Australia now makes AI firms prove their safety — what should going-global companies keep on record?

Australia has shifted the burden of proof on AI safety onto companies themselves. That single change matters more to firms doing business overseas than most of the day's headlines about chips and computing power. The direction is consistent across three developments in the same week: compliance is moving from "don't touch the red line" to "show us your records."

What actually changed is direction, not severity

On 8 October 2026, Australia's Assistant Minister for Science, Technology and the Digital Economy, Andrew Charlton, announced at a trust and safety summit in Sydney that the federal government would introduce systemic regulation of frontier AI. Companies would be required to build their own risk-management processes and to demonstrate that their safety systems are genuinely effective. A national standards programme is due to be completed by the end of 2026, with supporting legislation to be introduced to parliament in 2027.

That Australia intends to regulate AI is not news. What matters is how: the burden of proof sits explicitly with companies, rather than regulators listing prohibited behaviours one by one. Based on available public information, few national-level frameworks take this approach.

The announcement drew little attention on the Chinese internet. We think it deserves a read from any business owner going global — more than any financial headline that day about computing power or chips.

The reason is simple. With a list of prohibitions, you just avoid the listed items. With a self-certification regime, you first need something to certify.

Three events in one week, one message

On 9 October, the Central Committee of the Communist Party of China and the State Council issued a document on developing new quality productive forces. It calls for the full implementation of the "AI Plus" initiative, faster adoption of next-generation intelligent terminals such as AI phones and computers, and the construction of systems for technology monitoring, risk warning and emergency response to ensure AI is safe, reliable and controllable.

Platforms are tightening too. Anthropic updated its usage policy on 8 October, banning users from persistently abusing its AI models without cause from 12 November. Violations may lead to rate limits or account suspension.

States are writing rules; platforms are writing rules. All three point to the same move: the centre of gravity is shifting from "what you may not do" to "what you did, and how you prove it."

"This is aimed at large-model companies, not me"

This is the most common objection, and the easiest way to get caught out.

The burden of proof follows the business flow, not company size. Suppose you run a cross-border SaaS business and your customer-service bot handles conversation data from Australian clients. Or you run a factory using AI for quality inspection and production scheduling, with data stored on local servers. When someone asks how you ensure it is safe, the party answering is you — not the vendor supplying your API.

There is a subtler failure mode. For years, corporate compliance meant "don't touch the red line, don't leave a handle for anyone to grab." As a result, many firms kept almost no proper internal records. The new rules want precisely those records: risk-management processes, assessment conclusions, and the actions taken when something goes wrong. Without records, there is no compliance in a regulator's eyes.

There is a concrete way in. One technology publication distilled the enterprise-level requirements of "AI Plus" into three questions: can it run, which data may leave, and can you keep it under control after use. We suggest export-focused teams answer these three questions against their own operations first. Where you cannot answer, that is where the risk is.

How to implement: make traceability an internal standard

Rather than studying statutes country by country, make one thing a habit first: wherever AI is used, leave a checkable trace. Who, at what time, fed in which data, reached what conclusion, and on what basis. It sounds plodding, but it works equally for Australian-style self-certification and for what China calls controllability.

These requirements are not abstract, and products already address them. SHEYU ZHISHEN (舍予AI智审) offers a useful reference in how it handles applications for science and technology programmes: it reads tens to hundreds of pages in seconds, every extracted value carries a page number and the original text, dual engines cross-check each other, and no result is passed by default. Data does not leave the internal network.

It solves a specific problem in document review: on what basis do you claim a conclusion is correct. When a regulator asks a company how it proves its AI is safe, it is asking the same question.

Export-focused firms need not wait for legislation to land in 2027. Start with three small things: log AI-related operations, map your data flows clearly, and retain the review trail for anything you output. The cost is low. Whether you can stay at the table may depend on it.

Half a step further

What this round of AI regulation changes is not the technology threshold but the evidence threshold.

Customers will grade you on how well you use technology. Regulators will grade you on whether you can answer, substantively, how you ensure it is safe. For companies that can answer, the rules are guardrails. For those that cannot, the rules are a wall.