← Back to recent questions

Mostly for the same unglamorous jobs every other large organisation now uses AI for. Drafting speeches and press releases. Summarising submissions and Hansard transcripts. Translating campaign material into community languages. Crunching polling data. Drafting social media posts. The bit that gets the news coverage, the AI-generated ad and the occasional deepfake of a rival, is real but small. It is the visible tip of a much larger and much more boring iceberg.

The 2025 federal election was the first national vote held in Australia after generative AI became cheap and easy. Both major parties used it. So did several minor parties and some independents. The interesting question is not "are they using it" (yes, all of them, mostly behind the scenes) but "where can a voter actually see it, and what does it mean for the way you read the campaign".

The boring uses, which is most of it

Inside party offices, AI now does the same kinds of things it does in any modern workplace. Drafting first cuts of speeches, op-eds, fundraising emails, social posts, and media talking points. Summarising long documents like budget papers, Senate committee transcripts, and policy submissions so that staffers can brief their MPs faster. Producing community-language versions of leaflets and ads at a fraction of the old translation cost.

There is a second category that voters never see directly but feel the effect of: data work. Parties have always tried to read the electorate. AI lets them do it in much greater detail. Sentiment analysis across thousands of social posts. Clustering of voters by issue using public polling and commercial data. Predictive modelling of which seats are moving and why. Personalised messaging, where an undecided voter in one suburb sees an ad about cost of living and the same voter's neighbour sees one about hospital waiting times. The Australian Strategic Policy Institute's pre-election briefing covers this layer in some detail.

None of this is illegal. None of it is new in kind, only in volume and precision. It is the same direct-mail logic the major parties have used since the 1970s, with sharper tools.

The visible uses you actually see on your phone

This is the bit that the news writes about. In the 2025 campaign, both major parties pushed AI-generated material into social feeds, especially Instagram and TikTok where younger voters live. The Liberal Party produced what was widely reported as Australia's first fully AI-generated political ad, using a synthetic version of ACT Chief Minister Andrew Barr to deliver an attack-line voiceover, and they repurposed a viral AI mouse character ("Tim Cheese") into their own social videos. Labor produced AI-edited videos of opposition leader Peter Dutton intended to make him look weird or out of touch.

Independents joined in for a different reason. ACT independent senator David Pocock released AI-generated clips of the prime minister and the opposition leader appearing to support a ban on gambling advertising, precisely to demonstrate how easy the technology had become and to push for stronger laws.

And then there is the layer the parties themselves did not necessarily make, but which moved around in their orbit. During the campaign, a deepfake of Peter Dutton speaking Mandarin and proposing to ban the Aboriginal flag circulated on RedNote, the Chinese social platform. Nobody has been charged with making it. The reach was small, but the example is useful: in 2026, this is now trivially possible from any bedroom in the world.

What about the deepfake stuff, is it allowed?

Mostly yes, as long as the rules around authorisation are followed. The Australian Electoral Commission has been clear about this: there is no specific ban on AI-generated content in election campaigning. The Commonwealth Electoral Act 1918 has not been updated for the deepfake era. What it does require is that "electoral matter" (essentially, paid campaign material or material from a registered political participant) carry an authorisation statement, the "Authorised by P. Smith, [party], [address]" line at the bottom of a leaflet, post, or video.

That requirement does some work. It means a legitimate political ad cannot pretend not to be a political ad, even if the content inside it is AI-generated. It does not mean the content has to be true. It also does not catch user-generated AI content that does not come from a registered participant; private citizens making deepfakes of politicians for entertainment, satire, or mischief sit in a different bucket.

The separate criminal offence of misleading or deceiving an elector about how to cast a vote does still apply. So a deepfake telling voters that polling day has been moved, or that they have to vote online, could trigger a charge. A deepfake making a politician look silly, or putting words in their mouth that they did not say, almost certainly cannot, as the law stands.

