Australia's AI standards: report rogue incidents, bring your own power
Australia published the draft of its national AI rules this week, and the ask runs in both directions — labs that train models onshore would have to report when those models misbehave, and the data centres running them would have to supply their own energy and water.
Under a consultation paper released on September 18, frontier developers authorised to run large-scale AI training in Australia would face minimum security and safety requirements, including "disclosing defined reportable AI incidents to relevant Australian authorities" — a rule Australia does not currently have. The paper declines to define what counts as reportable and asks instead whether developers should disclose proactively, on request, or publicly. The same document sets capacity tiers for data centres: facilities below 30 megawatts are excluded, 30–100MW gets baseline requirements, and anything above 100MW — or a 150MW cumulative portfolio — faces the heaviest obligations, including published compliance plans and proof that building and grid-connection approvals are already in hand. Assistant science minister Andrew Charlton framed the trade plainly: "We welcome investment, but on Australia's terms."
Two design choices in the paper are the ones labs will fight over. The first is retroactivity: the government is considering applying the standards to data centres that are not yet operational but are already moving through approvals, which would catch a pipeline of projects financed under the old rules. The second is that training here comes with conditions beyond safety — developers would be expected to reserve computing capacity for Australian researchers or pay directly for their work, and the Attorney-General is running a parallel consultation on copyright that the paper ties to "incentivising AI training while providing strong copyright protections" for Australian creators.
The infrastructure half is a resource argument wearing a regulatory costume. Data centres covered by the standards would have to retire renewable electricity certificates equal to their annual consumption from the day they start operating, with a ramp toward all-new renewable generation, and would have to prioritise recycled or non-drinking water for cooling. That second clause has teeth in one specific market: Sydney Water estimates data centres could consume up to 20 percent of Sydney's drinking water by 2035.
The timing is not accidental. Prime Minister Anthony Albanese said in Washington this week that Australia will introduce new AI standards by the end of 2026, telling the ABC the risk is "that AI develops in a way in which humans are no longer in control of what AI is producing." He made the remarks after meeting Apple's Tim Cook at the company's California headquarters, where the two discussed the under-16 social media ban Australia has run for ten months — a policy the government says closed as many as 5 million accounts — and the digital duty of care and algorithm opt-out rules Labor plans to legislate before Christmas. If you want the background on how far these instruments reach compared with the EU's and China's, our explainer on what AI regulation actually covers is the place to start.
What to watch: whether "reportable incident" gets a real definition before the legislation is drafted — OpenAI disclosed six incidents of unexpected behaviour in six months without being required to, and the consultation closes October 9.
Should a country be able to condition AI training on reserved compute for its own researchers, or does that just push the training somewhere else? Tell us in the comments.
