For New Zealand voters trying to make sense of election promises, AI offers a tempting shortcut.
Ask it to explain a policy, compare two proposals, or find the costings, and it can save time working through documents. The question is how much of the original evidence survives in the answer.
AUT’s 2026 Trust in News report found that approximately 12 percent of respondents used ChatGPT as a news source. That does not tell us whether they used it to decide how to vote. It does show that AI has joined the mix of sources people consult.
There is plenty it could help with. A business owner looking at a tax proposal might want to know who qualifies, when it would begin, and what remains undecided. An AI tool that searches current documents and directs readers to relevant passages can make those details easier to find.
The difficulty is knowing when an explanation goes beyond its sources. A projected saving can become an expected result. A condition attached to a proposal can disappear from the summary. Generative AI can also produce incorrect information in language that gives the reader little reason to hesitate.
We already have an uneasy relationship with news. New Zealand media organisations are frequently accused of leaning left, and AUT’s 2024 research recorded more allegations of left-wing than right-wing bias in written comments about distrust. Those were respondents’ perceptions, rather than a finding about the political position of news coverage. The 2026 report included complaints about bias in both directions.
It is reasonable to seek another account of an issue. But if we question a newsroom’s judgement, we should take the same interest in how an AI answer was assembled. Which documents did it use? Are they current? Can we find the passage supporting its conclusion?
Then there is the person delivering the argument. An ABC News Verify investigation traced an apparently Australian political commentator on YouTube to a creator in Brazil. AI had been used to give him an Australian accent and alter his mouth movements to match the speech. His channel frequently favoured One Nation, which said it had no association with him, and had accumulated more than three million views.
The investigation did not establish an equivalent operation influencing New Zealand’s election. It showed that a voice sounding comfortably local was not sufficient evidence of a local connection. That is information a viewer might reasonably want before weighing the commentary.
Advertising Standards Authority chief executive Hilary Souter explained where its role begins.
“The ASA’s jurisdiction for advocacy and election advertising is focused on paid advertising. This means the ASA will consider complaints where an advertiser has paid a media company or platform to distribute content to an audience. This includes paid digital advertising, sponsored or boosted social media posts, and other paid election advertising formats.”
“If a creator independently posted political commentary online and there was no payment made to distribute the content, it would generally sit outside the ASA’s remit.”
Souter said AI use in advertising within its jurisdiction would be considered under the Advertising Standards Code’s truthful presentation principle. Jurisdictional questions remained fact-specific and would be assessed case by case.
AI can help us get closer to the detail of an election promise. It can also create a speaker whose apparent identity makes that promise sound more persuasive. As we use AI more, knowing where an answer came from becomes part of understanding what it means.
