Expert consultation
Prepared by
Muhammad Deckri Algamar, AIGP
&
Prof. Abu Bakar Munir
It cannot be recovered from an output afterwards. That fact decides who answers: the party that ingests the corpus, because it alone chose the sources and alone holds the records. Everything else in this submission follows from it.
Malaysia is doing something few jurisdictions have attempted. Rather than legislate for artificial intelligence in isolation, MyIPO has opened its copyright statute in a single consultation, from treaty accession to collective management to the training question itself. Reform on that breadth is harder than a single AI provision and considerably more useful, because it lets the answers be built to fit each other.
The timing matters beyond Malaysia. Almost everything written so far on generative AI and copyright has been written in Munich, London, Delaware and Delhi, for markets whose statutes and remedies differ from Malaysia's and from each other's. A CPTPP party in Southeast Asia legislating deliberately and in the open gives the global south a worked position rather than an inherited one.
That is also the case for a WIPO role. These questions are being answered jurisdiction by jurisdiction with no instrument aligning them. The member state proposals at SCCR/47 and SCCR/48, on licensing performers' voice and image rights and on compensation for AI training use, are the beginning of the forum where they could be.
One claim enters at the artificial intelligence topic and is carried by the rest, each supplying a component it needs.
Provenance cannot be reconstructed from an AI output after the fact. It can be established with certainty at the point of collection, at the point of registration, or through a licensed corpus. Because individual enforcement is available only to rights holders who can fund it, the mechanism that reaches everyone else is collective.
Malaysia already holds the components. Section 16B creates a statutory equitable remuneration right that arises without individual negotiation. Section 16B(3A) sends the rate to the Copyright Tribunal where parties do not agree. Section 36(1) already reaches a person who does, or causes another to do, a controlled act. None of the three is connected to AI.
MyIPO has opened its copyright statute across four klusters, and one question runs underneath most of what it asks. Any answered topic opens its response.
Each topic is one specimen, and its resolution records how settled Malaysian law is there. A rule in force resolves sharply. A proposal is coarser. Where nothing is in force the image barely resolves at all and the marks drift off the edge.
Only one topic resolves cleanly, and it is the one this submission asks MyIPO to protect.
Whether training on protected works requires permission is unresolved. Four incompatible models are put to respondents.
Rights management information is the one record that survives collection. The circumvention limb could settle the acquisition question by accident.
The safeguard set here is the best in the consultation, and it sits in a slide rather than in a provision.
Whether safe harbour reaches output a provider's own system generates is unaddressed. Automated matching detects similarity, not entitlement.
Article 5 protects distortion of a performance that exists. A fabricated performance modifies nothing, so nothing is engaged.
A diligent-search route to cultural heritage material. Training cannot be undone, so contingent compensation does not work here.
The only mechanism through which most Malaysian creators will ever be paid for AI use. Nothing connects the reform to AI.
Section 16B(3A) already gives it rate-setting power. Without a forum, silence never resolves into case law.
The mandatory declaration that a work was created without AI assistance, backed by a revocation clause. Malaysia's one concrete AI rule, and simplified documentation puts it at risk.
The difficulty of establishing provenance is an argument about who bears the burden of proof, not about whether the entitlement exists.
Topic 2.1, box 2 · only the developer holds the training records
Apakah manfaat paling utama yang anda jangkakan sekiranya Malaysia mempunyai dasar hak cipta AI yang jelas?
Box 1 · Main benefit of a clear AI copyright policy
Malaysian copyright law is silent on whether training an AI system on protected works requires permission, and the silence is not empty. The provision a court would start from, , is purpose-open: its fair dealing purposes follow the word "including". Whether training is fair dealing in Malaysia is undecided, not foreclosed.
Undecided cuts both ways. A rights holder must assume the reserved-rights default, on which training without a licence infringes. A developer can read the open words the other way, and the Delhi High Court has just done so with the same word in of the Indian Act, holding at interim stage in ANI v OpenAI that storing works to train a model falls within fair dealing.
Two defensible safest readings point in opposite directions. That deters lawful development inside Malaysia while doing nothing to prevent training conducted offshore on the same Malaysian works.
A clear rule, in whichever direction, converts uncertainty into a cost that can be priced. Developers can budget for licences or rely on an exception. Rights holders can enforce or negotiate. Investors can assess exposure before committing. Certainty is worth more to both sides than a favourable rule nobody can rely on.
The second benefit is coordination. The National AI Office consulted in parallel on an AI Governance Bill built on five principles, the fifth of which governs data: quality, integrity, provenance and security of the data used in AI systems. Principle 5 has no determinate content on training data until this consultation answers the question, because provenance in the copyright sense is precisely what is unsettled. Resolving it here gives the AI framework something to point to.
Apakah kebimbangan utama anda berkaitan AI dan hak cipta di Malaysia?
Box 2 · Main concern
Both concerns run to the same place: a rule that arrives after the corpora were built.
The first is that Malaysian rights holders whose works have already been ingested will have no route to compensation. Any rule this consultation produces operates prospectively. The corpora already exist. Malaysian music, film, journalism and literary works are in them now, and the consultation paper records creative industry and media bodies asking for precisely this: consent, compensation and transparency.
The second is that the difficulty of establishing provenance becomes the reason for that outcome. It is genuinely hard to determine, from an AI output alone, which protected works contributed to it. That difficulty is real, but it is an argument about who bears the burden of proof, not about whether the entitlement exists. Only the developer holds the training records. A rule requiring a rights holder to prove what entered a model they cannot inspect is a rule no rights holder can use.
The courts have not settled what a trained model retains, and this submission does not need them to. The Regional Court of Munich I treated works as reproduced within model parameters in GEMA v OpenAI on 11 November 2025, a decision now under appeal.
In Getty Images v Stability AI the point was not even contested: Getty did not allege that the model stores copies, its secondary infringement claim failed on that footing, and permission to appeal was granted in December 2025 on what counts as an infringing copy.
Malaysia should not conclude that because provenance is hard to verify after the fact, compensation is impractical.
Pada pendapat anda, apakah langkah yang perlu diambil Kerajaan bagi memastikan keseimbangan antara inovasi AI dan perlindungan hak pencipta?