The post-election picture is consistent. The AEC's post-2025 assessment was that AI did not change the result and did not appear to have substantially fooled voters. A subsequent four-person review panel did find that AI was used to spread disinformation during the campaign, including unlabelled images and videos, but again the volume was not enough to move the dial. The next election may look different.

How do I tell what I am actually looking at?

This is the part that matters more than the political question. The trust problem in 2026 is not unique to elections; the world-without-trusted-media question on this site covers the underlying shift, and the spotting AI images question covers the visual side. The election-specific version of the same problem is this.

When you see a political clip in your feed, four things tell you most of what you need to know. The first is the authorisation line. Real party material carries it. Anonymous attack content does not. The absence of the line is not proof of fakery, but it does tell you nobody has put their name to it.

The second is the source. If the clip is on the party's own YouTube channel or pinned on their official social account, it is theirs. If it is only on a third-party page with a name like "AustraliaTruthHQ", you are not watching the party, you are watching someone with a barrow to push.

The third is the cross-check. A reverse image search of a still from the clip will usually find either the original (showing where it came from) or other people who have already debunked it. Google Lens and TinEye both do this well and are free.

The fourth is context. AI-generated political content is rarely designed to inform you. It is designed to make you feel something fast: anger, contempt, fear, ridicule. If the clip is doing that and you are about to share it, that is the moment to slow down.

What about all those watermarks and labels?

The technology industry has built two main systems for marking AI-generated content. Content Credentials (the standard managed by the C2PA, the Coalition for Content Provenance and Authenticity) embed a tamper-evident "nutrition label" into images and videos, recording the camera, the editor, and any AI involvement. SynthID is Google's invisible watermark for content generated by Gemini, Veo, and other Google models.

Both are useful in principle. Neither is reliable enough to lean on for elections yet. Most social platforms strip the metadata that carries Content Credentials when you upload. Most AI tools do not embed them in the first place. SynthID only finds Google's own watermark, not Midjourney's, not Sora's, not the open-source models. The honest position in 2026 is that if you find a credential, treat it as a signal; if you do not, do not treat the absence as proof of anything.

What I would avoid

I would avoid sharing political content from accounts you do not recognise, even when the content matches your existing view. That is precisely the moment you are easiest to fool, and the moment AI-generated content does the most damage. I would avoid trusting any clip in which the politician is doing something out of character, whether it makes you laugh or makes you angry; that combination is exactly what targeted AI material is designed for. I would avoid arguing with a strange clip in a quote-tweet or repost before you have looked at the source channel; you may be the one amplifying it. And I would avoid the assumption that an obvious AI label, like the "Tim Cheese" mouse, means the content is harmless. Surreal humour is a delivery mechanism, not the message.

A simple test before you share

Before sharing any political clip, run a thirty-second check. Open the official channel of the party or politician shown (a quick search of their name plus "official" or "verified" finds it). If the same clip is there, it is theirs and you can react accordingly. If it is not, search for the most distinctive line of the script in quote marks; reputable fact-checkers like AAP FactCheck and RMIT ABC Fact Check will usually have written about it within a day if it has spread. If neither test finds it, treat the clip as unverified and do not pass it on. That single habit, repeated over the run-up to the next election, will do more for the quality of your feed than any amount of detector software.

The verdict

Australian political parties are using AI in roughly the same way Australian businesses are. Mostly invisibly, mostly for drafting and analysis, and the productivity gains are quietly significant. The dramatic uses, the AI-generated ad and the occasional deepfake of a rival, are a small and visible slice that will probably grow. The law has not caught up yet, and the AEC's job in 2025 was effectively to teach voters to think critically rather than to police the content itself.

For a voter, the practical answer is unromantic. Slow down before sharing. Look for the authorisation line. Check the official source. Reverse-image-search the still. Treat surprise and anger as a prompt to pause, not as a prompt to repost. That is the same critical-reading discipline good newspaper readers used to apply in the 1980s. The medium has changed; the discipline has not.