a corporate entity as a collective management organisation (CMO), which may thus issue licensing schemes for copyright works of the CMO’s members. The CA 1987 also establishes a Copyright Tribunal, which has a supervisory and quasi-judicial role in overseeing licensing schemes operated by CMOs, including resolving disputes concerning licence refusals, remuneration, unreasonable terms and licence expiry.

Malaysia also has no licensing regime or dedicated framework specifically for orphan works, out-of-commerce works or data trusts, although it is noted that MyIPO is currently contemplating provisions for orphan works in its current initiative in updating the CA 1987.

Malaysia does not currently recognise a specific statutory regime for AI training opt-outs, TDM reservations or machine-readable notices.

Malaysia currently does not have any AI-specific legislation which requires model developers, providers or deployers to document training data sources, filtering processes, deduplication measures, removal requests, rights reservations, synthetic data generation or model development records.

Regarding the issue of cross-border infringement, under the CA 1987, it is generally understood that an action for copyright infringement normally relates to the infringement having occurred within the jurisdiction of Malaysia. It is arguable, if the infringing activity is a combination of various acts done within and outside Malaysia, to what extent the acts committed outside Malaysia will be considered an infringement liable to action in the Malaysian courts. For example, if the combination of the acts comprises a single transaction originating from Malaysia (such as extracting a copy of the copyright work) for making further copies outside Malaysia for AI training purposes, it is likely to be argued that each of those acts would be considered part of the infringement and all parties involved in those acts would be considered infringers and liable to legal action in the Malaysian courts.

A claimant must first identify a copyright-protected work and establish that the defendant carried out, or caused to be carried out, an act restricted by copyright in Malaysia. Depending on the facts, this may include reproduction of non-transient copies, adaptation, communication to the public or another exclusive right reserved to the copyright owner.

For training, fine-tuning and evaluation activities, the central issue is whether the relevant process involves the making of a copy of a copyright work, or a substantial part of that work.

Where the AI training occurs without the creation of an infringing copy, an argument can be made that no copyright infringement has occurred. Conversely, where copyright works are reproduced in the AI system, a claimant would generally need to prove:

• ownership or subsistence of copyright in the relevant work;
• objective similarity between the copyright work and the reproduced material;
• a causal connection showing that the defendant had access to and used the claimant’s work; and
• that a substantial part of the work was reproduced.

In the context of agentic AI, the fact that a particular act was carried out automatically by an AI agent, autonomous system or tool-using model does not necessarily provide a defence under Malaysian copyright law, as a claimant is not required to prove that the model developer or provider intended to infringe or was aware that infringement would occur. Similarly, where a defendant has caused or procured the infringing act through the operation of an AI system, the autonomous nature of the technology does not, in itself, negate liability.

There are no Malaysian statutory provisions, judicial decisions or regulatory guidelines specifically addressing whether model weights, parameters, embeddings, caches or other internal model artefacts constitute infringing copies, adaptations or extracted parts of copyright works or databases.

It can be argued that model weights and parameters are unlikely, in themselves, to constitute infringing copies of copyright works.

Embeddings present a more nuanced issue. Although an embedding may represent aspects of a copyright work, infringement under Malaysian law depends on whether a substantial part of the work has been reproduced, which is assessed qualitatively rather than quantitatively. It is therefore arguable that an embedding would not ordinarily amount to reproduction of a substantial part of a copyright work, although the position remains untested.

Caches and other temporary storage mechanisms may involve the creation of copies of copyrighted material. The key issue is whether such copies satisfy the transient and incidental copying exception under Section 13(2)(q) of the CA 1987. Where the copy is genuinely transient and incidental, infringement may not arise.

There are currently no Malaysian cases or regulatory decisions addressing memorisation, regurgitation, substantial similarity or de minimis copying in the AI context. Accordingly, courts would likely apply established copyright principles, including the qualitative assessment of substantial reproduction. The extent to which memorised content, embeddings or other internal model artefacts may constitute actionable reproductions therefore remains uncertain.