Box 3 · The central box · steps to balance innovation and creators' rights
The rule turns on three decisions about who answers, and four about how the answer is made to work.
One. Decide acquisition and use separately. Singapore's exception turns on whether access was lawful. Japan's turns on the purpose of use and prejudice to the rights holder. These answer two questions. A workable rule sets lawful acquisition as a precondition, then applies a purpose and market-harm test to the use.
Two. Place input-side liability on the party that ingests. The entity that copies protected works into a training corpus, whether described as a model provider or a general-purpose AI developer, does an act controlled by copyright under and is therefore already within . The amendment should confirm this rather than leave it to inference. That party alone chose the sources, alone holds the records, and alone can prove lawful acquisition. Liability should follow the ingestion regardless of label, so that a party who fine-tunes on a corpus it assembled is treated the same as one who trained from scratch.
Three. Confine output-side liability to substantial reproduction. A deployer or user who prompts a system and publishes the result should be liable only where the output reproduces a protected work or a substantial part of it, which is the existing test in . A user cannot audit a corpus they cannot see. Making them answer for its lawfulness is neither fair nor enforceable, and would deter ordinary lawful use while leaving the ingesting party untouched.
Four. Put the record-keeping duty where the records are. Require providers to retain training-source records and to disclose them to a rights holder with a credible claim or to the Controller. Where records have not been kept, the absence should tell against the provider rather than defeat the claim. This is the practical answer to provenance: it cannot be reconstructed from an output, but it is fully knowable at the point of collection.
Five. Build the remuneration route on machinery Malaysia already has. Section 16B creates a statutory equitable remuneration right payable without individual negotiation, with the Tribunal setting the rate where parties do not agree. That is the architecture a training remuneration right needs. This consultation is separately strengthening collective management in Kluster 4 and expanding the Tribunal's royalty jurisdiction. Those reforms should be connected to AI expressly rather than developed alongside it.
Six. Say what the rule reaches, or the allocation is avoidable. Placing liability on the ingester answers who is responsible. It does not answer where. A developer can train wherever the law is weakest, and the choice-of-law rules make that ordinary rather than exotic: under of the Rome II Regulation the applicable law is that of the country for which protection is claimed, which fixes the analysis on the act rather than the actor, and nothing in the EU AI Act closes the resulting arbitrage.
Courts have gone both ways on whether a national court may reach foreign training at all. The Regional Court of Munich I took jurisdiction over training conducted in the United States in GEMA v Suno. In Getty Images v Stability AI the training claim failed in November 2025 for want of evidence that any training took place in the United Kingdom, and the surviving secondary infringement claim, on whether an imported model can be an infringing copy, is under appeal. The Munich approach has been criticised as a breach of comity, and the criticism notes that the European Union itself argued against extraterritorial reach in Abitron Austria v Hetronic.
Malaysia does not need to resolve that dispute, and should not try to. It needs to state its own position rather than inherit one by silence. The workable connecting factor is the one already proposed at topic 2.4: the rule should attach where protected Malaysian works are ingested, or where the resulting model is offered to users in Malaysia. That reaches the conduct that matters without asserting jurisdiction over foreign training as such.
Seven. Address use that has already occurred, and give one body express carriage of the overlap with the AI Governance Bill, stating which instrument owns the training-data transparency duty. A duty imposed in both is a duty in neither.
Cadangan atau ulasan lain berkait AI
Box 4 · Other comments
Three points bear on how MyIPO should read the comparative material below.
Recent decisions narrow the range of defensible answers, though not to a single line. The Regional Court of Munich I has now ruled twice: on 11 November 2025 in GEMA v OpenAI, and on 31 July 2026 in GEMA v Suno, the first European ruling on a generative AI music tool. The Suno judgment is the more significant for Malaysia. The court restrained reproduction for training carried out in the United States, took jurisdiction through a venue rule available to collecting societies, applied United States copyright law to those acts and rejected the fair use defence.
It also held that the provider of the system, not the user, is liable, and that compliance with the EU AI Act affords no copyright defence. Both judgments are on appeal to the Munich Court of Appeal and neither binds Malaysia. They indicate direction, not settled law.
The Court of Justice of the European Union is expected to rule in Like Company in September 2026 on questions bearing on the text and data mining exception, which will arrive after this consultation closes but before any amendment is finalised, and is worth awaiting on that specific point.
India has supplied a third direction. In ANI v OpenAI the Delhi High Court declined an interim injunction and held, prima facie, that storing protected works to train a model falls within fair dealing under , on a four-factor analysis whose fourth factor is the public interest. The order is interim and trial is pending. It matters here because the Indian provision shares its load-bearing word with Malaysia's: both extend fair dealing to purposes "including" the enumerated ones, which is why Box 1 treats the domestic question as open rather than foreclosed.
Across Germany, India and the United States no common doctrine is emerging, and the one lesson that does hold is structural: jurisdictions distinguishing acquisition from use are producing workable outcomes, while those treating AI training as one undifferentiated question are not.
The paper is right that the threshold question is whether copyright exists in AI output rather than who owns it. The test already exists. Section 7(3)(a) makes a literary, musical or artistic work eligible only where sufficient effort has been expended to make it original in character, and that is the provision an AI-assisted output must satisfy. What is missing is any statement of how it applies when part of the effort is the machine's.
The question is live in the region: writing in the on the Indonesian position, it has been argued that originality requires not merely labour but skill and creative effort, and that this standard is problematic to apply to works generated with AI assistance.
Codifying the Malaysian test improves on the present position, where the question is met case by case with no published guidance and no case law. Examination guidance should be published alongside any amendment.
This submission cites Bartz v Anthropic for the distinction between acquisition and use, and that distinction holds. The wider proposition sometimes drawn from it, that training is transformative and therefore fair, does not. In Kadrey v Meta the same district court reached the opposite emphasis in June 2025: Chhabria J held that harm to the market matters more than the purpose of the copying and that transformative use carries no automatic immunity, criticising the weight Alsup J had placed on transformativeness.
In Thomson Reuters v Ross the District of Delaware held in February 2025 that training on copyrighted legal material was not fair use, because what it produced was a market substitute rather than something genuinely transformative. That decision is on appeal to the Third Circuit, argued in June 2026, and may become the first appellate ruling anywhere on AI training. No ruling had issued when this submission was closed on 10 August 2026.
