What happened

A man with a terminal illness wrote to his local MP, Liberal backbencher Andrew Hastie, to say he'd decided to end his life through voluntary assisted dying. When Hastie went to reply, Microsoft Copilot suggested he open with "congratulations!" Or "great to hear from you!" Or, worse, "that is wonderful news!"

Hastie told this story at a parliamentary inquiry on Friday, September 18, and he told it on purpose. The hearing is examining the Albanese government's push to bring Anthropic and OpenAI into Australia, and Hastie used the Copilot exchange as his case for the opposite: AI that's run by Australians, not Americans. Same hearing, different witnesses made the opposite argument. Australia's cyber-defence chief said cutting off American models would be "catastrophic." That tension is the actual story here, more than the chatbot's terrible bedside manner.

What's documented vs. what's disputed

Documented:

• Hastie's testimony, reported by Guardian journalist Luca Ittimani, naming the specific phrases Copilot suggested.
• ASD director general Abigail Bradshaw telling the inquiry her agency is "entirely reliant" on US-housed processing and would "dramatically" lose capability if cut off in a conflict.
• Defence chief AI officer Chris Crozier confirming a Google contract to build local, US-disconnected datacentres for the department.
• The government's active consultation on standards that would require AI companies to report breaches and reserve compute for local research.

Disputed or simply unverified:

• The Copilot exchange itself. What's public is Hastie's retelling at the inquiry, not a published transcript or screenshot. I'd want the actual chat log before I fully trust the exact wording, and neither the Guardian nor anyone else has published it.
• Which Copilot deployment or configuration produced the suggestion, and whether it was drafting from the constituent's original email or from a summary of it. Context matters a lot for a failure like this, and none of that context made it into the coverage.
• The implicit claim doing the political work here, that an Australian-run model would have handled this better. Nobody tested that. It's an assumption wearing the clothes of a conclusion.

Timeline

• Undated — a terminally ill constituent emails Hastie's office describing his plan to use voluntary assisted dying.
• Undated — Copilot, used somewhere in the reply-drafting process, suggests celebratory language.
• 2026-09-18 — Hastie recounts the episode at the parliamentary inquiry into sovereign AI. On the same day, Bradshaw and Crozier testify about the operational risk of losing access to US-hosted models, and the Office of AI outlines forthcoming reporting and local-compute standards.

For builders

• If you're building or deploying any assistant that drafts replies to real people in emotionally loaded situations (caseworkers, healthcare intake, HR, constituent services), sentiment-mismatch on grief or serious illness is a documented failure mode. Test for it explicitly with adversarial examples before shipping, don't assume general alignment training caught it.
• If you sell into Australian government or enterprise, the Office of AI's consultation on breach reporting and reserved local compute is worth tracking now. It's pre-legislation, but it's the shape of what's coming.
• Copyright law reform is the actual blocker keeping frontier labs from training models on Australian soil. If that unlocks, it changes the local fine-tuning and hosting landscape faster than the sovereignty debate will.

Further reading

• The Guardian — Andrew Hastie says AI advised him to reply 'congratulations!' to man who planned to end life with assisted dying