Malaysia presently has no AI-specific statutory framework governing the liability of model providers for infringing acts committed by users, downstream applications or autonomous AI agents. There are also no reported Malaysian cases addressing whether model providers may be held liable for copyright infringement arising from the operation of generative AI systems, retrieval-augmented generation tools or agentic AI deployments.

Under established Malaysian copyright principles, liability is more likely to arise where a defendant has authorised, procured, induced or otherwise facilitated the infringing conduct of another person. The degree of knowledge, control and involvement exercised by the model provider would therefore be relevant.

Where the model training process itself constitutes copyright infringement, for example because copyright works have been reproduced without authorisation and retained as part of training datasets, liability may arise against the model developer or provider from the outset. If users subsequently generate infringing outputs derived from those infringing datasets, both the user and the model provider may be exposed to liability in respect of the resulting infringement.

If AI training in the system did not give rise to an infringement, but instead infringement arises solely from the output being prompted by the user of the AI system, it is clear that the user is liable for causing the infringement due to the infringing output, but it is arguable under current law whether the AI provider is also liable for the infringement by enabling the user in that regard.

Malaysia does not currently have AI-specific safe harbour provisions addressing model developers, model providers or autonomous AI systems.

Malaysia does not have AI-specific legislation governing the use of confidential information, trade secrets or proprietary data in AI training, fine-tuning, prompting, retrieval, memory storage, tool use or model evaluation. Nor are there any reported Malaysian cases directly addressing the application of breach of confidence or trade secret principles in the context of generative AI systems.

Claims are most likely to arise where a person with access to confidential information discloses or uses that information without authorisation by inputting it into an AI system. Examples may be:

• an employee entering confidential business information into a publicly available AI service;
• a contractor using proprietary datasets to fine-tune a model contrary to contractual restrictions;
• a user uploading confidential documents to an AI platform for analysis;
• an AI agent accessing confidential materials through connected databases, software tools or retrieval systems without proper authority; or
• confidential information being used to evaluate, improve or train an AI model without the consent of the information owner.

Unlike copyright law, where liability generally depends on the occurrence of a restricted act such as reproduction, trade secret and confidentiality claims are principally concerned with the unauthorised use or exploitation of confidential information. Accordingly, the use of confidential information itself may constitute a breach regardless of whether there is any reproduction thereof.

In a breach of confidential information claim, the claimant would have to show that the defendant knew or ought to have known that the information was confidential. Where the defendant is a model developer or model provider, the claimant may have challenges depending on the circumstances in proving this point as it would depend on the model developer’s or model provider’s accessibility to and their actual or constructive knowledge of the nature of the subject matter which is being protected as confidential information.

The usual defences in respect of a claim for copyright infringement are that:

• the act complained of was not an act controlled by copyright; and
• the act complained of comes within an exception to infringement under the CA 1987.

Regarding the first type of defence, the model developer and model provider may seek to argue that no copies were made during the training process as the various elements involved do not result in a copy being made in material form. Ultimately, this is a technical analysis and dependent on the nature of the training process and how the AI model is structured.

Regarding the second type of defence, the model developer and model provider may seek to argue that even if copies of the copyright work were made, they were of a transient and incidental nature which constitutes an exception under the CA 1987. Alternatively, it may also be argued that even if they were not transient copies, they fall under the fair dealing exception of the CA 1987.

The above defences have been more thoroughly discussed in 2.2 Copyright in Software and Model-Related Materials and 3.1 Use of Copyright Works for Training.

In addition to the above substantive defences in a case of copyright infringement, there is also the procedural defence of an infringement action being barred by limitation. This defence would be applicable for all IP infringement actions, and the limitation period will be either five or six years from the date of infringement depending on the type of IP and the respective legislation and law governing that IP.

So far, there has been no reported cases on copyright infringement involving AI-generated outputs.

