AI chatbots are providing unreliable answers when asked for federal government information, research has found.
When ChatGPT, Microsoft Copilot, and Google AI Overviews were asked 102 questions on topics such as tax, welfare, biosecurity, and citizenship only one in five answers (20.9 per cent) were reliable enough to act on. Almost half (49 per cent) were partly reliable, and 29.7 per cent were unreliable. For 51 per cent of questions, none of the three tools gave a reliable answer and 4.9 per cent of answers contained high-risk errors that could lead someone to miss out on support, breach a legal requirement or make a costly decision.
“The biggest problem wasn’t wrong information – it was missing information,” said digital consultant Lauren Leamon. The research – conducted by Avian – found key facts were often missing, and the omitted detail was often the condition, exception or recent change that decided what the person should do. “These failures matter because people don’t just use government information to learn something. They use it to make decisions: whether to apply for a payment, register for tax, bring medicine into Australia or challenge a dismissal,” Leamon said. “An answer can sound helpful, contain correct facts and still lead someone to make the wrong decision.”
The research found that results differed significantly between tools. ChatGPT gave the most reliable answers (33.3 per cent), followed by Google AI Overviews (16.7 per cent) and Copilot (12.7 per cent). Google AI Overviews – which appear at the top of search results by default – was among the least reliable and cited social media in 57.8 per cent of its answers.
Answers that referenced official sources were far more likely to be reliable (29.0 per cent) than those citing no official source (2.3 per cent). But official sources weren’t always completely accurate. “While not as common, government content did also present some issues,” Leamon said. “In several cases we found official pages that contradicted each other, or that still described programs and rules that had since ended. AI tools repeated these with confidence.” Keeping information accurate and consistent across government websites and services is an ongoing challenge, Leamon said. “Payment rates, thresholds and program details often need to be updated across multiple sources.”
While agencies can’t control how AI tools generate answers, they can shape the information those tools draw on, Leamon said. “We recommend agencies test how AI tools answer questions about their services, check that related official pages agree, make changes and closures explicit, state clearly when figures apply, and prioritise content where errors could cause harm. These steps make government information more reliable for everyone, whether they use AI or not.”