A jurisdiction whose trial courts disagree with each other, and whose first appellate word has not been spoken, is not a model to copy. It is a reason for Malaysia to legislate its own answer rather than wait for one to arrive. The single point on which both American lines agree is that market harm does real work, which is why the test proposed above turns on it.
I have indicated support for a hybrid on a specific understanding: lawful acquisition as a precondition, a purpose and market-harm test on use, and a remuneration route for use falling inside the exception. A hybrid left undefined is not a policy. If a hybrid is not adopted, a lawful-access model is preferable to full opt-in, because an opt-in regime is unenforceable against offshore training while imposing its full cost on developers operating inside Malaysia.
One government has already taken this route. Australia's Attorney-General stated on 26 October 2025 that the government will not entertain a text and data mining exception, a position reaffirmed on 1 April 2026 in the response to the Senate Select Committee on Adopting Artificial Intelligence. What is being built instead is a paid collective licensing framework for AI training under the Copyright Act, now under consultation by the Copyright and Artificial Intelligence Reference Group and weighed against voluntary licensing.
That is the remuneration limb of the hybrid, run through a collecting society. It is also why this answer and my answer at topic 4.1 should be read together: the collective management reform proposed there is the machinery a licensing route here would need.
The scholarship is moving the same way, and it supplies the frame the Delhi court reached for. argues that balancing generative AI against the rights of authors and performers requires recalibrating what public interest means, so that creators receive equitable remuneration while access to knowledge is preserved. The same author, applying to the training-data question, notes that the legal debate has otherwise run almost entirely on fair use and text and data mining. Malaysia is being asked to make that recalibration in statute rather than in argument.
This submission is made in a personal capacity as a researcher in international trade law and AI governance.
Apakah manfaat paling utama yang anda jangkakan daripada cadangan pindaan ini?
Box 1 · Main benefit
The main benefit is provenance, and provenance is the precondition for compensation.
Rights management information is the only mechanism in copyright law that travels with the work and records who made it. The central practical obstacle to remunerating Malaysian rights holders for AI training use is that the link between a work and an output cannot readily be established after the fact. Protecting RMI from removal and alteration preserves the one record that survives that process, and it does so without deciding any contested question about AI.
This matters most for works already in circulation. Where RMI has been stripped during collection or preprocessing, a Malaysian rights holder loses the evidence needed to establish that their work was used at all. A prohibition on removal, with effective remedies, protects the evidentiary basis of every future claim and of any collective licensing scheme built on it.
The TPM limb has a narrower benefit. It gives rights holders a means of controlling access at scale rather than work by work. The paper's framing of a shift from copy control to access control is the correct diagnosis, because bulk access is now the exposure.
This also has forward value. A licensed training corpus, in which rights are cleared at the point of collection, has established provenance by construction and does not depend on reconstructing the link between work and output afterwards. Intact rights management information is what makes assembling such a corpus possible.
Apakah kebimbangan utama anda berkaitan cadangan perlindungan TPM dan RMI?
Box 2 · Main concern
My concern is scope. A prohibition on circumventing access controls, drafted broadly, reaches conduct unconnected to infringement: security and encryption research, interoperability work, accessibility adaptation, library preservation, and analysis of lawfully obtained material.
There is a specific risk for this consultation. If circumventing an access control in order to obtain works for analysis becomes an offence regardless of whether the analysis is itself lawful, then the acquisition limb of the AI training question is settled restrictively here, inside a topic not framed as being about AI, while the AI topic is still asking respondents which model Malaysia should adopt. A decision of that consequence should be taken in the topic that is asking the question, not arrive as a side effect of another one.
Apakah cadangan anda bagi memastikan perlindungan TPM dan RMI dapat dilaksanakan secara berkesan tanpa menjejaskan pengguna yang sah?
Box 3 · Effectiveness without prejudicing lawful users
The obligation can be met without reaching lawful users, in three moves.
Tie liability to infringing purpose. Circumvention should be actionable where it is done to infringe or to enable infringement, rather than as a standalone offence untethered from the underlying right.
Enact express exceptions in the statute rather than in guidance, covering security and encryption research, interoperability, accessibility, library and archival preservation, and analysis of works to which the user already has lawful access.
The treaty permits this, and one point of precision is worth making first. The consultation paper cites CPTPP for both limbs. governs technological protection measures; governs rights management information.
The distinction matters here because provides expressly that a Party may adopt limitations and exceptions to the anti-circumvention prohibitions in order to enable non-infringing uses, where their adverse impact on those uses is determined through a legislative, regulatory or administrative process. This consultation is that process.
Parties have used it. New Zealand's Copyright Act 1994, section 226D, allows a qualified person to circumvent a technological protection measure on a user's behalf so that the user can exercise a permitted act, and the permitted-act limb of that section was inserted by New Zealand's own CPTPP implementing amendment in December 2018. Exceptions are not a concession against the obligation. They are part of its design.
Provide a route to obtain access where a technical measure blocks a use the Act otherwise permits. An exception that cannot be exercised because the lock remains in place is not an exception.
Maklum balas terbuka mengenai cadangan injunksi kehakiman dinamik.
Box 1 · Open feedback on the dynamic injunction proposal
I support the power. The one-site-one-order position described in the paper is not a policy choice, it is an artefact of procedure, and a rights holder who must return to court each time a domain changes has a remedy in name. The design of the power decides whether it is a remedy or a blocking tool.
One. Put the safeguards in the Act, not in the consultation paper. The paper states them already: judicial oversight at every stage, tests of necessity, proportionality and public interest, a right to be heard for affected parties, and targets confined to clearly infringing sites. That set is correct and I would not change it. But it presently exists in a slide. A court applying the amended Act will apply the words of the amended Act. If these tests are not on the face of the provision, the provision is a broad blocking power with a good explanatory memorandum.