Under the CA 1987, copyright is infringed when any person commits (or causes any other person to commit) an act controlled by copyright without the consent of the copyright owner. Such acts controlled by copyright under the CA 1987 include, among others, reproducing the copyright work and communicating the copyright work to the public. The nature and type of AI outputs will determine whether infringement has occurred, as described below:

• Outputs that are substantially similar to the training works: The liability attached to outputs which are substantially similar to the training works would depend on the liability attached to the training works. If the training works were themselves substantially similar to a copyright work and therefore an infringement of that copyright work, then it follows that the output being substantially similar to that infringing training work would also likely be an infringement of the copyright work. Conversely, if the training works were not infringing any copyright work, it should follow that in the absence of other relevant factors, the output would likely not be an infringing output.
• Outputs that have reproduced a protected expression from the prompts or reference images: In the case where the outputs have reproduced identical copies of the protected expression or reference images, since such protected expression and/or reference images enjoy copyright protection, prima facie the outputs would be infringing copies by way of reproduction without consent of the copyright owner unless the user of the output is able to come within the exceptions to infringement under the CA 1987.
• Outputs that imitate style, genre, voice or artistic technique without literal copying: Generally speaking, the elements of style, genre, voice and artistic technique do not come under the rubric of copyright protection as they may not easily come under any category of works protected by copyright and may also be regarded as not being in a material form. It is also a trite principle in copyright law that copyright only protects the expression of an idea and not the idea itself. Accordingly, the above types of output are unlikely to be regarded as copyright infringement.

Unwittingly Generating, Using, Publishing or Commercialising AI Output

Whether knowledge or awareness would affect liability for infringement would depend on the type of IP. Where the infringement of IP does not require knowledge or intention as a requirement, the fact that the infringing act was done unwittingly would not matter in determining liability for infringement.

For example, in terms of copyright infringement, it is generally held that since it is a strict liability, lack of knowledge that the act is infringing a copyright work would not be a defence to a claim for infringement. Hence, a user or deployer who uses an AI system to produce infringing outputs would be liable for infringement regardless of whether they know of their acts being an infringement.

In contrast, where the output involves the use of trade secrets/confidential information, liability will generally only arise if the user or deployer knows or ought to know that the information is confidential in nature.

Configuring or Deploying an Agentic System That Then Carries Out the Infringing Act

It is likely that an infringing act done by an agentic system would be attributed to the user of the agentic system due to the fact that the agentic system was acting within the scope of the user’s instructions (no matter how wide the instructions may have been) and that the user had overall control over the agentic system. Such user would arguably be deemed as the person who caused such infringing act to be done.

There is currently no specific legislation on how AI may affect liability arising from the use of trade mark and other brand-related AI-outputs.

We are of the view that the existing legal principles and legislation on trade mark infringement, passing off, advertising and trade description should apply to any signs or works that imitate or suggest association with third-party marks, logos, product get-up, brand styles, character merchandising or celebrity/personality indicia regardless of whether such signs or works were generated using AI or created with the assistance of AI. Such liability will be attributed to the user of the AI output.

Under Malaysian law, a patent is infringed when a person exploits a patented invention without the consent of the patent owner. The meaning of “exploit” for purposes of infringement includes making, selling, importing, stocking and using the patented invention. The PA 1983 also provides for “imminent infringement” whereby a person who performs acts which make it likely that an infringement would occur will also be liable for such infringement.

Hence, if the AI output contains information which if used may lead to the patent being infringed, a person would be liable for infringement if they use such information in such way. In addition, it is possible that the AI developer/provider may be exposed to an action for imminent infringement if the facts are such that it is likely that an infringement would occur in view of the information provided in the AI output.

Under the Malaysian Industrial Designs Act 1996 (“IDA 1996”), a registered industrial design is infringed when a person applies the industrial design to any article or sells or imports any articles to which such industrial design is applied without the consent of the owner of the industrial design. There is also a concept of imminent infringement similar to the one for the PA 1983 explained above. Hence, for infringement of industrial designs, similar considerations apply mutatis mutandis to those for patents.

