Malaysia’s AI law will be judged by enforcement, not enactment

Passing the AI Governance Bill will be the easy part. The real measure will be whether a wronged citizen can trace responsibility, a board can demonstrate oversight and a small developer can comply without a legal department.

An artificial intelligence (AI) agent that can read your email, approve payments and update records is no longer merely a tool you consult. It is an actor operating with your credentials.

When it errs, the question of who answers for it cannot be left to a vendor’s terms of service.

This is why Digital Minister Gobind Singh Deo’s announcement that an AI Governance Bill is being drafted is timely.

The minister told the Dewan Rakyat on Oct 28 that the government was aware of risks such as agentic misalignment and shutdown avoidance. These findings must be viewed in context. They emerged from controlled simulations, not live deployments, and do not establish that AI is becoming uncontrollable.

The more credible lesson is that systems given broad permissions and open-ended goals can take actions their operators never intended. The answer lies in sound engineering discipline: restricted access, logging, tested shutdown mechanisms and human escalation. Governance should be in place before an incident, not introduced in its aftermath.

Accountability must remain with people. Developers are responsible for design and testing. Vendors must be transparent about capabilities and limitations. Deploying organisations are responsible for configuration, permissions and monitoring. Boards must determine their risk appetite. Responsibility should follow control.

“The algorithm did it” cannot be a defence. AI governance is not merely an IT matter; it is a matter of organisational governance. The minister rightly highlighted impersonation and false information. Cloned executive voices, synthetic identities that defeat onboarding checks and fabricated videos undermine the verification on which banking, public administration and elections depend.

Detection tools alone will struggle to keep pace with these threats. Law and technology must work together through stronger authentication, content provenance, out-of-band verification for high-value instructions, mandatory incident reporting and public awareness. Sound regulation can also support innovation by giving investors and customers greater predictability.

The minister said the Bill would take a risk-based approach, with obligations determined by potential harm rather than industry label. That is sensible.

Malaysia can draw lessons from the Organisation for Economic Co-operation and Development (OECD) AI Principles, the National Institute of Standards and Technology (NIST) AI Risk Management Framework and the European Union’s tiered model. These offer useful guidance, not templates to be copied wholesale.

Malaysia also has foundations on which to build, including the Cyber Security Act 2024, the amended Personal Data Protection Act, the AI Code of Ethics and the MY-AISafe initiative.

Gobind’s pragmatic approach to data sovereignty also deserves support. The plan is to establish a Sovereign AI Cloud under Budget 2027, with “meaningful, reasonable and proportionate” control.

In my view, sovereignty cannot be achieved simply by placing servers on Malaysian soil. It depends on who holds the encryption keys, who has administrative access, which legal jurisdiction governs the provider, how resilient the supply chain is and whether essential services can be restored if a supplier fails or is compelled to withdraw.

Malaysia should protect what is genuinely sensitive while remaining open to cross-border collaboration in other areas.

The harder work, however, lies ahead. The Bill has not been tabled. The minister has said it is expected to be tabled by the first quarter of 2027 at the latest, and its text is not public.

Five tests will determine whether it works.

First, it must translate principles into controls. Ethics codes express aspirations, technical standards establish measurable requirements, and legislation makes obligations enforceable.

For high-risk systems, I would expect pre-deployment impact assessments, independent testing, human override, auditable logs and prompt incident reporting. Reassessment should be triggered whenever models, data, integrations or permissions change. An agent approved in January may behave quite differently by June.

Risk tiers should reflect the potential consequences of failure, the degree of autonomy and the sensitivity of the data involved. The Bill must also clarify how it will interact with existing sectoral regulators, so that banks and insurers do not face overlapping obligations without a clear lead authority.

Second, institutions need the capacity to enforce the rules. The Malaysia AI Safety Institute and AI Malaysia will need technical expertise, operational independence and effective coordination with regulators, universities, practitioners and civil society.

Success should be measured by outcomes: testing coverage, the speed of incident reporting, time taken to remedy failures and measurable improvements in organisational risk management. The number of guidelines issued is no measure of effectiveness.

Third, compliance costs must be proportionate. A multinational cloud provider and a 10-person start-up have different resources and risk profiles. Phased implementation, practical compliance guidance and shared testing facilities would protect the public without pricing Malaysian innovators out of the market.

Vague or excessive requirements would stifle local innovation, while weak safeguards would erode public confidence. Both would represent regulatory failure.

Fourth, literacy matters. Legislation alone cannot create an AI-literate society. The minister’s willingness to coordinate the development of school learning modules is welcome. Directors and senior managers also need training. They need not be technical experts, but they must know what a system can access, who can stop it and how to determine whether it has failed.

Fifth, the Bill must be enforceable in practice. It must clearly define responsibility, investigative powers, notification duties, remedies and penalties, and empower a regulator to act on them. Malaysia is right to act, but legislation without effective enforcement will offer little protection.

Passing the Bill will be the easy part. The real measure will be whether, a few years from now, a wronged citizen can trace responsibility, a board can demonstrate oversight and a small developer can comply without a legal department.

Meeting those tests would make Malaysia a trusted AI nation, not merely a busy one.

Murugason R. Thangaratnam is a cyber security practitioner. 

The views expressed here are the personal opinion of the writer and do not represent those of Twentytwo13.