Two. Define the addition step, which is the whole mechanism. Adding a mirror or successor site to a live order without a fresh suit is what makes the injunction dynamic and is also where the error risk sits, because the addition is the one stage a court has not individually examined in advance. The Act should say who may add a target, on what evidence, whether an officer of the court approves before or after the block takes effect, how the operator of an added site is notified, and how it applies to be removed. Approval after the fact is defensible if the review is real and fast. Approval by nobody is not.
Three. Specify the blocking method and prohibit collateral effect. An order that does not state its technical method invites the cheapest one, and the cheapest methods are the least precise. Blocking at the level of an address shared by unrelated services suppresses lawful sites that were never before the court and are not party to anything. The order should identify the method, and an added target should be removed from the order once it ceases to meet the factors that justified adding it. Orders should not run indefinitely by default.
Four. Publish what the mechanism does. The list of targets added under a live order, and periodic reporting on additions, removals and successful challenges, should be public. This is the same discipline recommended at topic 2.4 for automated takedown claims, and for the same reason: a removal mechanism that produces no record cannot be audited for accuracy, and accuracy is the only thing that distinguishes it from censorship.
The point I would most ask MyIPO to notice. Set this topic beside the next one. Blocking an entire website, the more drastic act, is proposed with judicial oversight at every stage, tests of necessity and proportionality, and a right to be heard. Removing a single work, the lesser act, is proposed with a twelve-hour clock and none of those things.
The asymmetry runs the wrong way, and it runs the wrong way against volume: the court-ordered remedy will be used rarely and deliberately, the notice remedy constantly and mostly by software.
Whatever safeguards this consultation is willing to write into the remedy a judge supervises, it should be willing to write a proportionate version of into the remedy nobody supervises. The safeguard set in this topic is the model. It should be carried into topic 2.4, scaled to the lesser interference.
One boundary. This mechanism is designed for sites whose purpose or effect is infringement. It should not be extended, by the elasticity of same purpose or effect, to reach general-purpose AI services or the models behind them. Those raise the questions this consultation asks at topic 2.1, where the liability rule is still being decided, and they should be answered there on the merits rather than arrive through an enforcement provision drafted for pirate site mirrors.
Cadangan Penambahbaikkan?
Box 1 · Suggested improvements
Five changes would make the regime workable for the parties who actually use it.
One. The takedown timeframe. The paper states an expedited twelve-hour window described as aligned with UK, EU and US practice. None of the three regimes sets a fixed clock.
In the United States, 17 U.S.C. section 512(c)(1)(C) conditions the hosting safe harbour on responding expeditiously to remove or disable access upon notice, and states no period. In the European Union, Article 16 of the Digital Services Act requires notices to be processed in a timely, diligent, non-arbitrary and objective manner, and requires the provider to notify its decision with reasons and with information on available redress.
In the United Kingdom, regulation 19 of the Electronic Commerce (EC Directive) Regulations 2002 makes the hosting defence contingent on acting expeditiously once the provider has knowledge. Each of the three is a standard calibrated to the notice and the provider. None of them is twelve hours. This form still asks respondents to choose between six, twelve and forty-eight hours, so the figure appears to remain open.
The difference is not presentational. A short fixed clock forces removal before assessment, which is the failure described next. An expeditious standard with a stated outer limit, paired with the reasoned decision and redress route the Digital Services Act already requires, delivers the same speed with fewer wrongful removals.
Two. Automated matching must not, by itself, remove content or redirect revenue. Where takedown is operated at scale it is operated by software, and matching systems detect similarity, not entitlement. They cannot tell an original from a copy of it.
The consequence is documented. In March 2026 a third party filed automated ownership claims against an independent folk musician's own video recordings, using material generated by cloning her voice from those same recordings. Her revenue was redirected to the claimant. Claims were asserted even over public domain material dating to the 1870s. They were withdrawn only after the matter attracted public attention, not because the process identified the error. The distributor observed afterwards that her recordings sat outside the fingerprint databases these systems rely on, which is what made her vulnerable.
That exposure is not evenly distributed. Independent artists, traditional performers and ethnic repertoire are the least likely to be enrolled in commercial fingerprint databases, and are therefore the most exposed to having their own work claimed by someone else. This consultation separately identifies ethnic works as needing clarity in collective management. The same repertoire is the most at risk here.
Accordingly the amendment should provide that an automated match may flag content but may not by itself effect removal or redirect revenue; that a claimant must attest to entitlement, with consequences for claims made without a reasonable basis; that where an uploader disputes a claim, disputed revenue is held rather than paid over until the dispute is resolved; that no claim may be sustained over material in the public domain; and that service providers report periodically on claim volumes and reversal rates. Speed without accuracy transfers income from creators to whoever files first.
There is very recent support for restraint, from the highest level. In Cox v Sony the Supreme Court of the United States held unanimously on 25 March 2026, with a concurrence from Sotomayor and Jackson JJ, that contributory copyright liability no longer runs on knowledge of infringement combined with a failure to take reasonable steps. Secondary liability now requires either inducement, meaning affirmative encouragement, or a service tailored to infringement, and internet access, with its substantial lawful uses, is not.
The decision does not bind Malaysia and concerns liability rather than takedown procedure. Its relevance is the direction of travel: the premise that knowledge plus inaction equals responsibility is being narrowed in the jurisdiction that built it, at the moment this consultation proposes to tighten one here.
Three. Safe harbour should not shelter output a provider's own system generates. Safe harbour exists because a host does not choose, see or control what users upload. That rationale does not transfer to a provider whose model produces the material. Where output reproduces a protected work or a substantial part of it, the provider should answer under , which already reaches a person who does or causes another to do a controlled act. In GEMA v Suno the Regional Court of Munich I held on 31 July 2026 that the provider, not the prompting user, bears liability.
Cox does not cut against this. Cox concerns secondary liability for someone else's act. Where a provider's own model produces the material, its liability is primary, arising from its own conduct, and already reaches it.
Four. Protect the good-faith deployer and let recourse run upward. A deployer using a licensed system in good faith, whose output does not reproduce a protected work or a substantial part, should not be liable, and should have recourse against the provider where the corpus proves unlawful. This creates commercial demand for warranted training data, which is how rights holders actually get paid.
Five. Reach providers offering models into Malaysia from abroad. A rule confined to domestic providers would advantage offshore ingestion of Malaysian works. Offering a model to users in Malaysia should be sufficient connection.