There has not been any decided case on IP infringement or related liability where an AI system or AI agent autonomously takes steps such as scraping or retrieving third-party content, using software tools or application programming interfaces (APIs), generating code or content, uploading or publishing outputs, making product or design choices, or otherwise interacting with IP-protected material.

That said, it is unlikely the courts will see the autonomous behaviour of an AI system or AI agent as a ground for the user or provider of the AI system or AI agent to disclaim liability for the infringing acts of the AI system or AI agent. This is because the AI system or AI agent will be regarded as being under the overall control of the said user or provider, and therefore, arguably the principles akin to strict and vicarious liability will come into play.

Generally, copyright protection in Malaysia is premised on human authorship based on the definition of “author” as found in the CA 1987. Furthermore, the CA 1987 requires sufficient effort to make a work original in character, which is best understood as requiring sufficient human effort.

On the other hand, the CA 1987 is silent on whether such effort must be solely human or may be assisted with tools and technology. Therefore, a possible reading of the CA 1987 on this issue is that a work may still enjoy copyright, even if the authorship thereof is with the assistance of tools and technology (such as in the form of AI), provided that in the final analysis, the originality in the work can be shown to be a result of sufficient human effort involving, inter alia, aesthetic choice, personal judgement and intellectual investment.

Having regard to the above, it can be seen that how the prompting is done in the use of the AI system and the manner of use of agentic systems in the creation of the work will determine whether human authorship has been sufficiently proven, which in turn will confirm whether copyright will subsist in the work being created.

The views expressed above have not yet been tested by the courts and may well be the subject matter of proposed amendments to be made to the CA 1987.

In contrast, and as mentioned in 1.4 Courts, IP Offices and Regulators, MyIPO currently seems to take a stricter view about the exclusive need for human authorship, suggesting that any form of AI assistance in the creation of a work will adversely affect the subsistence of copyright in the same. This is evidenced by how MyIPO processes applications for voluntary notification of copyright, which is provided for in the CA 1987. In processing such applications, MyIPO requires a declaration that “the work was created without AI assistance”, failing which such notification will not be allowed.

As explained in 6.1 Human Authorship and Copyright Protection, the CA 1987 does not recognise computer-generated or AI-generated works where there is no human author or where human contribution is minimal.

It is interesting to note that the position is not the same for the registration of industrial designs. The IDA 1996, in fact, recognises that an industrial design may be wholly generated by a computer with no human author, that such an industrial design may nevertheless still be accorded protection under it (provided that all other criteria are satisfied), and that the person by whom the arrangements necessary for the creation of the industrial design are made shall be deemed to be the author of the industrial design.

In the event that AI outputs are generated from existing works and the threshold of human effort and contribution are not met, and as such human authorship is challenged, there can be implications on the existence of copyright relating to the said outputs. For example, if the outputs were meant to give rise to a joint authorship between the author/owner of the existing works and the user who generated the said AI outputs from the existing works, it is arguable that the whole work (ie, existing work plus new work from the AI outputs) may not enjoy copyright protection.

Similarly, if a derivative work is wholly based on AI outputs without sufficient human effort, the derivative work will not enjoy copyright protection in terms of the adaptations and transformations made to the existing work, although the original existing work may well continue to enjoy the pre-existing copyright.

In respect of copyright, as mentioned above, when processing an application for voluntary notification of copyright, MyIPO requires a declaration that the work was created without the assistance of AI, failing which the application will be rejected. None of the other IP legislation and regulations require disclosure of AI involvement when registering or enforcing IP rights.

A proper reading of the PA 1983 suggests that reference to “inventor” must mean a human inventor, although there is no express provision in the PA 1983 defining “inventor” as a natural person. Therefore, following this interpretation of the PA 1983, an AI system cannot be named as an inventor, co-inventor or creator. However, like the CA 1987, the PA 1983 does not expressly prohibit the inventor from using AI to assist in making their invention, subject to our comments below.

