This new SAA report (available in PDF here) goes through the key issues around Australia’s national decision making around AI, the business demands of OpenAI and Anthropic and how this all stacks up in the wider international and technological picture around AI, policy, security and law. Here’s a flavour of what the report covers.
The national security advantages of undertaking frontier AI training in Australia are not dependent upon access to the datasets involved in copyrighted materials. From a policy perspective, the enhanced national security benefits of frontier AI training can readily co-exist with existing copyright protections. Loosening Australia’s copyright laws would benefit the AI companies’ business models — it is a commercial demand, not a national security requirement.
National security and copyright live in separate worlds. National security protects Australia’s sovereignty, people, institutions and critical infrastructure against espionage, cyberattack, terrorism and military aggression. Copyright protects the works of creators. Weakening protections in one to strengthen the other is tenuous at best and demands a compelling explanation. It is hard to see how training on lyrics, literature or visual art would materially improve AI protection of sovereignty, counter-terrorism or critical infrastructure — and the burden of proof sits with those seeking the exemption, not with requiring copyright holders to prove its absence.
Cyber is the foundation of all other national security uses of AI — without information and system security, everything else is vulnerable. AI-enabled hacking is already happening, from OpenAI’s agents escaping onto the internet and hacking company and government sites to broad cyber disruptions in Japan, South Korea and France. Cyber capability depends on particular datasets: computer science, coding, cryptography, technical details of systems, vulnerabilities and exploits, threat libraries and industry cyber intelligence. It does not depend on copyright-protected creative works. Wider national security uses — counter-terrorism, intelligence, surveillance and targeting, autonomous weapons — need broader datasets than cyber but not copyrighted works. These are private datasets locked up in government holdings and in proprietary databases inside companies, militaries and intelligence agencies – they are not affected by copyright.
Open weight models give Australia options and argue against monopoly lock-in. Open weight models from Moonshot AI (Kimi), Alibaba (Qwen) and DeepSeek are closing the performance gap with the top US models — the UK’s AI Security Institute found the gap had narrowed from 6–10 months through 2025 to 4–7 months by July 2026. These are already powerful tools, and the future is likely to be a hybrid environment: the US itself is not locking out Chinese AI, and General Purpose Artificial Intelligence (GPAI) models will be in perpetual competition with focused, purpose-built models tailored for particular attacks, systems and targets. Open weight models are akin to the Android operating system for phones, while OpenAI and Anthropic’s closed models are more like the proprietary Apple approach to its IOS operating system.
Making Australia a permissive data-harvesting jurisdiction will attract more than ‘friendly’ firms. It would also draw malign actors — cyber hackers and organised crime — and may draw negative reactions from close partners (America, the EU, UK, Japan). The EU AI Act shows the international movement is in the opposite direction to what the US firms want Australia to take.
‘Models can’t distinguish copyright material, so no payment regime is workable’ — is this right? No. The international trend is towards protecting data from AI harvesting and seeing data as a valuable resource. This will require the companies to find ways to identify and pay content creators, and established models for negotiating payments already exist in sectors like music streaming.
The argument that AI is powerful enough to cure cancer, break the world’s cyber security and solve frontier maths, but simply cannot work out how to pay content creators, looks contradictory — and very hard to take seriously.
Other governments and entities like the EU will create approaches to force payment. Australia can remain outside these arrangements or partner with like-minded countries.
Don’t we need to change Australian laws to get OpenAI and Anthropic to work with our security and intelligence agencies? No. Right now, agencies like ASD or the Australian military could access OpenAI’s or Anthropic’s frontier models on satisfactory commercial terms — and it is very likely the US Government would support access to top-end (if not absolute best) models for Australian national security agencies, particularly if Australia is willing to pay the requested price. Those models will be developed in the US regardless.
Foreign training in Australia doesn’t deliver ‘sovereign AI’. There is an assumption in public debate that ‘AI model training in Australia’ creates sovereign Australian AI. It doesn’t. The models would remain proprietary corporate products, and our military, ASD or any other agency would still be just customers. Using Australian copyrighted data in ‘local’ training would deliver no more cyber or national security benefit than using the same data for training in the US. The secret sauce is collaborative model development and company-supported deployment for focused security uses — and embedding ASD, military or agency personnel in the companies’ US training operations could deliver the same benefits without any domestic law change.
This is the Exec Summary to SAA’s new report, commissioned by ARIA.
Read the full report in PDF at this link.