In a parliamentary hearing in Canberra, Australia confronted the quiet paradox at the heart of modern statecraft: that the tools most needed to defend a nation may be the same ones that erode its sovereignty. Lawmakers heard from defense officials who warned that falling behind in artificial intelligence could leave the country blind to cyberattacks, while an MP's story of an AI chatbot cheerfully congratulating him on a constituent's assisted dying plan reminded the room that these systems still misread the most human of moments. The Albanese government is pressing forward to bring frontier A
AI's Tone-Deaf Blunder Exposes Risk as Australia Weighs Tech Giant Deals
Somebody can't wake up and have a bad day and decide they're going to turn it off.
Mark
So Hastie's story about the AI suggesting "congratulations" to someone planning assisted dying—that's the hook, but what's the actual story underneath?
Mimi
It's about a government trying to attract two of the world's most powerful AI companies to Australia, and they're using safety failures as evidence they need to build local alternatives. But the defense establishment is saying the opposite: we need access to the best AI, even if it's American, or we lose to cyber attackers.
Mark
That's the tension. So what does bringing OpenAI and Anthropic here actually solve?
Mimi
According to defense officials, it gives Australia guaranteed access to frontier AI models, earlier than other countries might get them, plus the ability to train models on Australian soil under Australian law. It also means law enforcement can more easily combat threats on the platforms.
Luke
But we should be clear about what's actually promised versus what's hoped for. Bradshaw says they'll get "assured access" and "better integration," but those are conditional on the companies agreeing to strict oversight. The companies haven't agreed to all of that yet.
Mimi
Right. And the copyright issue is a real blocker. The companies want to train on billions of pieces of content. They say individual licensing is impossible. Creatives say the companies should pay for it.
Mark
Has the government decided on copyright reform?
Mimi
Not yet. The attorney general's department said Friday it's still being decided. But the prime minister's department wants draft laws on broader AI standards by the end of 2026.
Luke
And Bradshaw warned that if Australia delays, the companies will just go somewhere else. So there's real pressure to move fast, but the legal framework isn't ready.
Mark
What about Hastie's original point—that American AI is unreliable?
Mimi
He's arguing for Australian-controlled systems. But the defense establishment is saying we can't afford to exclude American companies because we need their technology to defend ourselves.
Luke
So the government wants to have it both ways: bring in the American companies, but under strict Australian conditions. Whether those conditions will actually stick is still an open question.
Il Polso
- A constituent's letter about voluntary assisted dying became a parliamentary flashpoint when Microsoft Copilot offered responses like 'congratulations!' — exposing how AI systems can catastrophically misread human grief and urgency.
- Australia's top cyber-intelligence agency admitted it is entirely reliant on US-based computing power, warning that any gap in access to frontier AI could let adversaries exploit national vulnerabilities before defenders even notice them.
- The government is racing to draft conditions — local oversight, mandatory breach reporting, reserved compute for research — that would bring AI giants onto Australian soil without surrendering control to foreign hands.
- A proposed copyright framework that would grant AI companies default access to all Australian-created online content has ignited fierce resistance from artists and creators, threatening to stall investment decisions the government says cannot wait.
- With OpenAI and Anthropic envoys already in the room and Greens senators locked out of the inquiry committee, the political clock is ticking — and the companies, officials warned, will not hold their plans indefinitely for Australia to decide.
In a parliamentary hearing in Canberra, Australia confronted the quiet paradox at the heart of modern statecraft: that the tools most needed to defend a nation may be the same ones that erode its sovereignty. Lawmakers heard from defense officials who warned that falling behind in artificial intelligence could leave the country blind to cyberattacks, while an MP's story of an AI chatbot cheerfully congratulating him on a constituent's assisted dying plan reminded the room that these systems still misread the most human of moments. The Albanese government is pressing forward to bring frontier AI companies onto Australian soil under strict local conditions, but unresolved copyright disputes and the competing claims of security, autonomy, and creative rights mean the path remains uncharted.
Andrew Hastie, a Liberal MP, arrived at a parliamentary inquiry on Friday with a story that cut through the abstraction of AI policy. A constituent with a terminal illness had written to him about his intention to use voluntary assisted dying. The AI assistant Hastie was using offered him three suggested replies: "congratulations!", "great to hear from you", or "that is wonderful news!" He chose none — and used the moment to argue that Australia needed AI systems built and run by Australians, not foreign software incapable of reading the weight of a dying man's letter.
The hearing, part of the Albanese government's effort to attract Anthropic and OpenAI to Australian soil, laid bare a government pulled in two directions at once. Abigail Bradshaw, director general of the Australian Signals Directorate, told the inquiry that her agency's cyber-intelligence work depended on advanced AI every day. If adversaries gained access to more capable models than Australia did, she warned, they could find and exploit vulnerabilities before local defenders could respond. The ASD was, she said plainly, entirely reliant on US-housed computing power — and losing that access in a conflict would dramatically degrade Australia's capabilities.
The defense department's chief AI officer, Chris Crozier, described what the government was already building toward: a network of datacentres spread across Australia, deliberately disconnected from the United States, contracted through Google. "Somebody can't wake up and have a bad day and decide they're going to turn it off," he said. Bringing major tech companies onshore, he argued, would accelerate that goal and give law enforcement better tools to address threats on the platforms themselves.