It should be noted that in the patent application, the inventor needs to be named and the inventor needs to be human. Hence, if, for instance, the invention was entirely or mainly made or conceived by the AI system, then naming the human user of the AI system as the inventor in the patent application would be a false statement, making the patent application liable to be invalidated.

The question therefore arises as to what degree of AI assistance is acceptable for the human inventor to remain legitimately so called. In this regard, there is currently no Malaysian authority on AI-generated or AI-assisted inventions. However, it is likely that the Malaysian courts will be guided by the reasoning in the English case of Thaler v Comptroller-General of Patents, Designs and Trade Marks [2024] 2 All ER 527, where it was held that only the natural person who devised the inventive concept may be named as the inventor and, therefore, although an AI system may assist in generating ideas, identifying solutions or performing technical tasks in the process of making the invention, the human inventor must be the one who makes the substantial intellectual contribution that leads to the inventive concept being used for the patent application.

At the time of writing, the use or availability of AI does not change the legal tests for patentability in Malaysia. An invention must still be new, involve an inventive step (ie, one not obvious to a “person skilled in the art”) and be capable of industrial application. The “person skilled in the art” continues to be a hypothetical human, not an AI system.

However, AI may make it more challenging and difficult to demonstrate that an invention involves an inventive step. A solution that would once have required significant human effort may now be devised more easily and within a shorter period of time with the use or availability of AI. As a result, patent examiners may apply a higher benchmark for an invention to satisfy the inventive step requirement in a patent application.

There are currently no special rules on AI-generated disclosures, synthetic datasets or automatically generated technical documents. The definition of prior art under the PA 1983 is broad and includes information made available to the public anywhere in the world before the filing date. Therefore, AI-generated materials may form part of the prior art if they are publicly available through an identifiable source and sufficiently disclose the relevant invention, but if they remain confidential, then they do not form part of the prior art.

Protection and enforcement of the items below are subject to the relevant laws and legislation governing the IP as listed:

• Designs are protected by copyright and industrial design.
• Product configurations are protected by copyright and industrial design.
• Graphical iser interfaces are protected by copyright and industrial Design.
• Icons are protected by copyright and trade mark/passing off.
• Avatars are protected by copyright and trade mark/passing off.
• Virtual goods are protected by copyright.
• Trade dress is protected by trade mark/passing off.

The above IP and the criteria required under their respective legislation and law to confer IP protection have been further discussed in other parts of this guide. The fact that the above items may have been AI-generated may affect the availability of IP protection depending on the relevant IP legislation/law, also as discussed in other parts of this guide.

AI-created brand assets, including names, logos, slogans, sounds and motion marks, may be registered and protected as trade marks, subject to the requirements for registrability under the Trademarks Act 2019. Trade mark protection does not depend on human authorship of the mark, but on whether the mark is capable of distinguishing the goods or services of one undertaking from those of others.

Proprietorship of a registered trade mark would generally vest in the person who is named in the trade mark register as the proprietor of the trade mark, which would ordinarily be a legal person. The fact that the mark was generated by AI or with the assistance of AI does not, by itself, prevent registration, provided that the mark satisfies the requirements for trade mark registration.

If the AI-assisted works have met the threshold for human contribution as discussed above and are therefore conferred copyright under the CA 1987, then the rules pertaining to moral rights, attribution rights and integrity rights shall apply in full force with no difference as compared to works which are not AI-assisted. If, however, the threshold for human contribution is not met for the AI-assisted works, then none of the said rights will be applicable since the said works in the first place did not enjoy copyright protection under the CA 1987 due to not having the requisite human authorship.

