Australia's AI watchdog has a big job and a small budget

The Australian AI Safety Institute is up and running, testing frontier models with a A$29.9 million budget. It can study the labs, but it cannot tell them what to do.

Australia's newest government body is already at work. The Australian AI Safety Institute, announced in November 2025, began operating in May 2026 and has started testing frontier AI models with technical partners, according to Assistant Minister for Science, Technology and the Digital Economy Andrew Charlton.

The institute sits inside the Department of Industry, Science and Resources. Its job, as the government describes it, is threefold: analyse and test new AI models and applications, help regulators and agencies respond to emerging risks, and shape how AI is developed and governed, including internationally. It has signed agreements to share testing methods and expertise with its counterparts in the United Kingdom and Canada, and it is working with the Gradient Institute on risks from AI agents and with the CSIRO on human oversight of AI systems.

Leading it is Dr Kate Conroy, appointed in May 2026 as general manager of AI governance. She has worked on AI risk rules for the Queensland government and on responsible AI for the Royal Australian Air Force.

"Safety is not the brake on the AI opportunity, it is the enabler," Charlton told the AI Safety Forum in Sydney in July 2026.

The money is modest. The institute has A$29.9 million over four years, according to the ABC, or roughly A$7.5 million a year. The United Kingdom launched its equivalent in 2023 with £100 million in initial funding, and the companies it would be testing spend billions a year developing their models.

"The big challenge for the AI Safety Institute is being able to provide meaningful support and advice given the level of funding it's got," Toby Walsh, chief scientist at the UNSW AI Institute, told the ABC.

There is also the question of teeth. The institute is an advisory and testing body inside a government department, with no regulatory powers of its own. It cannot fine a company, block a product or compel a lab to hand over a model. Enforcement still rests with existing regulators such as the Australian Competition and Consumer Commission, the Office of the Australian Information Commissioner and the eSafety Commissioner, under laws that were not written with AI in mind.

A A$7.5 million-a-year institute will not out-muscle the world's AI labs, and it is not meant to. Its value lies in giving government its own technical judgement, so policy is not written solely on the say-so of the companies being regulated. The testing partnerships with the United Kingdom and Canada matter more than the budget, because they let a small team borrow tools it could never build alone. The weakness is that advice without powers depends on someone acting on it. We expect the institute to become the technical engine behind whatever AI law Canberra eventually writes. Until then, it can raise the alarm but not stop anything.