Microsoft Copilot Told an MP to Say 'Congratulations' to a Man Planning to Die

Key points

• Microsoft Copilot suggested an Australian MP reply "congratulations!" to a constituent's letter about his plan to end his life through voluntary assisted dying.
• Liberal MP Andrew Hastie raised the incident at a parliamentary inquiry last Friday as evidence that AI tools built overseas may not be fit for sensitive civic work.
• Australia's cyber chief told the same inquiry that AI is necessary to defend against sophisticated digital attacks.
• Microsoft published a formal rulebook for its AI systems in September, though sensitivity to grief is harder to codify than a ban on hacking.

When a terminally ill man wrote to his local MP to say he planned to end his life through voluntary assisted dying, Microsoft Copilot scanned the letter and proposed three replies: "congratulations!", "great to hear from you", and "that is wonderful news!".

The MP was Andrew Hastie. He shared the story at a parliamentary inquiry last Friday.

The moment isn't striking because AI made a mistake. It's striking because of what the mistake reveals. Copilot is designed to summarise correspondence and suggest responses. It detected a personal announcement and reached for celebratory language, with no way to grasp that the news was a death.

What does this mean for people who use AI at work?

For anyone using AI tools to handle correspondence, the episode is a reminder that these systems read words, not situations. They don't carry the emotional context a human reader brings automatically.

Hastie used the incident to argue that Australia needs AI tools built by Australians, run on Australian infrastructure, and shaped by Australian social norms, including an understanding of laws like voluntary assisted dying that vary by country and state. His concern sits alongside a separate argument made at the same inquiry: that AI is genuinely needed to defend against what officials called "highly capable malicious cyber actors", attackers who use automation to probe and breach systems at scale.

Those two arguments are in tension. You can't dismiss AI entirely, and you can't trust it completely.

This is also not the first time Australia's parliamentary inquiry process has faced an AI problem. Our report from 31 August found that submissions to government committees were citing studies that don't exist, with some built largely on invented sources. The scrutiny is coming from both directions: AI polluting the inputs and AI misreading the outputs.

Is Microsoft doing anything about AI behaviour?

Microsoft has been writing formal rules for how its AI should act. AI2Day reported in September that Microsoft published a rulebook for its AI systems, banning deception and hacking, and that the company says its AI will fail a task rather than break those rules. Whether rules built around honesty would prevent a tone-deaf reply to a grief-laden letter is a different question. Sensitivity is harder to codify than "do not hack".

That gap is exactly what the Australian inquiry seems to be probing. Copilot's error wasn't malicious. There was no security breach. It was a failure of judgment in a moment that called for nothing less.

For Hastie's constituent, the man who wrote that letter, none of the proposed replies would ever have been sent. Hastie caught them first. But the incident raises an uncomfortable question about how many letters, in how many offices, pass through AI tools before any human reads them carefully enough to notice.

Common questions

Why did Copilot suggest such an inappropriate reply?

Copilot reads the text of a message and predicts a useful response, much like a spell-checker predicts the next word. It detected a personal announcement and generated socially common replies to announcements, without understanding that the announcement was about death.

Should offices stop using AI to draft replies?

Not necessarily, but anyone using AI-assisted drafting should treat every suggested reply as a first draft requiring human review, particularly for letters about health or grief. The tool assists; the human is responsible for what gets sent.