Generally speaking, works that imitate the style of a human author may not be an infringement if there is no substantial similarity with a work of that human author. However, where the works imitate the style of a human creator or create a synthetic performance or falsely present themselves as human-created or created by a particular person, there could be other forms of IP infringement, such as false trade description or passing off. In such a case, it would not matter whether such works are AI-assisted or otherwise. The person creating and presenting the said works by themselves or using AI assistance would be liable for the said infringement.

There is a combination of IP and other laws which protect against the unlawful use of a person’s name, image, likeness, voice, performance, persona or digital replica, and which laws are triggered depends on the nature of the unlawful use. For example, a performer’s rights are enshrined in the CA 1987, giving the right to the performer to control the communication and reproduction of the performer’s performance. Similarly, the use of the name, likeness, voice, persona or digital replica of a famous person without that famous person’s consent may result in an action for passing off in view of the goodwill attached to the notoriety of that famous person.

In terms of available venues, claims relating to AI and IP issues/disputes would not be different from a general IP claim.

In Malaysia, the usual mechanism for litigants to obtain evidence is by way of general civil court procedures for the discovery, procurement and preservation of evidence, including seeking an order for discovery and, in certain cases, seizure and preservation of the evidence.

The general rule in civil court procedure is that evidence being sought to be discovered must be relevant to the facts and pleadings of the case. As long as the relevance of the evidence being sought to be discovered can be demonstrated, the said discovery procedures may apply to training data, model development, prompts, logs, outputs, source code, model weights, filters or evaluation results, including tool-call histories, browsing records, API logs, agent instructions, memory stores and orchestration logs.

Furthermore, if a claimant in the proceedings believes that the defendant may destroy such incriminating evidence, the claimant may apply to the court for a search and seizure order (also known as an Anton Piller order) whereby the claimant, under strict rules and as supervised by the court, is permitted to enter the defendant’s premises to undertake such search and seizure measures.

In order to protect the confidentiality of documents or trade secrets of a party that is ordered to produce evidence, the court has wide powers to instruct measures to be undertaken by all parties concerned to protect such confidentiality and trade secrets.

For example, in proceedings under an Anton Piller order, usually third-party independent solicitors and technical experts are directly engaged in the actual search and seizure process, and not the claimant, although the claimant or its solicitors may be present during the said process.

Additionally, during the court hearings relating to the case, the judge may order, usually at the request of either or both parties to the proceedings, that in view of the confidentiality of the evidence and matters being heard in court, the hearing may be conducted “in camera”, with the public gallery being vacated.

Under Malaysian law, generally, the High Court is empowered to grant interim injunctive relief that it deems proper and fit according to the circumstances of the case and on the principles of preserving the status quo having regard to the balance of convenience in view of the serious issue to be tried.

At the conclusion of the case, should the judge rule in favour of the claimant, a final permanent injunction may be granted against the defendant ordering it to cease the unlawful acts complained of.

In relation to AI and IP disputes, depending on the circumstances, the injunction granted may include orders to stop training, deployment or distribution, preserve evidence, remove works from datasets, disable outputs, apply filters/guardrails, delete or quarantine model versions, suspend agents, revoke tool access, disable autonomous publication or transaction functions, or provide corrective notice.

There have not been any decided cases in Malaysia to provide precedents on how monetary and non-monetary remedies will be granted for successful AI-related IP claims. However, there is no reason why the general rules relating to such remedies would not be applicable.

Typically, monetary remedies will include the following:

• Damages: This is generally understood to mean putting the claimant in the position monetarily as if the infringement had not occurred, and this would include losses and expense suffered or incurred by the claimant (such as royalties which would have been earned by the claimant). In addition, the court may order punitive damages to be awarded to the claimant if the conduct of the defendant was of a flagrant or recalcitrant nature.
• Account of profit: If the defendant had enriched itself in a manner beyond the loss suffered by the claimant, the claimant is also entitled to seek to disgorge the profits the defendant earned from the infringing act.

In addition to the above, the CA 1987 also provides for statutory damages as an option for the claimant in the form of a capped amount for each work which is infringed combined with an aggregate capped amount.

