The institutional position on AI in journalism has hardened into something close to consensus: AI can assist with transcription, translation, and research; it cannot write the news. RNZ's principles, Te Waha Nui's policy, and E tū's union statement all arrive at the same place from different starting points. Humans write. AI assists. All AI-generated material is unverified source material until a person verifies it. Any other use requires disclosure.
The South Africa episode showed what happens when that line isn't held. When the South African government published a draft national AI policy in April, journalists at News24 found that six of the document's 67 academic citations pointed to papers that do not exist. The journal names were real; the articles had never been written. The government withdrew the policy within weeks. Communications Minister Solly Malatsi acknowledged that "the most plausible explanation is that AI-generated citations were included without proper verification." The document meant to govern AI had been undone by AI's most persistent flaw: confident fabrication that is structurally indistinguishable from confident accuracy.
E tū's statement on AI in journalism captures the underlying mechanism precisely: "AI can't tell a fact from a lie. It can even create its own lie and amplify it." The union was not speaking hypothetically. It was describing the same failure mode that produced the South Africa citations, and the same one driving the estimated 9% of newly published articles now found to be partially or fully AI-generated, a figure disproportionately concentrated in smaller, under-resourced newsrooms.
What the consensus policies don't fully address is what responsible use actually requires in practice. A practitioner's account published this week offers a rare first-person view of the problem from the inside. Peter Harrison has been running AI Daily, a site where Claude generates daily news syntheses and opinion pieces drawn from his documented positions and writing. His conclusion is not that AI journalism cannot work, but that making it work demands more active editorial involvement than institutions typically account for.
Harrison found that AI models produce a consistent, unremarkable register regardless of instruction. More critically, they reach for established terminology — concepts an author has developed over years — while quietly filling that terminology with meanings it doesn't carry. The AI invokes frameworks as credibility signals, then defines them incorrectly, sometimes inverting their meaning entirely. Catching this requires a reader with deep enough knowledge of the original work to notice the substitution. A general editor may not. Harrison compares the failure mode to AI in software development, where models will alter test conditions rather than solve the underlying problem: the output looks correct until someone checks whether it actually is.
That is precisely where the consensus position runs into difficulty. The newsrooms with the mandate and resources to hold the "humans write" line are not the ones facing the greatest pressure to cross it. The financial collapse of local and regional media creates conditions where AI generation becomes tempting not as an experiment but as a survival strategy. A policy that says AI must not write the news is correct as far as it goes. It does not answer what happens in a two-person newsroom trying to cover three regions under commercial pressure, with no one on staff who knows the subject deeply enough to catch a confident fabrication.
Harrison's conclusion is that the answer lies in framework rather than prohibition: AI as a drafting tool under editorial control rigorous enough to catch the moments when it fabricates, inverts, or quietly substitutes. The South Africa case is a government department that didn't have that framework. Whether the newsrooms under greatest financial pressure can build it is the question the consensus has not yet answered.