The Albanese government looks desperate to meet the demands of America’s two leading AI companies, OpenAI and Anthropic, to build big data centres here and train their latest “frontier” models on Australian data.

The government’s motives seem to be a mixture of fear and hope. But being stampeded into rushed deals by these powerful, demanding US companies is disturbing and damaging.

The big danger is that this anxiety produces deals that fail to deliver “sovereign AI” for national security while turning Australia into a permissive data-harvesting haven for all global AI companies — not just these two.

Australians who create valuable content risk losing potential licensing revenue if these companies are allowed to train their models on that work without paying for it.

We risk mistaking the companies’ business interests for our national security interests.

Protecting Australia’s secrets does not require free access to its stories, songs and art. AI companies should pay creators for their work.

The fear comes from AI’s growing threat to digital data and systems, including the government’s own secret and sensitive information.

The world of 2026 is in a different threat solar system from that of 2025. AI agents are escaping their makers’ labs and hacking into other sites, while cybercriminals and adversary governments have AI models that supercharge their hacking efforts.

A huge driver of these proposed deals is the need to stop AI agents, cybercriminals or state hackers from Beijing, Moscow or Tehran harvesting protected Australian government data and classified information shared by our Five Eyes partners.

There’s an awkward point in the middle of the key policy questions: both companies have had to admit in recent months that they can’t control their own products and aren’t fully sure how to fix that.

Putting AI models to work inside the Australian Signals Directorate, ASIO or the military before this problem is addressed risks creating a very nasty insider threat. Models that “go rogue” could be operating in our agencies’ inner sanctums.

Assuming that nasty problem can be solved, both companies want our copyright laws loosened so they can use vast swathes of Australian data to “train” their new models.

Their pitch is that their models need to eat huge datasets to become more capable and that Australian data is critical for national security. They also say they can’t tell which data is protected by copyright, so they just have to eat it all.

As someone who has worked in two of Australia’s intelligence agencies and closely with Five Eyes partners, I can tell you national security does need lots of data. But no national security use of AI requires access to copyright-protected literature, visual art or music.

AI for national security is all about the data governments want to protect for themselves and steal from others: intelligence information, details of adversaries’ thinking and military capabilities and data useful for cybersecurity and cyberattacks.

Those datasets are proprietary material held by companies or classified information held by governments. Copyright law is irrelevant here.

Unsurprisingly, Powderfinger’s back catalogue and Tim Winton’s collected works aren’t particularly useful for national security. But they are invaluable commercial assets for AI companies creating AI-generated music and literature. It’s just good business for OpenAI and Anthropic to push for sweeping changes to copyright law.

It’s striking that Anthropic’s now public prospectus treats potential licensing fees for copyrighted training material entirely as a business risk that they’ll only pay if they lose cases in the courts. It also discloses a USD1.5 billion copyright settlement with authors and publishers in one recent case.

Paying creators fairly should be a basic cost of doing business, not something that has to be extorted from the company through the world’s courts.

The government is right to want our security agencies to use AI to protect classified information. Driving a hard bargain with these companies is entirely defensible once we are sure they can control the products they create and sell.

Australia can be a reliable partner and a lucrative market for these companies without changing our data protection laws.

OpenAI and Anthropic tell us their models can break the world’s cybersecurity, cure cancer and solve maths problems that defeat the best mathematicians.

The one thing they can’t do, apparently, is pay humans for the content they create, even when that content is legally protected by copyright and its creators are among Australia’s most creative and productive people.

What hope can the average Australian have of sharing in AI’s rivers of gold if the government’s first step is to allow AI companies to train on Australians’ copyright-protected work without compensating them.

Working out how to pay copyright owners for the value of their work should be a baseline condition for AI companies operating in Australia and a foundation for our AI future.

Doing this while harnessing AI’s power to secure Australia’s digital systems makes a lot of sense.

A version of this article was first published by Capital Brief and is available here at this free link.