On 5 October 2026, Digital Minister Gobind Singh Deo told the Dewan Rakyat that the proposed AI Governance Bill will be tabled in early 2027, not 2026. The bill is still being drafted. It is meant as a cross-sector framework that sets obligations by the potential harm of an AI system rather than by industry alone, and it would complement existing sectoral laws. Gobind tied it to the National AI Action Plan 2026-2030 and said more powerful, autonomous systems would need stronger safety controls, testing, monitoring and human oversight. On 10 September he had said the draft was complete and hoped to table it by Q1 2027 at the latest.
The case for moving fast
The strongest argument for urgency is that Malaysia already regulates AI harms, just reactively. On 3 January 2026, the Malaysian Communications and Multimedia Commission (MCMC) said misuse of AI tools to manipulate images of women and minors was an offence under Section 233 of the Communications and Multimedia Act 1998. It also said it would investigate X users and call in X's representatives (MCMC media statement). Enforcement after the fact, under a 1998 statute, is a poor substitute for clear duties set in advance. A regulator that improvises has more discretion than one working from a dedicated statute, and developers cannot plan around discretion.
That is a fair case for a dedicated law. It does not follow that the speed of tabling matters more than the quality of drafting.
What the design gets right
The July 2026 consultation paper from the National AI Office (NAIO) was released on 10 July, with written feedback due 31 July. According to Baker McKenzie's summary, it proposes three tiers. Tier 1 prohibits systems developed or deployed with intent to cause harm. Tier 2 covers systems with a risk of harm and attaches risk assessment, documentation, human oversight and monitoring duties. Tier 3 leaves systems with no foreseeable material harm to a baseline responsibility with lighter compliance.
Three features are defensible on pro-innovation grounds. First, obligations attach to function and harm, not to a technology label. Second, the paper exempts personal, family and household use, as well as systems used solely for national defence or security. Third, it proposes AI sandboxes aimed at SMEs and startups, and an enablement function that issues guidance, templates and training. A regulator that is also required to help firms comply is better than one that only polices them.
The harm test is also narrower than in many frameworks. Baker McKenzie reports it is assessed against four categories: death, bodily injury, unlawful deprivation of fundamental liberty under the Federal Constitution, and contravention of written law.
Where the draft is still vulnerable
The fourth category is the problem. "Contravention of written law" is open-ended. Malaysia has a large body of written law, including provisions that criminalise broad categories of speech. If any AI system whose output could contribute to a legal violation can be classed as high risk, the tiering stops being a proportionate filter and becomes a way to import every existing content offence into AI compliance. For a publication that values free expression, that is the clause to watch. If the intention is to regulate AI systems rather than the content they generate, the statute should say so in its text, not leave it to ministerial remarks.
A second gap is the intent line between Tier 1 and Tier 2. Prohibition by intent is easy to state and hard to prove. An open-weight model released for general use may be misused by someone else. The bill should say plainly that a developer's liability does not extend to downstream misuse it did not intend and could not reasonably foresee. The consultation paper defines Developers and Deployers as separate roles, though one organisation can hold either or both, so duties should be assigned by role with care.
Third, there is enforcement. The proposed Central AI Authority would run AI safety, investigation and enforcement, and AI enablement, working with existing Sectoral Leads that may be delegated specific powers. Combining supervision, investigation and enablement in one body invites conflict of interest unless there is clear internal separation and a right of appeal.
Lessons from the Online Safety Act
Communications Minister Fahmi Fadzil said in July that the AI bill would be tabled after discussions across ministries, agencies and Cabinet were complete, and he pointed to the Online Safety Act 2025 process of forums, committees and engagement as the model (The Vibes). That model deserves scrutiny as well as praise. The Online Safety Act [Act 866], in force since 1 January 2026 per MCMC's own page, carries short deadlines. Mayer Brown reports a one-hour acknowledgement window for user reports and penalties of up to RM10 million for failing a Part III duty, with possible personal liability for directors and officers. If the AI bill copies that enforcement posture, compliance cost will fall hardest on small domestic developers, which is the opposite of what the Action Plan wants.
What the delay should buy
A slip of a few months is not a crisis. A badly drafted law that is hard to amend is. Three things would make the extra time worthwhile:
- Define "harm" so that contravention of written law counts only where the AI system is the proximate cause of a serious statutory violation.
- Publish the second-round draft, not only the first-round paper, with a comment window longer than the roughly three weeks the July consultation ran.
- Write the sandbox and SME proportionality rules into the statute itself, not into later regulations.
The core idea, regulating by harm rather than by sector, is sound. Whether it is also pro-innovation depends on definitions that have not yet been published in final form. The ministry has bought itself time, and it should use it to narrow the draft, not just finish it.