Artificial intelligence, explained.
Sam Altman, Chief Executive Officer of OpenAI, briefs the Security Council meeting on Artificial Intelligence and International Security, 23 September 2026 (Ariana Lindquist/UN Photo)
AI arms control will fail without the means to check on commitments, just as nuclear disarmament would.
Subscribe to The Informer for monthly expert analysis, and to Events for advance notice of visiting world leaders and distinguished guests.
You may unsubscribe from Lowy Institute newsletters at any time. For information on our privacy practices and how to unsubscribe, see our Privacy Policy.
AI needs trust – and Australia can provide the verificationAI needs trust – and Australia can provide the verification
Competitive pressures in the AI race make it difficult for any one frontier lab or nation to slow down development unilaterally for fear of granting an advantage to competitors – no matter the calls for restraint (Opens in new window).
Instead, a slowdown must be coordinated and collective, codified in an international agreement to pace frontier AI development.
The major obstacle to such an agreement is the ability of each signatory to be confident that the other signatories are adhering to the rules of the agreement. Without the ability to verify compliance, each signatory is incentivised to assume the worst of others and cheat in turn.
The same challenge of verification was the central problem of Cold War nuclear arms control. Analogous technical and governance tools must now be built for AI.
Australia should invest in the capability to help build them. For example, an agreement might cap the amount of compute that could be spent on training new AI models. Verification would let all signatories confirm that no one had crossed that cap.
In the same way that a state without nuclear weapons rescued the test ban, states without frontier labs can be political leaders in AI verification.
Australia is particularly well-placed to lead here. It has no frontier AI models of its own, and while there are efforts to develop domestic models, these models are far behind the frontier.
For verification, that is an advantage. A verifier that also hosts frontier labs is less trusted: it could go easy on its own labs, or use an inspection as an opportunity to spy on rivals’ methods. By virtue of having no labs of its own to favour, and no frontier program that a rival’s secrets could enhance, Australia has no such incentives.
Australia’s alliance with the US means that it is unlikely to be viewed as a neutral party equally trusted by Beijing and Washington. Yet its few conflicts of interest do position it well as a credible contributor to a multilateral verification body.
It has also done this before. In 1996, Australia led the adoption of the Comprehensive Nuclear-Test-Ban Treaty (Opens in new window). Talks stalled because negotiation by consensus allowed a single state to block the text indefinitely. Australia broke the stalemate by taking the draft text directly to the United Nations General Assembly.
In the same way that a state without nuclear weapons rescued the test ban, states without frontier labs can be political leaders in AI verification. Australia presently hosts 21 International Monitoring System facilities, the third most of any country. These include seismic, hydroacoustic, infrasound and radionuclide stations, covering not just Australia but the Indian Ocean, the Southern Ocean, the Pacific approaches and Antarctic territory – regions with few alternative hosts. Thirty years on, the initial investment has paid a dividend in technical capability and political standing.
Warramunga Seismic and Infrasound Research Station located near Tennant Creek in the Northern Territory, Australia, run by ASNO to monitor for underground and atmospheric nuclear tests (CTBTO)
Verification institutions already exist in Canberra. The Australian Safeguards and Non-Proliferation Office (ASNO) regulates the government’s nuclear and chemical test ban obligations, and researches new verification techniques in collaboration with Australian universities, the Commonwealth Scientific and Industrial Research Organisation, and the Australian Nuclear Science and Technology Organisation. This does not mean that ASNO should verify AI treaties – rather, that the bureaucratic and political know-how is already in place and an AI equivalent does not need to be reinvented.
On the AI side, the institutions are younger and smaller, but growing. The AI Safety Institute (AISI) was announced (Opens in new window) in December 2025, with $29.9 million of funding over four years. It is an advisory body not supported by any AI-specific regulation, but it exists and is plugged into frontier labs and its international counterparts. A memorandum of understanding between the Australian government and Anthropic commits the company to joint safety evaluations with AISI, in conjunction with institutes in the US, UK and Japan. ASNO’s remit covers nuclear and chemical verification, while AISI has no explicit verification mandate. Nothing joins the two. That gap is where AI verification capability would sit.
The hardware is in Australia, too. Verification methods need testing under real operating conditions. Australia already hosts 160 data centres, with 90 more in development (Opens in new window). One concrete research avenue is power monitoring; early results (Opens in new window) suggest that patterns of electricity use can help distinguish AI training from other kinds of computation, a possible path for checking whether a data centre has run training banned by an agreement.
Running these experiments at scale in working facilities is a natural next step. New facilities also offer an opportunity to design verification in from the start. For example, the government could support voluntary pilots that install verification instrumentation and give researchers access.
Canberra should fund a dedicated AI verification research program, jointly scoped by ASNO and AISI, and open to universities, start-ups and data-centre operators. Novel verification techniques will not generate the political will for the US and China to forge an agreement to pace frontier AI. But credible, scalable methods of verification make it easier to come to such agreements. Building the techniques and the political momentum is a long-term task that Australia should start now.
About the author
Tom Kimpson
Dr Tom Kimpson is a Research Fellow at the University of Melbourne and a Fellow of the Machine Alignment & Theory Scholars (MATS) program in Berkeley. His research covers AI verification, mathematical biology and astrophysics.
The most-pressing world events explained by Lowy Institute experts and global contributors, in your inbox, every Thursday.
You may unsubscribe from The Interpreter at any time. For information on our privacy practices and how to unsubscribe, see our Privacy Policy.
