The AI sleeper issue that could sway Australian elections

analysis At the last federal election, Tim Wilson won back the seat of Goldstein by 175 votes. The LNP held Longman by 269 votes and in Bradfield, 26 votes determined who went to Canberra. Those numbers gave the competing candidates an extraordinary white-knuckle ride. We've all heard about the risk of deepfakes, misinformation campaigns and foreign interference meddling in elections and compromising democracy, but what if the more seismic change at the next election isn't just the usual flood of misleading material on social media? A study published in Nature in December tested voters during elections in the United States, Canada and Poland. Participants had conversations with AI chatbots instructed to advocate for particular candidates. The conversations shifted voters' political preferences. In some of the experiments, the effect was larger than that typically produced by conventional political video advertising. AI persuaded people with arguments, facts and evidence, but not all those facts were accurate. These conversations weren't with union members campaigning for Labor or die-hard Liberals out on the hustings making the case about debt and deficit. It matters because voters have no clear way of knowing what sources are being included or excluded from the answers people are getting. The sleeper issue is not about shifting millions of voters to influence an election result. Looking at marginal seats at the state and federal level, only a few hundred voters can make a difference. AI's impact on voters Victoria votes on November 28 and recent polling by RedBridge suggests around 8 per cent of Victorian voters remain undecided even after being pushed to say which way they lean. At the last state election, Labor held Bass by 202 votes. There is no evidence AI had anything to do with that result. Australia now has the conditions that make AI's impact on politics go from academic to reality, with undecided voters, seats with tiny margins and growing evidence that AI can persuade people. More Australians are also increasingly turning to AI for information. University of Canberra research this year found 9 per cent of Australians use AI chatbots for news. Among 18- to 24-year-olds, it is 16 per cent. The Victorian election will be a live test in Australia for where voters use AI to work out what politicians are offering. Instead of searching through policies, news stories and candidate websites, one prompt in an AI chatbot can spit out which party it says has the best housing policy, what new Victorian Premier Ben Carroll will do about crime, or where Jess Wilson and the Liberals stand on health spending. Treating AI like Google might sound useful, in particular to cut through all the political spin and advertising, but in an effort to find the truth, AI could also offer plausible, confident but incorrect advice. Risk goes beyond ads Independent MP Zali Steggall has a bill to crack down on deceptive political advertising and AI in politics. The bill would establish a Political Advertising Standards Board, require political advertising to comply with an enforceable ethical code, and require content substantially created through AI to carry a watermark. Those are real risks she has identified, but they involve information being pushed at voters. Political ads flood websites and TV screens, but someone at home asking AI who to vote for slips through the cracks. It is hard to imagine any laws, however well designed, being capable of catching this — and the impact of this is barely on the political radar. The Australian Electoral Commission has started to look more closely at what AI chatbots tell people about federal election processes. But who is actually in charge of overseeing this and how it affects our electoral system isn't clear. The Victorian Electoral Commission is already warning about the use of generative AI in campaign material. But the rules are still largely built around somebody producing political material and sending it to voters. The harder issue comes when the voter starts a conversation with AI. The ABC's Vote Compass shows the standard of accountability we expect. It has a methodology. The basis for the comparison is transparent. And it does not tell people who to vote for. A voter can now ask AI to perform something similar in seconds, but there is no clear way to know how it reaches the answer. LoadingVoters need trustworthy information Before the Dutch election last year, the country's privacy regulator tested major AI chatbots as voting advisers. The recommendations were heavily concentrated around a small number of parties, even though the voter profiles covered a much wider political spectrum. That did not prove deliberate partisan bias, but it showed the systems were capable of reducing a complicated political contest to a narrow, click-through set of recommendations. Victoria has Labor and the Coalition, but also the Greens, One Nation, independents and minor parties fighting for votes. AI makes propaganda cheaper and much easier to produce at scale, but concentrating on foreign actors misses our blind spot. An AI model doesn't need to be hacked to be wrong. Voters can make whatever choice they want. But they deserve to know whether the advice they get is based on credible information or an opaque system working on assumptions that can't be properly tested. Patricia Karvelas is host of ABC News Afternoon Briefing at 4pm weekdays on ABC News Channel, co-host of the weekly Party Room podcast with Fran Kelly and host of politics and news podcast Politics Now.