DAYS before Johor voted this month, an AI generated poster of a prominent political figure appeared online. It carried words she says she never spoke. There was no label to tell voters it was synthetic, no quick way to take it down, and no offence that squarely fitted. By the time she responded, the ballots were cast. The episode exposed a plain fact: Malaysia has no rulebook for AI in its elections.
The timing matters. The National AI Office is consulting the public on a proposed AI Governance Bill, and the consultation closes on July 31. The sixteenth general election is expected within a year. The rulebook is being written now. The question is whether elections will be in it.
Credit where it is due, Malaysia is not starting from nothing. The National Guidelines on AI Governance and Ethics have existed since 2024, though they bind nobody. The consultation paper published on July 10 is a more serious step. It proposes a Central AI Authority, sets out five governance principles anchored partly in the Federal Constitution, and sorts AI systems into three risk tiers. It requires incident reporting, including near misses, and it creates a regulatory sandbox for innovation. As institutional design, this is a well-shaped skeleton.
I read the paper looking for elections. I found silence. The word election does not appear in it. Nothing addresses campaigns, political advertising or synthetic content. The omission matters because of how the paper defines harm. Its risk framework rests on four baseline categories: Death, Bodily injury, Unlawful deprivation of liberty and Contravention of written law. Now place a convincing deepfake of a candidate against that list. Nobody dies. Nobody is injured. It is far from clear that any written law is broken, because no Malaysian law was written for synthetic content in campaigns. Under the proposed framework, deceiving an entire electorate may not count as harm at all. That is the most consequential omission in the paper.
Fake content damages electoral trust twice. The first damage is obvious. Fake things get believed. A convincing fabrication released at the right moment can move votes before anyone proves it false. In an election, corrections arrive after the votes are counted. The second damage is worse. Once voters know convincing fakes exist, a politician caught on a genuine recording can simply call it AI. Real things get denied. Scholars call this the liar’s dividend. Every scandal and every recording now carries a built-in escape hatch.
These questions are no longer abstract. The prime minister has just unveiled an AI double of himself, built by a Malaysian company. It speaks Malay, English and local dialects, and the benefits are real. A citizen can now question the highest office at midnight, in her own language. The questions themselves are public sentiment no survey can match. But the avatar is an initiative of a political party, not of government. When the avatar speaks, whose words are they? Where does the party end and the office begin? At what point does a leader’s own PR tool become, in effect, a deepfake of himself? No Malaysian law clearly answers any of these questions.
There is a quieter problem beside these. The tool reaches only those with a smartphone, a data plan and the confidence to use them. The elderly voter, the estate worker without coverage and the grandmother in a kampung stand outside the conversation. A tool meant to widen engagement can narrow whose voice is heard. A government that treats an avatar’s chat logs as the voice of the people is listening to the connected and mistaking them for everyone. Disclosure rules must therefore be paired with digital inclusion.
Malaysia would not be inventing anything by acting. The European Union requires disclosure of AI generated content from Aug 2. South Korea bans deepfake campaign material in the ninety days before polling, with penalties of up to seven years in prison. China has required labels on AI-generated content since September 2025. Japan approved election labelling rules this month. The United Arab Emirates requires every candidate in its national elections to declare any AI used in campaigning. Disclosure in politics is becoming the global norm. Malaysia would not be an outlier for adopting it. We would be an outlier for declining.
The same comparisons carry a warning. Brazil banned chatbots from giving voting advice. Six months from its election, they were reportedly still giving it. South Korea employs hundreds of election monitors. They say each new model makes the fakes harder to catch. Japan’s new rules carry no penalties, and its own press doubts they will bite. Rules without enforcement change nothing.
The hardest governance lesson of the year came from America. An American AI company set red lines against mass surveillance and autonomous weapons, and held to them under government pressure. It lost its federal contracts and was designated a security risk for its trouble. What protected it was not goodwill but a court applying written law. When Washington later cut foreign access to its most capable models overnight, governments everywhere learned two things. Voluntary arrangements are fragile. Foreign dependence is fragile too. The things that held were statutes, independent courts and rights of appeal. Our Bill should be built on things that hold.
Three steps would close the gap. None is radical, and the first needs no new machinery at all. Widen the Bill’s harm framework to recognise societal harms, with voter deception named among them. Require labels on AI-generated political material. Put a named accountable person behind every political avatar. Then use the paper’s own architecture. It proposes Sectoral Leads to translate the framework into sector rules. Appoint the Election Commission as the Sectoral Lead for elections. The Bill stays clean, and the campaign work sits where it belongs. Second, the Election Commission should act before GE16 anyway, since no Bill will be law before the next campaign. Issue guidance on synthetic content. Require labelling of AI campaign material. Create a takedown channel that works in hours, not weeks. Third, use the momentum Parliament already has. A bipartisan parliamentary group has called for the Election Offences Act 1954 to be amended before the election, including takedown powers over campaign material on social media. Add synthetic content to the same amendment. One amendment closes both gaps.
None of this threatens legitimate speech. A labelling rule regulates deception, not opinion. Satire and criticism remain untouched. A fabricated video must simply say what it is. The consultation paper’s own first principle commits the framework to protecting human agency. An electorate misled by fabricated content has lost exactly that.
The consultation closes on July 31. Prevention costs less than the first crisis. The Bill’s architects have done serious work, and they have asked the public to complete it. The Public Consultation Feedback can be accessed at https://upc.mpc.gov.my/view-consultation/264.
The most important completion is also the simplest. Remember the ballot.
Dr Chithra Latha Ramalingam is senior lecturer at Monash University Malaysia's School of Business, specialising in governance, law, ethics and climate change policy.