Can AI Train on Your Work in Malaysia? What Copyright Permission Means

About this guide: This explainer uses an incident reported on 2026-08-04 as a starting point. The legal rules below are general; they do not decide anyone's guilt, liability or individual case.

Two-track checklist, decision tree and fill-in wording

AI permission and proof pack

1. Creator track: keep original files, drafts, export history, publication dates, the suspected output, lawful screenshots, URLs and the platform terms that applied when you uploaded the work.
2. AI-user track: give every dataset item a row showing its source, apparent owner, permission, permitted training use, exclusions, territory, duration, onward use and expiry.
3. Apply three checks: public access is not permission; paid access is not training permission; and an existing licence must cover the actual intended use.
4. Ask in writing: ‘May [user] use [work] for [training purpose] for [duration/territory], with [payment] and [onward-use terms]?’ Do not imply that silence is consent.
5. For AI-assisted output, log your drafting, selection, arrangement, editing and other creative choices so your human contribution can be explained.
6. If use is suspected, preserve evidence and identify the work and provable use before considering licensing, takedown or legal options. Get Malaysian advice for commercial datasets, claimed exceptions or real disputes.

What happened

On 4 August 2026, Deputy Domestic Trade and Cost of Living Minister Fuziah Salleh told the Dewan Negara that copyrighted material used for AI model training is subject to approval from the copyright holder. Bernama reported her as saying that licensing and royalty arrangements are to be agreed between the relevant parties under the Copyright Act 1987. She referred to licensing, assignment and collective management organisations as ways rights can be managed.

The same report said Malaysia had not yet recorded a court case involving copyright infringement arising from AI use. The government had not completed a dedicated impact study on jobs and earnings in the creative industries and was continuing stakeholder engagement before considering new policy or legal frameworks. Free Malaysia Today later explained the practical point for creators: putting a photograph, song, illustration or article online does not by itself remove copyright protection.

This was a current official position delivered in Parliament, not a new Act, amendment or Malaysian court judgment. That distinction is important. It gives creators and AI businesses a clear compliance signal — do not assume online availability equals permission — while leaving fact-specific questions about a particular dataset, technical process, licence or statutory exception to be resolved under the existing law.

The legal insight: what law applies

MyIPO's official copyright guide says copyright is an exclusive right that lets an owner control a creative work for a period under the Copyright Act 1987. Protected categories include literary, musical and artistic works, films, sound recordings, broadcasts, published editions and derivative works. The guide also says sufficient effort must make the work original in character and the work must be written down, recorded or otherwise put into material form. An idea, procedure, method of operation or mathematical concept is not protected merely as an idea.

Permission is not a single universal switch. A licence can define which work may be used, for what purpose, in which territory, for how long and on what payment terms. An assignment transfers rights and is different from a limited licence. The deputy minister referred to section 27 and collective management organisations when discussing licensing and royalties. Before copying material into a training dataset, the practical legal question is whether the person doing so has authority covering that actual use.

The government's statement should not be stretched into a claim that every interaction with an AI system is automatically infringement. Ask: which protected work is involved, who owns it, what did the system do with it, what permission covers that use, what can be proved, and might a legal exception apply? A legal exception is a limited rule that may permit a particular act without the owner's permission. Training systems and datasets differ, and no reported Malaysian AI copyright case had been identified by the government as at 4 August. It would therefore be unsafe to promise how a court would decide a particular technical workflow.

There is a second issue: whether an AI-assisted output is itself protected. Fuziah said the present framework protects work showing originality through human effort, skill and intellectual contribution, while a fully AI-generated work without significant human creativity does not qualify. That does not make every work made with an AI tool unprotected. It means the creator should be able to explain and preserve the human choices — drafting, selection, arrangement, editing or other creative contribution — rather than assuming a prompt alone settles authorship.

How does this impact me?

If you are a creator, publishing your work on a website or social account is not the same thing as abandoning copyright. However, the platform's terms may contain a licence governing hosting, display, processing or other uses. Keep the version of the terms that applied when you uploaded the work and do not describe a platform licence more broadly than its actual words.

If you build, fine-tune or procure an AI system, keep a source register for training material. Record where each item came from, who appears to own it, the permission or licence relied on, permitted uses, exclusions and expiry terms. A vendor's general assurance that data is 'public' or 'open' is not a substitute for checking the underlying rights and contract.

If you suspect your work was used without permission, preserve your original files, drafts and publication dates, the suspected output, relevant prompts or screenshots if lawfully available, and the provider's current terms. Do not send an accusation before identifying the work and the use you can actually prove. A lawyer can then assess ownership, evidence, the appropriate recipient and whether a takedown request, licensing discussion or court remedy is realistic.

What this incident teaches us

The first lesson is to label the source of a legal proposition. A deputy minister's parliamentary answer is authoritative evidence of the government's current position, but it is not the same as statutory text or a binding judicial decision. A careful guide can report that position without pretending Malaysia already has a complete body of AI copyright case law.

The second lesson is that consent must be specific enough to be useful. 'The owner agreed' tells you little unless the record shows which material, which acts and which commercial or technical purpose were covered. Clear licences reduce uncertainty for both creators and developers; vague provenance merely moves the dispute to a later date.

Finally, copyright is only one layer. Platform contracts, confidentiality, personal-data rules and other rights may affect a dataset even where a copyright licence exists. Conversely, proving that an output resembles a work is not by itself a complete infringement analysis. Separate each legal issue and avoid treating one permission as clearance for everything.

The verdict

Malaysia's message to AI users is now difficult to ignore: online creative work is not automatically free training material. Get and record permission that covers the intended use. Creators should preserve evidence of both their original work and their human contribution. But keep the conclusion measured — the August statement applies existing copyright principles; it did not enact a new AI copyright offence or decide how every future Malaysian case will turn out.

What can I do if this happens to me?

• Keep original project files, drafts, export history and dated publication records for work you create.
• Save the platform terms and privacy or content licence that applied when you uploaded the work.
• Before training or fine-tuning a model, create a dataset register linking every source to its permission and use limits.
• Use a written request: ‘May [user] use [work] for [training purpose] for [duration and territory], with [payment] and [onward-use terms]?’ Do not treat silence as consent.
• If you find a suspected use, preserve the exact output and surrounding evidence before contacting the provider or owner.
• For a commercial dataset, a claimed legal exception or a real dispute, obtain Malaysian legal advice instead of relying on a general AI policy summary.