MALAYSIA has seen the weapon. It has not measured the wound.
On Sept 10, an American technology company published an uncomfortable admission. Anthropic said it had disrupted a commercial platform aimed at Malaysian voters. The operator was traced to a technology firm based in Istanbul. Access was sold as a service, in the way that software is sold. The platform even presented itself as a tool against disinformation.
The report describes voter profiling across all 222 parliamentary constituencies. It describes real census and electoral data, and millions of voter records. It describes targeting organised around race, religion and royalty. It describes roughly a thousand automated X accounts built to evade detection. It describes a fabricated outlet called Malaysia Pulse rewriting real local journalism. It describes Russian and Chinese state media republished with the attribution quietly stripped out. It describes fabricated dossiers aimed at an opposition figure and civil society groups.
Two denials belong in any honest account of this. The report says the operator pursued a contract with our national communications regulator. Anthropic found no evidence that the pursuit ever succeeded. The Malaysian Communications and Multimedia Commission denies involvement and is seeking clarification. It says such activities fall outside its functions and mandate. The matter is now under examination by the regulator, with calls for a full investigation. One observed request sought a million artificial views in support of the prime minister. Those engagement figures were self reported by the operator, Anthropic notes. Anthropic also found no evidence linking the operation to the prime minister or his office. Both denials deserve to be reported as clearly as the allegations.
A third qualification belongs here, and it is the strongest argument against alarm. Anthropic rated the campaign at category two on the Brookings breakout scale. The assets were spread across platforms without breaking into authentic communities. In plain language, the company does not claim this reached real Malaysians. Note carefully what that rating does not say. It does not say the operation was hypothetical or unbuilt. The accounts existed, the outlet published and the dossiers were written. What is unproven is that any of it reached authentic communities.
That qualification deserves a careful answer rather than a dismissal. The operation was caught upstream, during production, before it built any audience. Detection worked, which is a success and not a finding of harmlessness. The next operator may not use a model whose maker publishes threat reports. The capability has already been built, priced and marketed to somebody. An empty audience this time tells us nothing about the next attempt.
So consider carefully what the report actually establishes. It establishes capability. It establishes commercial intent. It establishes that industrial political manipulation can now be bought.
Consider just as carefully what it does not establish. It does not establish that one Malaysian voter changed their mind. It does not even say when the operation ran, or for how long.
That silence is the real problem. We hold evidence of supply and almost no evidence of effect.
The international research is not reassuring, and it is not ours. A study published in Nature last December tested conversations with models advocating candidates. Across American, Canadian and Polish samples, candidate preference moved measurably. The effects were larger than those normally reported for video advertising. Work in Nature Communications found generated messages shifting attitudes on policy. Further work found personalised machine argument outperforming human argument.
None of that research was conducted here, and none of it modelled our politics. Malaysian voters do not read a message only as information. They read who they believe is speaking, and to whom that person belongs. Party loyalty, language, faith and community do much of the interpretive work. A synthetic message rejected by an undecided voter may persuade a committed one. A disclosure that wounds one politician may cost another almost nothing.
We are therefore importing foreign effect sizes into a very different electorate. That is not evidence. That is a guess wearing the clothes of evidence.
In July I argued on this forum that our AI Governance Bill forgot the ballot (The bill that forgot the ballot). The Bill frames harm as physical injury or as breach of existing law. Electoral deception sits outside that frame entirely. I proposed three repairs to it. Recognise the deception of voters as a harm in its own right. Require clear labels on campaign material generated by machine. Name the Election Commission as sectoral lead before the next general election.
I hold all three positions and September has only hardened them. It has also exposed two gaps that I did not press hard enough. The first is evidence. None of those three repairs can be calibrated without knowing what moves Malaysian voters.
Take labelling, the remedy that everybody reaches for first. Nobody has tested what a label does to a Malaysian voter. It may restore trust. It may simply move trust from the person to the party. It may even reward the practised politician who discloses early and appears candid. A mandate built on an untested assumption is a mandate built on hope.
A darker possibility sits beyond the reach of any label at all. As fabrication becomes ordinary, the genuine recording loses its authority. A politician caught on tape may simply call the tape synthetic. Scholars call this the liar dividend. It costs nothing to claim and a great deal to disprove.
The report contains one detail that should trouble anyone who trusts technical safeguards. The model refused several requests, including material it identified as political defamation. The operator then negotiated softer wording and kept building towards the same capability. Safety controls held at the sentence and gave way at the system. Anthropic says it has since tightened its filters against such negotiation.
The second gap is how we came to know any of this at all. Malaysia learned of an operation against its voters from an American company. We did not learn it from our regulator, our Election Commission or any domestic capability. No Malaysian law required Anthropic to tell us a single thing. It published because it chose to publish, and for no other reason. It says it shared intelligence with authorities, without naming which ones. Should the next operator use a provider that publishes nothing, we learn nothing. So add a fourth repair to my three. Require providers to report election-related misuse involving Malaysia to a named domestic body.
So here are the two unglamorous proposals. Before the next general election, Malaysia should fund the measurement and compel the reporting.
This is neither expensive nor slow by the standards of public policy. A controlled study can show voters human, assisted and synthetic political messages. It can measure trust, authenticity, perceived accountability and perceived autonomy. It can test whether disclosure helps, harms or merely relocates our trust. It can test whether effects differ by party attachment, age and language. Interviews can then ask the question that no dataset answers. Who did you think was speaking, and who should answer for it?
Regulators cannot govern a harm they have never sized. Platforms will not act on a risk nobody has quantified locally. Parties will not restrain themselves without a number they fear. Evidence is the cheapest instrument of governance available to us. It costs a fraction of the first scandal that it prevents.
The September report tells us what the machinery is able to do. It cannot tell us what that machinery would do to us. The next election will not be decided by the cleverness of the technology. It will be decided by whether we troubled ourselves to measure its effect.
We can automate the message. We can never automate responsibility for the message.
We were mapped seat by seat, by people who never set foot here. The frightening part is not that it happened. It is that we would never have known.
Dr Chithra Latha Ramalingam is senior lecturer at Monash University Malaysia’s School of Business, specialising in governance, law, ethics and climate change policy.