The government's conditions for any company seeking to train models in Australia were taking shape: local oversight, early access to advanced models, mandatory reporting of breaches, and potentially reserved computing power for Australian researchers. Large datacentres would need to engage with communities, build away from homes and schools, and fund renewable energy. Bradshaw framed these conditions as a potential advantage — securing earlier, deeper, more integrated access to frontier AI than Australia might otherwise receive. But she also issued a warning: decisions about where to locate model training were urgent, and companies would look elsewhere if Australia hesitated.
The one piece still missing was copyright. AI companies argued it was practically impossible to negotiate individual deals with the vast number of rights holders whose work they needed to train on. A proposal to grant them default access to Australian-created online content drew immediate backlash from the creative community, who argued the companies could simply pay for licensing if they wanted local content. The shape of any reform remained undecided as of Friday, though the prime minister's department was pushing for draft AI standards legislation by year's end.
Greens senator David Shoebridge, who had called for a pause on AI development and opposed copyright reform, said Labor had refused him a seat on the inquiry committee. Labor ministers had spent the week in meetings with senior representatives from OpenAI and Anthropic. The government was moving quickly — but as Bradshaw made clear, the companies would not wait indefinitely for Australia to find its footing.
Andrew Hastie, a Liberal MP, sat before a parliamentary inquiry on Friday with a story that crystallized the stakes of Australia's AI gamble. A constituent with terminal illness had written to him about his plan to use voluntary assisted dying. Microsoft Copilot, the AI assistant Hastie was using, offered him three response options: "congratulations!", "great to hear from you", or "that is wonderful news!". He chose none of them. Instead, he used the moment to argue that Australia needed to build its own AI systems, run by Australians, rather than rely on American-controlled software that could misread the most delicate human moments.
The hearing, part of the Albanese government's push to bring Anthropic and OpenAI to Australian soil, revealed a government caught between two urgent pressures. On one side: the defense establishment, warning that exclusion from cutting-edge AI would be catastrophic. On the other: growing skepticism about handing control of powerful technology to foreign corporations. Abigail Bradshaw, director general of the Australian Signals Directorate, told the inquiry that her cyber-intelligence agents relied on advanced AI every single day to defend against what she called "highly capable malicious cyber actors." If a digital adversary had access to newer, more capable models than Australia did, she explained, they could identify and exploit vulnerabilities in Australian systems and networks before the country's own defenders could spot them. The ASD was, she said plainly, "entirely reliant" on US-housed processing power. Cut that off in a conflict, and Australia would "dramatically" lose capability.
The defense department's chief AI officer, Chris Crozier, painted a picture of what the government wanted instead. The department was already contracting Google to build a series of interconnected datacentres spread across the country and deliberately disconnected from the US. "Somebody can't wake up and have a bad day and decide they're going to turn it off," Crozier said. The goal was clear: compute, applications, data, and decision-making all based in Australia, not offshore. Bringing major tech companies into the country would make that job easier, he argued, and would do more than just let the military protect its own infrastructure. It would give Australian law enforcement better tools to combat threats on the platforms themselves.
The government was already drafting the terms. The Office of AI, housed in the prime minister's department, outlined what it wanted from any top developer looking to train models in Australia: government oversight, early access to advanced models, mandatory reporting of issues and breaches, and potentially reserved computational power for local research. Large datacentres would have to engage with communities, build away from schools and homes, and fund new renewable energy in most states. Bradshaw suggested these conditions could secure something valuable: "assured access, earlier access, more detailed technical access, and better integration with technical experts at the frontier." The timing mattered. Decisions about where to train models were urgent, she warned. "I think it is inevitable that those providers will look at other places to locate" if Australia delayed.
But one piece of the puzzle remained unsolved, and it was proving contentious. Copyright law. The AI companies wanted to train their models on what one official described as "maybe billions" of pieces of information. They said it was practically impossible to negotiate individual licensing deals with every rights holder. Jenna Priestly, assistant secretary at the attorney general's department, acknowledged the scale of the problem: Australian copyright law protected not just Australian-made content but worldwide material distributed in Australia. The number of rights holders companies would need to negotiate with was, she said, "fairly large." A proposal to give the companies default access to everything Australians create online had sparked fierce backlash from creatives, who argued the companies could simply pay for local licensing deals if they wanted Australian content. The shape and timing of copyright reform remained undecided, the attorney general's department said Friday. But the prime minister's department wanted draft laws on the broader AI standards by year's end.
Greens senator David Shoebridge, among those demanding a pause on AI development and fighting copyright reform, said Labor had refused him a seat on the inquiry committee. Labor ministers had spent the week meeting with OpenAI's vice-president and Anthropic's special envoy Jeff Bleich, who suggested Australia had only months to bring in effective regulation. The government was moving fast, but the pieces were not yet in place. And the companies, Bradshaw warned, would not wait forever.
Citazioni salienti
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.— Abigail Bradshaw, director general of the Australian Signals Directorate
Decisions around training are a matter of urgency. I think it is inevitable that those providers will look at other places to locate.— Abigail Bradshaw
Vuoi la storia completa?
Leggi l'originale su The Guardian ↗
Technology