Australia's most recent national accounts, on Toby Walsh's reading, say the country would have gone into recession were it not for the money being spent on building AI data centers.
The public argument about AI risk is an argument about extinction. Walsh says that part is very improbable, and that the risks worth governing are nearer and duller: cyber attacks, jobs, political discourse, and who teenagers are talking to at night.
"Well, it's very improbable that AI is going to kill us all, but AI is starting to have the capability to do some serious harm, particularly things like the cyber risk."
Walsh is Scientia Professor of Artificial Intelligence at the University of New South Wales in Sydney, and he was speaking on the day the US president dismissed the industry's own calls for a slowdown as a conspiracy.
The full segment is covered here so you can skip it.
Here are the 6 arguments that matter.
👤 Guest: Toby Walsh, Scientia Professor of Artificial Intelligence at the University of New South Wales in Sydney
🎙️ Host: Haidi Stroud-Watts, a Bloomberg Television anchor in Sydney
👥 Also on: Doug Krisner, who anchors the Bloomberg Daybreak: Asia Edition podcast and introduced the conversation
🧩 Other segments: Sunny Bangia, who covers Asia equities for Bloomberg News
📰 Published: 14 September 2026 on the Bloomberg Daybreak: Asia Edition feed
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ length not available
Key Takeaways
The extinction argument is the wrong one to have, and the cyber one is already live
Anthropic's own report describes attempted bioweapon research and attempted drone-swarm building
Data-center construction is what kept Australia out of recession in the latest national accounts
The risk Walsh puts first is employment, not annihilation
His second is what AI does to political discourse, after what social media did to it
200,000 Australian teenagers have had an AI boyfriend or an AI girlfriend
That these founders agree on anything is the signal, because they agree on almost nothing
Oversight should look like drug and aviation regulation, not self-certification
Europe, Canada and Australia together are a bigger economic zone than the US or China alone
1. Not Extinction, But Harm
Stroud-Watts opened by asking whether safeguards are needed, whether it is already too late, and who could realistically impose them given how hard regulators find it to keep up.
Walsh separated the two questions immediately. "Well, it's very improbable that AI is going to kill us all, but AI is starting to have the capability to do some serious harm, particularly things like the cyber risk."
The host's setup came from Australia's own signals agency: its director general had said that day that the country needs an early warning system for AI. Walsh agreed it is something that needs warning about, and noted the agency is lifting its own game.
Stroud-Watts came back to the framing later, having read his notes. "I like that in your notes you're saying the good news is I very much doubt we're all going to die because of rogue AI." Her objection was that this is a very low bar, given what AI could still do to the economy, the financial system and the social fabric.
2. What Anthropic Found
The evidence Walsh reached for is not a forecast but a disclosure from one of the companies asking for a slowdown.
"There was actually a very troubling report from Anthropic last week showing that Iranians had tried to do bioweapon research using the Claude Chatbot, and Russian hackers had tried to build some autonomous swarms of kamikaze drones."
Both examples are attempts by state-linked users against a commercial model, which is why he treats cyber and weapons misuse as the live category rather than a speculative one.
3. AI Is Holding Up GDP
Stroud-Watts put the Australian political problem to him: both sides of government are trying to work out how much of the AI opportunity the economy can take, whether through data-center build-out or by hosting model training.
Walsh's answer is that the build-out is not an opportunity in prospect, it is already load-bearing. "The latest ABS GDP figures said that if it weren't for the spend on building AI data centers and the like, we would have gone into recession."
That is why he does not fault the government for chasing the growth. His point is that the same government has to hold the longer-term risks in view at the same time.
He does not think those risks are national ones. "I mean, AI is an international race. We've got to work out how we coordinate risks." That includes how the United States and China manage them.
On the American position he was blunt: "And it's a bit troubling that Trump is dismissing the risk completely."
4. The Insidious Risks
Asked what actually worries him, Walsh moved away from the catastrophic scenarios entirely.
"I'm actually much more worried about some of the longer-term, more insidious risks, the risks it's going to have on people's employment."
The question he attaches to it: "Is it going to create an underclass of people who are put out of work?"
The second is political discourse, and he argued it from a case already settled. "We've already seen what social media did to that. It polarised us and turned politics into a mudfest."
