Australia has no law against AI lies in federal elections

By Diana Trent · 2026-08-24

What the AEC can and cannot do about AI political content

The Australian Electoral Commission's own submission to the parliamentary committee on artificial intelligence is worth reading carefully. Its powers to investigate or act on AI-generated electoral disinformation are limited to the authorisation requirement under section 321D of the Commonwealth Electoral Act 1918^[1]. That's it. A disclosure label. The content itself, however fabricated or targeted, falls outside the commission's reach.

Section 321D requires electoral material to carry an authorisation: a name, a suburb, a postcode. It was designed for how-to-vote cards and corflutes, not for synthetic audio cloned from a candidate's voice or an AI chatbot quietly working undecided voters in marginal seats in the final 72 hours of a campaign. The AEC knows this. Parliament knows this.

It is currently legal to lie in a federal Australian political advertisement^[2]. There is no federal truth-in-advertising law that reaches political content. South Australia and the ACT have their own regimes; the Commonwealth does not.^[3] The practical effect is that a campaign operative can commission a synthetic video of an opponent making a statement they never made, attach an authorisation box in ten-point type, and distribute it freely. The AEC can confirm the authorisation is present. It can do little else.

The overseas warning signs Australia has already ignored

The New Hampshire primary of January 2024 produced a case regulators and legal scholars will be citing for years. A deepfake robocall using an AI-generated voice indistinguishable from President Biden's was sent to Democratic voters ahead of the primary. The message, documented in the League of Women Voters of New Hampshire's legal complaint, told recipients: "Voting this Tuesday only enables the Republicans in their quest to elect Donald Trump again."^[4] The caller ID was spoofed. The voice was synthetic. The instruction was to stay home.

The call used a fabricated Biden voice and spoofed caller ID to tell New Hampshire Democrats not to vote in the primary^[4]. Consider that same call, not in English but in Vietnamese or Mandarin, targeted at specific electorates in western Sydney, landing the night before polling day. The AEC could check whether it carried an authorisation. It could not ban the content.

South Korea moved faster. Its National Election Commission detected 388 pieces of AI-generated media breaching its law banning deepfake political videos within 90 days of a poll, all ahead of its April 2024 election.^[1] That figure comes from a country with a functioning enforcement mechanism, proactive detection capability, and a legal prohibition that actually exists. Australia has none of those things. The AEC's own submission cited South Korea's experience as context for Australia's current exposure, and parliament read the submission. Parliament has not acted.

Why compulsory voting makes micro-targeted deception unusually potent here

There is a dimension specific to Australia that tends to get lost when the US debate gets imported wholesale. Compulsory voting means the relevant question is not turnout suppression of your opponent's base; it is preference flows among voters who would rather be anywhere else on a Saturday. Those voters are disproportionately low-information, disproportionately susceptible to last-minute framing, and disproportionately decisive in marginal seats under preferential counting.

A synthetic robocall telling New Hampshire Democrats to stay home works through non-participation. An AI campaign in Australia does not need to stop anyone voting. It needs to shift a few thousand second preferences in a handful of seats. The required sophistication is lower. The required investment is minimal. A Nature study published in December 2025 found that conversations with AI chatbots during real elections shifted voter preferences by more than conventional video advertising.^[5]

A campaign or a third party could deploy a chatbot that engages undecided voters in marginal seats, responds to their specific concerns, addresses their individual hesitations, and steers their preferences, simultaneously across thousands of conversations, in the final days of a campaign when the authorisation requirement is even harder to enforce. The AEC has no tool for this. Section 321D does not reach it.

What Steggall proposed and why both major parties have stalled

Independent MP Zali Steggall introduced a bill designed to close these gaps. According to her proposal, the bill would prohibit the authorisation of certain electoral and referendum matter that is inaccurate and misleading to a material extent, including material that has been modified using digital technology, such as 'deepfakes'; and require electoral and referendum matter that has been created or modified using digital technology (including AI) to carry a statement to that effect^[6].

That is a sensible two-part architecture: a prohibition on materially misleading AI-altered content, and a disclosure requirement when AI was used at all. Neither provision is radical. Both have analogues in state law for conventional advertising. Neither the Coalition nor Labor has moved to pass it.

Both Labor and the Coalition have run campaigns in an environment where misleading political advertising carries no federal sanction. Both parties have built campaign infrastructure premised on that freedom. Requiring disclosure of AI use would constrain both of them. Prohibiting materially misleading AI content would open the question of who adjudicates "materially misleading," which is a power neither party wants to vest in an independent body. The bill sits. Parliament rises. An election approaches.

What disclosure legislation needs to say, and when it needs to pass

If Steggall's bill or something like it does not pass before the election writs are issued, Australia goes to the polls with a framework designed for corflutes and how-to-vote cards. The AEC has told parliament as much. The parliamentary library has documented the absence of federal truth-in-advertising law. The overseas evidence is on the record. The Nature study is peer-reviewed. The New Hampshire complaint is public.

Workable legislation needs a definition of AI-generated or AI-altered electoral content broad enough to capture synthetic audio, video and text, but precise enough to survive a constitutional challenge on political communication grounds. It needs a mandatory on-material disclosure statement, machine-readable as well as human-readable. It needs a prohibition on content that is materially misleading as to the identity or conduct of a candidate, an enforcement mechanism with teeth rather than a referral pathway that ends at the authorisation box, and it needs to be law before the campaign period begins.

Both major parties know the window is narrow. Their calculation appears to be that the risks of AI disinformation are diffuse and the risks of disclosure legislation are concentrated on their own campaign operations. That calculation will look very different after the first major AI-generated electoral incident on Australian soil, and Steggall's bill remains before parliament in the meantime.

Diana Trent

Diana Trent writes about regulation, competition and the law as it meets technology. She reads the judgments and the regulator filings that most people skip, and finds the story in them.