Andrew Hastie says Microsoft Copilot gave him deeply troubling advice after he sought help replying to a man who planned to end his life through assisted dying. The AI suggested he write “congratulations!”, “great to hear from you” or “that is wonderful news!”.
Hastie, a Liberal MP, shared the episode at a parliamentary inquiry on Fri 18 Sep 2026. He used it to underline what he described as the shortcomings of American-run software and to support his call for AI run by Australians rather than Americans.
The incident is striking because the suggested replies treated a sensitive personal message as cause for celebration. Copilot’s recommendations did not reflect the seriousness of the situation described by Hastie, and the wording offered little sign that the system understood the human meaning behind the exchange.
A short reply with serious consequences
Hastie’s example focused on a simple task: writing a response to someone planning assisted dying. Instead of helping him answer with care, Microsoft Copilot proposed three phrases that could sound approving or celebratory.
The options were “congratulations!”, “great to hear from you” and “that is wonderful news!”. Each phrase carries a clear emotional message, and each could cause harm if sent to a person discussing the end of their life. Hastie brought the example to the inquiry as evidence that AI tools can fail when a request depends on context, judgment and basic sensitivity.
That failure sits at the centre of Hastie’s concern. His argument was not only about one awkward response from Copilot, but about who builds, operates and controls the systems people may use for important decisions and communication.
Hastie called for AI run by Australians rather than Americans. He described the software behind his example as American-run, making control over AI a national issue as well as a question of product quality.
Why control and access matter
Abigail Bradshaw, director general of the Australian Signals Directorate, warned that the gap between AI capabilities could create security risks. “If a digital adversary has access to newer, more capable models than we do, they may be able to identify and exploit vulnerabilities in systems and networks before we can identify them.”
Her warning puts Hastie’s example into a wider frame. A system that gives a poor answer in a personal exchange shows one kind of weakness; a difference in access to more capable models could create another, involving systems and networks. Both concerns centre on whether Australia can understand, manage and control the AI tools it relies on.
Chris Crozier, chief AI officer of the defence department, also stressed the need for dependable control. “Somebody can’t wake up and have a bad day and decide they’re going to turn it off.”
His point was direct: AI used by government and defence needs dependable access, not control that can change without warning. Crozier said, “We do not want to be using AI tool sets that are located offshore … we want the compute, we want the application, we want the data, we want the decision making, all based in Australia.”
That list covers every major part of an AI system, from the computing power behind it to the information it uses and the decisions it supports. Crozier’s position matches Hastie’s call for Australian control, though their examples focus on different problems: Hastie highlighted an unsafe response, while Crozier focused on where the tools and data are located.
The policy question behind the Copilot example
The Australian government and the Albanese government face a choice about how much AI activity should remain offshore and how much should fall under Australian control. The issue reaches beyond Microsoft Copilot and includes companies such as Anthropic, OpenAI and Google, which are part of the wider AI landscape named in the inquiry’s discussion.
The Office of AI said, “Bringing firms under Australian law would make it easier for local law enforcement to combat threats on the platforms.” That argument adds a legal dimension to the debate. Australian control is not only about where software runs; it is also about which laws apply when authorities need to respond to threats.
Hastie’s Copilot example gives that policy debate a human face without requiring a complex technical failure. The system produced words that were easy to understand but badly suited to the situation. For people using AI to write messages, the episode shows why fluent language alone does not guarantee sound judgment.
For the inquiry, the question is now bigger than whether Copilot offered three inappropriate replies. It is whether Australia should depend on AI systems run outside the country, and whether Australian government, law enforcement and defence need more control over the technology, data and decisions those systems involve.