The third is relationships, and it is where he produced the segment's hardest number. "Increasingly, people are starting to have relationships, synthetic relationships with bots. 200,000 teenagers in Australia have had an AI boyfriend or an AI girlfriend."
His conclusion is a two-part job for government: protecting citizens from these effects while seizing the economic possibilities, rather than choosing between them.
5. Marking Their Own Homework
Stroud-Watts asked whether it is significant that the warnings are now coming from the heads of the AI companies rather than from whistleblowers and activists.
Walsh said the agreement itself is the evidence. "the heads of these AI companies hate each other. They agree on almost nothing. So the fact that they can agree on this is actually quite telling." He named Elon Musk, Sam Altman and Dario Amodei.
He took that as a reason to act rather than to reassure: he said he definitely thinks the pace of development should be slowed.
His objection to leaving it with them is a phrase the host picked up on. "It's probably been a little late in the day for them still to be marking their own homework."
He noted that the companies have themselves suggested independent government oversight auditing what happens in the labs, and that the precedent is ordinary. "We do that with drug companies. We do that with aviation companies. We do that in any other business where there is significant risk to human life."
The conclusion is that voluntary restraint is not a governance model. "I think there needs to be better oversight. We can't rely on their goodwill."
Asked whether the public sector has the expertise to do it, he conceded the hiring problem first. "We're talking about pretty capable, competent people who can command very large salaries in the private sector. So it's going to be hard for government to do that."
But he thinks the institutions already exist. He pointed to AI safety centers set up by the UK government, the US government and Australia: "There's significant expertise. They maybe need some more resources. And they probably need some more law to add to teeth to what they do. But the governments have started to move down the right road."
6. Backing the Middle Powers
The last question was the geopolitical one: if the two biggest economies are both dismissing the concerns, what should a country like Australia aim for?
Stroud-Watts set it up with the two leaders' positions, noting AI discussions are expected when Xi Jinping and Trump meet, and that both have treated the warnings as fear-mongering. "President Trump, in his social media comments, statement saying it's a hoax, he thinks it's a conspiracy, right?"
Walsh's answer was to name an ally. "I'm backing people like Mark Carney. He's completely correct."
His argument is about combined weight rather than moral authority. "Middle powers like Canada, Australia and large powers like Europe can really set the agenda here."
And the number that makes it work: "We're actually Europe plus Canada plus Australia is actually a larger economic zone than the United States or China on their own."
What he wants those countries to do with it is set the norms and standards for everyone else, working together rather than separately.
Bonus Insights
The segment was framed by Doug Krisner rather than by the interviewer. He noted that Trump had dismissed the industry's calls for a slowdown that day, blamed a "sick conspiracy" for the voter backlash on AI data centers and the concern about frontier models, and that China had described the fresh warnings as fear-mongering — all while the industry itself considers common safety standards.
Stroud-Watts's own framing of the stakes went well beyond safety: she put the possible damage as seismic impacts on the economy, on financial systems and on the social fabric, and asked repeatedly who is actually in the right place to decide where the guardrails sit at any given moment.
Walsh's phrase for the oversight problem was picked up on air. Stroud-Watts told him she loved the marking-your-own-homework comparison, which is the line he used for letting the labs assess themselves.
He treats AI safety centers as an existing asset rather than something to be built, which is a different position from the one usually taken in these conversations — the gap he names is law and resources, not institutions.
Walsh's bottom line is that the extinction debate is a distraction from risks that are already here, that Australia cannot regulate AI on its own while its growth depends on building data centers for it, and that the practical answer is a bloc of middle powers writing standards the largest economies have declined to write.
Products, Companies & Tools Mentioned
Anthropic and Claude (The lab whose own report describes attempted bioweapon research and attempted drone-swarm building on its chatbot)
OpenAI (One of the companies whose leadership called for a slowdown, which Walsh reads as a signal precisely because these founders agree on nothing)
University of New South Wales (Where Walsh holds the Scientia Professorship in Artificial Intelligence)
Books & Resources Mentioned
Australian National Accounts – Australian Bureau of Statistics (The GDP figures Walsh cites for the claim that data-center spending kept Australia out of recession)
Anthropic's report on model misuse (Published the week before the interview; Walsh describes its findings but does not name or link it)
Watch the full episode:
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: