When AI Drafts Go Wrong and Why Canberra Wants Its Own System
Andrew Hastie told a parliamentary inquiry on 18 September that Microsoft Copilot suggested he reply to a terminally ill constituent with "congratulations!", "great to hear from you" or "that is wonderful news!" The Guardian
The man had written to Hastie to say he planned to end his life under voluntary assisted dying laws. Hastie said he fed the letter into Copilot, Microsoft's AI drafting helper, to prepare a draft reply. The suggestions did not fit the context.
Hastie raised the story on Friday to point to the limits of American-run software. He called for AI run by Australians. The Albanese government seeks to bring Anthropic and OpenAI to Australia.
The broader context here is that the politics is easier than the plumbing. That is a neat line from Hastie. It is also a complicated one for both sides of the House, because it still leaves open how Canberra builds, funds and powers a home-grown system at the speed agencies say they need it.
Australian Signals Directorate Director-General Abigail Bradshaw described the dependence in direct terms. She said the agency needs advanced AI every day to hold off "highly capable malicious cyber actors". She said the directorate is "entirely reliant" on processing housed in the United States and would "dramatically" lose capability if cut off in a conflict.
Defence Department chief AI officer Chris Crozier set out the plan to fix that. He said the government had contracted Google to build linked datacentres across Australia, large facilities packed with computers that run AI systems, disconnected from the US. He said Australia wants AI computing power, apps, data and decision-making all based in Australia rather than using offshore tool sets.
The Office of AI, based in the prime minister's department, set out the regulatory side. It told the inquiry that bringing AI firms under Australian law would help local law enforcement deal with threats on platforms. The government is consulting on standards that would force AI firms to report problems and breaches and potentially set aside some computing power for local research. Under planned rules, large datacentres in most states will have to consult communities, build away from schools and homes, and pay for new renewables.
In my view, electorate staff will read the Copilot story less as a tech tale than as a workflow warning. Offices already handle large volumes of sensitive mail. Voluntary assisted dying letters, veterans' matters, immigration cases. A draft tool that cannot read context is not a time saver. The question for MPs' offices and ministerial units is where the human check sits, who signs off, and what is logged when a draft goes wrong.
Looking at what this means for policy, the inquiry has now paired a human failure with a structural one. Hastie gave the personal example of offshore models getting it wrong. Bradshaw gave the capability example of offshore computing being unavailable at the worst moment. Crozier and the Office of AI offered localisation as the answer, through onshore datacentres, onshore legal presence, incident reporting, reserved research computing, and planning and energy rules. For practitioners, the test is delivery. Contracts, legislation, planning approvals and power deals will decide whether sovereign AI is an operating position or a press release.