On the relationship between this topic and the dynamic injunction proposal, see my response at topic 2.3. The safeguards the paper proposes there are the ones this topic needs.
Pada Pandangan anda, apakah manfaat paling utama yang anda jangkakan sekiranya Malaysia menyertai BTAP?
Box 1 · Main benefit of accession
Accession would close a long-standing asymmetry. The WIPO Performances and Phonograms Treaty gave performers protection in aural performances while leaving audiovisual performers substantially outside the international framework. The Beijing Treaty corrects that, and Malaysian actors, dancers and audiovisual performers would gain moral rights and economic rights in their fixed performances on the same footing as their counterparts elsewhere.
There is a second benefit specific to this moment. Performers are the class most exposed to generative reproduction of voice, image and persona, and the terms on which that will be governed are being negotiated now, at the forty-seventh and forty-eighth sessions of the Standing Committee on Copyright and Related Rights, where member states have advanced proposals on licensing performers' voice and image rights and on compensation for AI training use.
Accession places Malaysia inside the instrument through which those questions will be worked out, as a participant in the drafting rather than a recipient of the result.
Apakah kebimbangan utama anda?
Box 2 · Main concern
My concern is that the Treaty, taken alone, protects against a category of harm that is no longer the principal one.
Article 5 gives a performer the right to be identified and the right to object to distortion, mutilation or other modification of his performances where prejudicial to reputation. The economic rights in Articles 7 to 11 attach to performances fixed in audiovisual fixations. The architecture presupposes a performance that exists and a copy that can be identified.
Generative systems do not work that way. A model that reconstructs a performer's voice or likeness from statistical patterns learned across many recordings produces a performance that never occurred. It modifies no fixation, so the integrity right is not engaged. It reproduces no particular fixed performance, so the economic rights are not engaged. The output may be commercially valuable precisely because it is recognisably that performer, and yet fall outside every right the Treaty confers.
This is a structural feature of fixation-based protection rather than a drafting oversight, and it is common to copyright, related rights and passing off alike. Malaysia can accede to the Beijing Treaty and still leave its performers without a remedy against the exploitation they are most likely to encounter.
Adakah terdapat mana-mana aspek yang perlu diberi perhatian oleh Kerajaan sebelum Malaysia menyertai BTAP?
Box 3 · Aspects to attend to before accession
Four, all concerning implementation rather than accession itself.
Legislate for synthetic performances, not only for modified ones. Implementing legislation should reach the unauthorised generation of a performance in a performer's voice, image or likeness where that capability was derived from their performances, without requiring the claimant to identify a particular fixation that was reproduced. Without this the protection tracks a technology that has been superseded.
Make the right collectively exercisable. A right that can only be enforced by individual action is available in practice to performers who can fund litigation. The population most exposed is the opposite one: session performers, voice artists and independent performers whose recordings supply training material and who have no realistic route to a remedy. Statutory responses elsewhere, including the digital replica laws enacted in several United States jurisdictions, are real but reach only claimants who can already afford to litigate. Malaysia can avoid that outcome by making the right licensable and enforceable through a collecting society from the outset.
Connect this to the collective management and Tribunal reforms in this same consultation. Kluster 4 proposes strengthening collective management and giving the Copyright Tribunal jurisdiction over royalty disputes. Those are the collection and rate-setting layers a collectively exercised likeness right would require. They are being designed in the same exercise, and should be designed together.
Do not let takedown mechanisms substitute for the missing right. Where the correct right does not exist, pressure falls on notice and takedown to perform work it was not designed for, which produces both over-removal and under-protection. A defined performers' right resolves at the source what an automated matching system cannot resolve at the margin.
Cadangan/Ulasan Keseluruhan
Box 4 · Overall comments
One framing point.
Unauthorised AI reproduction of voice and likeness is generally discussed as a problem affecting well-known performers. The evidence points the other way. Prominent artists hold registered rights, retain counsel, and are monitored by rights management systems that detect misuse. The greater exposure sits with performers who hold the same rights in law and no means to invoke them, and who are frequently outside the commercial fingerprint databases through which infringement is detected at all.
A recent case in which a folk musician's own recordings were claimed against her by a party that had cloned her voice from those recordings, and who recovered her position only after public attention, illustrates how thin the protection is at that end of the market.
Treating this as a question about celebrities leads to a stronger individual cause of action. Treating it as a question about creative labour leads to an accessible collective one. The second is the design Malaysia should pursue, and the Beijing Treaty is the natural vehicle for it.
These observations draw on research I am currently completing on performers' rights and generative reproduction. This submission is made in a personal capacity.
Pandangan anda bagi memastikan mekanisme Orphan Works dapat dilaksanakan secara telus, adil dan berkesan?
Box 1 · Transparent, fair and effective implementation
I support the mechanism. Three conditions would determine whether it works.
First, state whether permitted use includes computational analysis. The consultation paper describes the beneficiaries as archives, libraries, universities and digital repositories, and the material as old books, historical photographs, manuscripts, audiovisual recordings, archival documents and cultural heritage material. That is a description of a training corpus as much as a reading room. The scheme will be read by developers as a lawful route to a large body of Malaysian material, whether or not it was drafted with that in mind. Silence on the point will be resolved by whoever moves first. The Act should say expressly whether computational analysis is inside or outside the permitted use, and if inside, on what terms.
Second, recognise that training is not reversible, and design for that. Orphan works schemes rest on an assumption of reversibility. If the owner appears, use stops and payment follows. This consultation adopts that assumption, and the questions in this section ask whether rights holders remain protected if they are later identified. For ordinary uses the assumption holds: a book leaves the shelf, a scan comes down. It does not hold for training. Once a work has been ingested, its influence on the trained model cannot be removed without retraining the model. Courts differ on how to describe what a model retains, and nothing here depends on that dispute: on either description, the ingestion cannot be undone. A reappearing owner therefore has no remedy in kind, only a remedy in money.
The consequence for drafting is that remuneration cannot be contingent on the owner coming forward. Where an orphan work is used for training, a sum should be set aside at the point of use and held, whether or not anyone ever claims it, with unclaimed amounts applied to the sector after a defined period. Contingent compensation works only where the underlying use can be undone.