In terms of non-monetary remedies, the following will be considered by the court:

• injunctive relief, which could be in the form of prohibitory or mandatory injunctions to prevent further infringing activities;
• order for delivery up of any infringing copies/goods as well as the articles used for the making of infringing copies/goods; and
• order for destruction of the above items.

In terms of territoriality, it is possible for the claimant to claim for losses and damages suffered outside Malaysia but arising from the infringing activity in Malaysia, for example, where the defendant has earned royalties or other revenue from the sale of the infringing copies made in Malaysia to buyers outside Malaysia.

Where the defendant’s assets are outside the Malaysian courts’ jurisdiction, it is possible for the claimant to execute the judgment in the jurisdiction where the assets are located outside Malaysia by registering the foreign judgment in that jurisdiction (where there is a reciprocal arrangement between the two jurisdictions) or otherwise by suing on the foreign judgment in that jurisdiction subject to the laws therein. The above principles apply vice versa, ie, where a judgment is obtained in a jurisdiction outside Malaysia and sought to be enforced in Malaysia.

The following are key provisions which should be considered for inclusion in licensing arrangements for models relating to the use of data and content:

• Scope and permitted uses: The licence should clearly define the licensed content, including the types of works, formats, scope of data and any exclusions. It should also specify the permitted uses, such as training, fine-tuning, evaluation and retrieval-augmented generation, and whether use is commercial or non-commercial.
• Permitted models, territory and field of use: The licence should identify the AI models covered, whether the licence is exclusive or non-exclusive, the permitted territories for deployment, and any restrictions on industry sectors or use cases.
• Remuneration: Payment may take the form of a lump-sum fee, royalties or a hybrid structure, depending on the nature and anticipated use of the content.

• Attribution: The licence should address whether attribution is required and prohibit the removal or alteration of attribution notices, metadata or other provenance information.
• Audit rights: Given the opacity of AI systems, licensors should have audit rights to verify compliance. These may include access to relevant records, datasets, logs and documentation, as well as technical measures such as provenance tracking, output testing and memorisation assessments. Audit findings should remain confidential.
• Opt-out and withdrawal rights: The licence should provide mechanisms for withdrawing content from training datasets, specify when opt-out rights may be exercised, and define the consequences of withdrawal.
• Deletion obligations: The licence should set out when licensed data must be deleted, such as on termination, on completion of training, or where removal is requested and technically feasible.
• Downstream outputs: The licence should address ownership of outputs, allocation of liability for infringing outputs, and any safeguards required to reduce infringement risks, such as output filtering, memorisation controls and watermarking.

MyIPO is currently spearheading public consultation relating to proposed amendments to the CA 1987, which includes discussing, among others, issues relating to the use of copyright works for the training of AI models and protection of AI-assisted or AI-created works. The public consultation is only at its preliminary stage, which aims to ascertain stakeholders’ views on what the gaps are or what changes to the law need to be considered to cope with the advent of AI.

The Ministry of Digital has announced its proposal for an AI Governance Bill to be considered for the nation and has issued a public consultation paper for this purpose. It was stated that the scope of the proposed AI Governance Bill is to ensure responsible and safe AI use in Malaysia by regulating the developers and deployers of AI systems. As this was only a high-level consultation paper, there is yet no firm details as to how the Bill will be drafted. As read, the said consultation paper does not seem to cover the IP aspects of AI, but this may change as the public consultation takes its course.

There have been no official reports about the Malaysian Government participating in any international efforts to harmonise AI and IP rules with international, regional or foreign entities. Regionally, there is an ASEAN Working Group on Intellectual Property Cooperation (AWGIPC) consisting of the directors/heads of IP offices in all the ASEAN countries and formed with the intention of rationalising, harmonising and streamlining IP laws and procedures in the ASEAN region. It is likely that this topic of AI may become one of AWGIPC’s agenda items to be discussed in the near future.