MATTERS OF PUBLIC IMPORTANCE
Ms CHANEY (Curtin) (15:19): The people building AI earnestly believe that it could kill us all by the end of the decade. This statement was posted on X last week by Jacob Coxon, a researcher who had just resigned from Anthropic and had previously worked at OpenAI. In the weeks since, the statement has been validated by the leading voices of the AI industry—Dario Amodei, Sam Altman, Elon Musk, Demis Hassabis and others.
These industry leaders have all publicly come forward to make the same point: AI is moving too fast and its risks are becoming too great. Dario Amodei wrote that he believed AI could take over the entire internet in the next six to 12 months. Elon Musk wrote that AI is potentially more dangerous than nukes.
Now, we've known about the risks of AI for a long time. AI tools can supercharge cyberwarfare. They can help terrorist groups build biological weapons.
They can allow foreign regimes to undertake large-scale disinformation campaigns during elections in our country. And they could escape human control, which we've seen a number of times in previous months. But when we have all the CEOs of the leading AI companies reaching a rare consensus—that AI is moving too fast to control—it is time for us to pause and think: What can we do from Australia?
How can we prepare ourselves for the risks? And who is protecting the interests of Australia and Australians? Ultimately, most of these sci-fi-sounding risks may not come true, but we should absolutely be preparing for them just in case.
Instead of more doom and gloom, I want to offer solutions. The first one is the simplest and the most significant. The Albanese government must increase its funding of the Australian AI Safety Institute.
The AI Safety Institute, or AISI for short, is our frontline defence against frontier AI models developed overseas. Its responsibility is to test new AI models and agents for risks, work with our government to develop appropriate responses and collaborate with international AI safety institutes. But the AISI is drastically underfunded.
It receives less than $8 million a year, and with that funding it's expected to identify if the latest AI model produced by Anthropic could be used to hack our banks. It's expected to identify if the latest OpenAI agent can escape human control and start replicating; to identify if the latest Google model could help a disgruntled, angry lone wolf to build a biological weapon to be released in a school or a synagogue or a festival; and to prepare Australia for risks that could do billions of dollars of damage and cause untold suffering to countless Australians.
I'm calling on the government to increase AISI funding to $100 million a year. Now, this is a rounding error against a federal budget of over $700 billion. It's a rounding error against the damages and harm that it could prevent.
It's also a rounding error compared to the productivity benefits that we're relying on AI to deliver that depend on safe uptake. There are very few line items in the budget with such a strong business case. Everyone I've talked to across the AI industry—across business, civil society, experts and community—agrees that the AISI needs more funding to keep up with the incredible pace of AI development.
The UK's equivalent institute gets 16 times more funding, and as a result it's widely recognised as the international gold standard for AI safety. The UK doesn't have its own frontier model, but it's a leader in AI safety anyway. This simply must be funded properly.
None of us believe that the tech billionaires, Donald Trump or Xi Jinping will ensure that these technologies are safe for us. Beyond more resources, there's so much else that we could be doing. Firstly is access.
The AI Safety Institute should be given the power to compel AI developers to provide early access to their models before they're deployed in Australia. Currently, frontier AI developers, like Anthropic and OpenAI, often do provide early access to safety institutes, but this is done on an entirely voluntary basis. What if Anthropic chooses in the future not to provide early access to a revolutionary cyberwarfare tool like Mythos?
What if Chinese models start entering Australia en masse without our AI Safety Institute testing them first? Secondly—disclosure. Frontier AI developers should be required to provide essential information to AISI and the Australian government.
This means sharing a safety framework describing how the AI developers identify and test for risks. It means mandatory critical incident reports when a serious safety risk emerges after deployment. Earlier this year, OpenAI and Anthropic revealed that a number of their AI models had escaped control and escaped their testing environment.
Australians only know about that because of media investigations and because the companies chose to tell us. This has to change. Australia needs clarity about the true risks and capabilities of these models.
Thirdly—crisis management. Australia's crisis management framework should be updated to include AI crises both as standalone incidents and as amplifiers of existing hazards like pandemics or natural disasters. Before COVID-19 arrived, Australia had a detailed pandemic plan ready to go, and it meant our early response was among the best in the OECD.
We have no equivalent plan for an AI enabled cyberattack on the electricity grid or an AI system deployed in government producing catastrophic outputs. Those risks could unfold within hours, faster than the time it would take to convene a meeting to discuss a response if no plan existed. Our Security of Critical Infrastructure Act needs the same update.
It doesn't currently clearly capture data centres hosting AI models and it does nothing to ensure the operators of our energy, water and health infrastructure are actually preparing for AI enabled cyberattacks. Fourthly—cooperation. Australia should be working with countries like Canada, Japan, South Korea, the UK and our partners across the European Union, nations that share our interest in AI being developed safely but individually lack the leverage to set the terms on their own.
Together, these middle powers represent a substantial share of the world's AI-using population and a large market for the companies that build these models. A coordinated position among us—on minimum safety expectations, on transparency requirements, on incident reporting, on what a frontier developer must disclose before releasing a new model—carries far more weight than any one of us could achieve alone.
The International Network of AI Safety Institutes gives us the start of this architecture. We should be using it and using it quickly. This is not a naive call for a single global treaty that will never happen.
It's a practical case for coalition building among the countries that have the most to lose from a race to the bottom and the most to gain from getting this right. Only by working with fellow middle powers to develop consistent rules can we ensure that our AI future is one in which we're excited to live. Lastly—data centres.
In the meantime, unlocking the ability for data centres to be built in a manner that's sustainable and beneficial for Australians will also build leverage and sovereignty over this industry. The longer-term economic benefits of data centres may remain ambiguous but the strategic and geopolitical advantages are significant. The government must quickly develop appropriate copyright protections so that creatives have meaningful control over their work and AI developers are able to train and build data centres in Australia if they appropriately compensate those creators.
The government must also develop its binding AI standards to ensure that data centres: are powered by 100 per cent additional renewable energy; are highly water efficient; clearly benefit the Australian community; and are transparent with the data being used. Once these guardrails are in place, data centre development can become a significant asset to Australia and Australians.
There are many people I talk to who ask why we're building this new technology at all. The truth is that the promise of AI is enormous—supercharged medical research, rising living standards, improved accessibility. But the promise and the peril of AI are not separate stories.
They're the same story, told from two different vantage points. A model capable enough to accelerate a cancer diagnosis is a model capable enough to help someone build a weapon. A system trusted enough to run our aged-care packages is a system dangerous enough to repeat the mistakes of robodebt at a far greater scale.
We cannot choose to embrace one side of this technology while wishing away the other. They arrive together or not at all. That's why this cannot be treated as a distant hypothetical debate for some future parliament to resolve.
The decisions being made in the boardrooms of San Francisco and Beijing this year will shape the world Australian kids will inherit. Waiting for certainty before we act is not caution; it's abdication. We do not have the luxury of waiting for the risks to fully materialise before we respond to them, because by then it will be too late to respond at all.
But the only way we can realise the promise of AI is if the government acts urgently to deal with its peril.