Third, publish the diligent searches. A public register recording what was searched, where, and what was found gives the scheme its transparency and gives a later claimant something to test. It also produces something of independent value: a corpus assembled through a recorded search process has documented provenance, which is exactly what cannot be reconstructed after the fact from an AI output.
A closing observation. Malaysia's orphan works pool is disproportionately its own cultural heritage, including traditional and ethnic material whose authorship was never individually recorded. Whoever obtains computational access to that pool obtains the ability to train on Malaysian heritage. That should be a deliberate national decision taken in this consultation, with terms attached, rather than an incidental consequence of a mechanism designed for library access.
One further limit, which the consultation paper does not reach. Some of that heritage was never within copyright at all. Guy Morrow has made the point about First Nations cultural knowledge in Australia, and it holds for Malaysian and Indonesian traditional knowledge equally: a framework resting on copyright cannot protect material copyright never covered, and a remuneration route built on copyright will pay nobody for it.
That is not an argument against the mechanism proposed here. It is a reason to be explicit that this mechanism does not reach the whole of the material MyIPO is describing, and that the remainder needs a separate instrument.
Pandangan anda bagi memastikan CMO di Malaysia dapat beroperasi secara lebih telus, cekap dan berakauntabiliti?
Box 1 · Transparency, efficiency and accountability
Four measures would make collective management accountable to its own members.
The problem is regional, and Malaysia is not solving it alone. Indonesia has put before the Standing Committee on Copyright and Related Rights, a proposal for a binding instrument on the governance of copyright royalty in the digital environment. It names limited institutional capacity within collective management organisations as a foundational challenge for developing countries, alongside the framework for collection and distribution and the absence of a standardised global database.
Work on Indonesian music remuneration, in , reaches the same diagnosis from the domestic side. The measures below are what that diagnosis looks like written into a Malaysian statute.
Publish distribution rules and outcomes. Accountability follows from members being able to see how a tariff was set, how collections were allocated, and what proportion was distributed rather than retained. Annual reporting should include collection and distribution totals, administrative deduction rates, and the value and age of undistributed royalties.
Regulate undistributed royalties expressly. Where a rights holder cannot be identified or located, the sums held should be recorded, published, and applied under a stated rule after a defined period. This is also the point at which the consultation's orphan works work and its collective management work meet, and they should be drafted consistently.
Give members a low-cost route to challenge a distribution. Accountability without an accessible remedy is reporting, not accountability. This should connect to the Copyright Tribunal reforms proposed elsewhere in this consultation.
Resolve mandate and repertoire overlap, including for ethnic works. Malaysia licenses music through three declared bodies with distinct mandates: Music Authors' Copyright Protection Berhad for composers, lyricists and publishers, Public Performance Malaysia for sound recording producers, and Recording Performers Malaysia for recording performers.
Malaysia has already tried consolidating them. Music Rights Malaysia was declared by MyIPO and began joint licensing on 1 January 2017, and it failed: its constituent bodies could not reach agreement on material operational matters, its licensing body status was revoked and it was dissolved in 2020, with licensing returning to the three bodies separately.
The lesson is not that a single window is wrong. It is that a joint vehicle with no rule for resolving deadlock does not survive its first one. Any mandate reform should state who decides when the bodies do not agree, which is the same gap the Tribunal reforms in topic 4.2 close elsewhere. A creator whose repertoire falls between mandates is a creator nobody collects for, and ethnic repertoire is the clearest instance of it.
The observation. These reforms are usually justified as domestic housekeeping. Their larger significance is that collective management is the only mechanism through which most Malaysian creators will ever be paid for AI use of their work.
The reason is distributional. Individual enforcement against a model developer requires identifying the use, establishing provenance, and funding litigation across borders. A well-advised recording company can do that. A session musician, a traditional performer or an independent artist cannot. Enforcement designed around individual action reaches the people who least need it. The recent case of the folk musician whose recordings were claimed against her by a party that had cloned her voice illustrates the point: she recovered her position through public attention, not through any mechanism available to her as of right.
Malaysia already holds the components. Section 16B creates a statutory equitable remuneration right that arises without individual negotiation, and sends the rate to the Copyright Tribunal where the parties do not agree. A collecting society is the natural collection and distribution layer for such a right. What is missing is a mandate extending to AI training use, on a blanket basis, with distribution calibrated to how much a rights holder's work or identity is actually used rather than to prominence.
This is a live multilateral question, not a speculative one. Member states advanced proposals on licensing performers' voice and image rights and on compensation for AI training use at the forty-seventh and forty-eighth sessions of the WIPO Standing Committee on Copyright and Related Rights, and Indonesia has tabled a proposal for a binding instrument on the governance of copyright royalties in the digital environment.
Two collecting societies have already moved, and the significant point is that neither waited for legislation. In Germany, GEMA published a generative AI licensing model in 2024 and has since launched a cleared training dataset for AI developers, while separately obtaining judgments against two AI providers.
In Korea, the Korea Music Copyright Association replaced its blanket refusal to register AI-assisted works on 4 August 2026 with a disclosure standard, enforced through its trust agreements by suspending royalties, recovering sums improperly paid and terminating membership, with a further revision proposed to allow a penalty of up to three times any amount wrongly received.
Both acted through their own rules and contracts. That is the practical case for regulating collecting societies well: a properly governed society can set and enforce an AI standard across its whole repertoire in months, where a statutory amendment takes years and reaches only what it expressly names. It is also the reason the governance measures above matter. A society with that much power over its members' income needs published rules, an accessible challenge route and real oversight.
Australia is going further still. Having rejected a text and data mining exception, it is now consulting on a paid collective licensing framework for AI training under its Copyright Act. That is the same architecture proposed at topic 2.1, and it would sit on precisely the collection and distribution layer this topic is reforming. A government that decides to pay creators for training use needs a competent society to pay them through.
If Malaysia strengthens collective management without connecting it to AI use, it will have built the machinery and left the largest new revenue stream outside it.
Sila berikan sebarang ulasan atau pandangan tambahan berkaitan cadangan pindaan Tribunal Hak Cipta. (Sekiranya ada)
Box 1 · Additional comments
I support the proposals, particularly the new royalty disputes jurisdiction and the provision allowing either party to bring a matter without the other's agreement. That second element does most of the work: a remedy available only by mutual consent is not a remedy against a party with no incentive to consent.
Three additions.
Confirm that the jurisdiction reaches AI-related licensing. Section 16B(3A) already empowers the Tribunal to determine equitable remuneration where parties do not agree. That is the correct architecture for AI training tariffs, and the amendment should make clear the new jurisdiction extends to disputes over remuneration for AI use, whether or not a licensing body and its member are on opposite sides.
Resource the Tribunal for valuation, not only for procedure. Setting a rate for AI training use is unlike setting a broadcast tariff. It requires evidence about corpus composition, usage measurement and comparable licensing. The sixty-day decision deadline is welcome but will only hold if the Tribunal has access to technical expertise.
Publish determinations with reasons. A body of published tariff reasoning is a public good. It also addresses a structural problem: where no forum determines these questions, the law stays silent, and silence is then read as permission or prohibition depending on who is arguing. A Tribunal that decides and publishes is how a jurisdiction stops being silent.
Cadangan penambahbaikan yang anda syorkan bagi sistem Pemberitahuan Sukarela Hak Cipta (CVN)?
Box 1 · Suggested improvements
I support simplification, with one qualification that runs the other way.
requires an applicant to declare that the work was created without the assistance of Artificial Intelligence (AI) and solely represents the author's own creativity and personal effort. The section is marked required, and the same declaration reserves to MyIPO the right to revoke or withdraw an issued certificate where the information submitted is found to be false or misleading. That declaration is, so far as I am aware, among the most concrete AI-specific requirements in any national copyright filing system. It should not be removed as part of reducing documentation.
It should, however, be refined, and the declaration already contains the material for its own refinement. It runs two limbs. The first, that the work represents the author's own creativity and personal effort, maps onto the statutory test, since makes a work eligible only where sufficient effort has been expended to make it original in character. The second, that no AI assisted, tests something the Act does not ask. My suggestion is to keep the first limb and grade the second.
The case for grading rather than holding the absolute is that a binary declaration is increasingly difficult to make honestly, since AI assistance now appears in ordinary creative tools, and here the difficulty carries a consequence. An applicant who declares inaccurately places the resulting certificate at risk under the revocation clause in the same section. A requirement that invites inaccurate answers protects nobody, and this one exposes the applicant who gives them.
Recording the nature and extent of AI involvement instead turns the declaration into evidence going to the question the Controller must decide, rather than a gate standing in front of it.
That design is now in operation. On 4 August 2026 the Korea Music Copyright Association reversed its blanket refusal to register AI-assisted works and replaced it with a disclosure requirement: a registrant must state which parts of the work involved AI, name the tools used, and describe their own creative contribution, and may be asked for supporting material including project files, editing histories, prompt records and generation logs. Works produced entirely by simple prompting remain ineligible. The test applied is whether a human played a substantial and leading role in writing, composing or arranging.
Two qualifications on that example. It is the registration policy of a private collecting society rather than Korean law, and it governs whether the association will administer a work rather than whether copyright subsists in it. It also covers music alone. Neither qualification weakens its value here, because Malaysia's question at this topic is likewise about what a registration system should record. It shows that a graded disclosure requirement is administrable, and that the evidence it produces is the kind a registry can actually assess.
A second example sits closer to home and answers a different question. On 20 July 2026 the Info-communications Media Development Authority of Singapore published Transparency Guidelines for Generative AI Chatbots, built around a Chatbot Information Card that sets out a system's capabilities and limitations, the measures used to manage safety and reliability, data handling practices, and the channels for reporting problems.
The drafting requirements are the useful part: plain language at a secondary-school reading level, easy to locate, and layered, so that a summary sits at the front with the detail one click away. Named organisations including DBS, OCBC and Singapore Airlines have said they will work to it.
These guidelines are voluntary, and I do not offer them as evidence that disclosure obligations are hardening across the region. Their value here is design. A notification form that asks an applicant to describe AI involvement will produce useful evidence only if the question is drafted so that an honest applicant can answer it, and Singapore has done that drafting work in an adjacent context.
Pandangan atau cadangan tambahan berkaitan cadangan pindaan sistem Pemberitahuan Sukarela Hak Cipta (CVN).
Box 2 · Additional comments
One further point, on what the register could become.
If notifications record the nature and extent of AI involvement, the Register of Copyright accumulates, over time, a national record of how AI is actually being used in Malaysian creative production. That has evidential value for individual disputes, given the proposal to make certified extracts admissible, and analytical value for policy, since it would allow the effect of whatever rule this consultation adopts to be observed rather than assumed.
This connects to the difficulty running through the AI topic of this consultation. Provenance cannot be reconstructed from an AI output after the fact. It can be recorded at the point of creation and at the point of registration. A notification system that captures it is building the evidence base that the training question will otherwise lack.
Every authority below is marked alongside its relevance.
| Authority | Citation | Held, and how it is used |
|---|---|---|
| GEMA v OpenAI | LG München I 11 Nov 2025 42 O 14139/24 |
Memorisation of lyrics within model parameters is reproduction, in the model and on output, outside the text and data mining exception. Used for the proposition that provenance problems are tractable, and that ingestion is a controlled act. On appeal to OLG München |
| GEMA v Suno | LG München I 31 Jul 2026 42 O 763/25 |
First European ruling on a generative AI music tool. Restrained reproduction for training carried out in the United States, took jurisdiction through a venue rule available to collecting societies, applied United States law and rejected fair use. Held that the provider is liable, not the user, and that EU AI Act compliance affords no copyright defence. Carries the liability allocation and the extraterritorial point. Appeal signalled |
| Bartz v Anthropic | N.D. Cal. Alsup J 23 Jun 2025 |
Training on lawfully acquired books was exceedingly transformative and fair use. Downloading pirated copies to build a central library was not. Carries the acquisition and use split. |
| Bartz settlement | N.D. Cal. final approval 20 Jul 2026 |
USD 1.5 billion, approximately USD 3,100 per work across more than 480,000 works. Used for the quantum and the method: the figure was reached by counting the corpus, not by tracing outputs. Creates no binding precedent |
| CJEU, Like Company | Ruling expected September 2026 |
Expected to address the text and data mining exception. Lands after this consultation closes but before any amendment is finalised. Pending |
| Beijing Treaty | Art 5 Arts 7 to 11 |
Moral rights attach to distortion of his performances. Economic rights attach to performances fixed in audiovisual fixations. The fixation architecture that synthetic replicas fall outside. |
| Copyright Act 1987 | s.13(1), s.36(1) s.16B, 16B(3A) to (3B) |
Substantial part test. Infringement by a person who does, or causes another to do, a controlled act. Statutory equitable remuneration. Tribunal rate-setting where parties do not agree. |
| Copyright Act 1987 | s.7(3)(a) | Eligibility requires that sufficient effort has been expended to make the work original in character. The test an AI-assisted work must satisfy. Carries the subsistence point at 2.1 and the registration design at 4.3. |
| Notice and takedown | 17 U.S.C. s.512(c)(1)(C) DSA Art 16 UK e-Commerce Regs reg 19 |
All three require expeditious or timely action and none states a period. Used to test the consultation paper's claim that a fixed twelve-hour clock is aligned with UK, EU and US practice. The claim does not hold |
| CPTPP | Art 18.68, Art 18.69 NZ Copyright Act s.226D |
18.68 governs technological measures, 18.69 governs rights management information; the paper cites 18.69 for both. Art 18.68(4) permits limitations and exceptions to enable non-infringing uses, and New Zealand's permitted-act exception was inserted by its own CPTPP implementing amendment. |
| Malaysian licensing bodies | MACP, PPM, RPM MRM dissolved 2020 |
Three declared bodies with distinct mandates. The joint vehicle declared from 1 January 2017 had its status revoked and was dissolved after its constituents could not agree on operational matters. Carries the mandate-overlap point at 4.1. |
| AI Governance Bill | Principle 5 of five NAIO consultation |
Data governance: quality, integrity, provenance and security of data used in AI systems. The principle whose content depends on the answer this consultation gives. |
| Form CR-1 | MyIPO Section G |
A mandatory declaration that the work was created without the assistance of Artificial Intelligence (AI) and solely represents the author's own creativity and personal effort, with a revocation clause attached. The provision at risk from simplified documentation. |
| GEMA licensing, PLAI | Sept 2024 Jul 2026 |
First collecting society to publish a generative AI licensing model, then to launch a cleared training dataset for AI developers. The collective route working end to end. |
| WIPO SCCR/47 and 48 | Dec 2025, May 2026 incl. SCCR/47/6 |
Member state proposals on licensing performers' voice and image rights and on compensation for AI training use. Indonesia's proposal for a binding instrument on the governance of copyright royalties in the digital environment. |
| Automated claim incident | March 2026 trade press |
Automated ownership claims filed against an independent musician's own recordings using material cloned from them, including over public domain works. Withdrawn after publicity, not by process. Not named in the submission |
| Getty v Stability AI | EWHC November 2025 final |
Copyright claims largely rejected. Getty withdrew its primary infringement claims at trial. The court accepted that the model does not contain copies, and declined secondary infringement because training occurred outside the United Kingdom. Used at 2.1 Box 3 for territoriality. Contradicts the memorisation finding used elsewhere in 2.1 and 3.3. Unresolved |
| Cox v Sony | US Supreme Court 2026, unanimous |
Contributory liability narrowed. Knowledge of infringement plus failure to take reasonable steps is eliminated as a route; only inducement or tailoring remain. Carries the restraint direction at 2.4, and is expressly distinguished there from primary liability for a provider's own output. |
| Kadrey v Meta | N.D. Cal. Chhabria J June 2025 |
Meta won on fair use, but market harm was held to matter more than the purpose of the copying, and transformative use carries no automatic immunity. Criticises the weight given to transformativeness in Bartz. Shows the American position is split inside one district. |
| Thomson Reuters v Ross | D. Del. February 2025 |
Training on copyrighted legal material was not fair use: a market substitute rather than something genuinely transformative. On appeal, Third Circuit, argued June 2026. May be the first appellate ruling anywhere on AI training |
| Australia, CAIRG | 26 Oct 2025 reaffirmed 1 Apr 2026 |
The government will not entertain a text and data mining exception. A paid collective licensing framework for AI training is instead under consultation. Carries the hybrid at 2.1 and the machinery point at 4.1. Web-search consensus. Primary texts not read |
| Rome II Art 8, comity | Journal analysis Edward Lee, Jul 2026 |
The applicable law is that of the country for which protection is claimed, which lets a developer relocate training to a weaker jurisdiction, and nothing in the EU AI Act closes it. Munich is criticised for breaching comity, citing the European Union's own contrary position in Abitron Austria v Hetronic. Carries the reach step at 2.1 Box 3. |
| Singapore, IMDA | 20 July 2026 voluntary |
Chatbot Information Card: capabilities and limitations, safety measures, data handling, complaint channels, in plain language at secondary-school reading level with layered disclosure. Used at 4.3 for design detail. Voluntary guidance. Not offered as stronger than KOMCA |
| Vietnam, Law 131/2025 | Effective 1 Apr 2026 | Reads as a controlled exception permitting AI training on legally published material, not the opt-out right an earlier source described. The source claiming an opt-out is offline. Unresolved. Cited in neither direction |
| ANI v OpenAI | Delhi interim |
Reported as a liberal reading of fair dealing with a public-interest limb inside the fairness analysis. Judgment held but unread. Not used anywhere |
This submission is filed by both authors in their personal capacities. It carries no institutional, employer or organisational affiliation, and no client interest is represented in it.
deckrialgamar@gmail.com
linkedin.com/in/muhammaddeckrialgamar
Correspondence on this submission, including from MyIPO, is welcome at the address above.
Submission to the public consultation on proposed amendments to the Copyright Act 1987 (Act 332), Intellectual Property Corporation of Malaysia. Consultation open 3 July to 14 August 2026.
M. D. Algamar and A. B. Munir, Submission to the MyIPO public consultation on the Copyright Act 1987, August 2026.
Quotation with attribution is welcome. The case docket and the answer key accompany this document and may be cited with it.