Two things were reported on 20 July. The federal government is developing new rules to make AI-driven automated decision-making in public services fairer and more transparent (The Mandarin, 2026a). On the same day, a former Services Australia employee, Jeannie-Marie Blake, began proceedings against the Commonwealth alleging mistreatment over her role as a Robodebt whistleblower (The Mandarin, 2026b).

The pairing is worth sitting with. One item is the institutional lesson on its way into law. The other is the institutional behaviour, three years after a royal commission examined that behaviour in forensic detail and reported on it. Both are true at once, and the gap between them is the subject of this piece.

Start with what is actually coming, because the timetable is firmer than most commentary suggests. The Privacy and Other Legislation Amendment Act 2024 introduced an automated decision-making obligation. From 10 December 2026, entities covered by the Australian Privacy Principles that use personal information in automated decisions capable of affecting rights or interests must set out in their privacy policies what kinds of personal information are used and what kinds of decisions are made. The Office of the Australian Information Commissioner ran a consultation on guidance for that obligation which closed on 15 June, and the Attorney-General’s Department is leading a broader whole-of-government transparency framework, a commitment made in the government response to the Robodebt Royal Commission (OAIC, 2026; Attorney-General’s Department, 2024).

That is genuine progress and deserves saying so. A statutory obligation with a commencement date beats a voluntary principle by a wide margin, and Australia has spent enough of the last three years discovering what voluntary produces. I traced that pattern through the AI transparency register in The Transparency Illusion, where compliance rose to ninety-four out of ninety-four agencies while the substance behind the statements stayed thin.

Here is the difficulty. What arrives in December is a disclosure obligation, and Robodebt was never a disclosure failure. The scheme was announced, and its method was described in budget papers and defended at estimates. Journalists wrote about the scheme, affected people wrote about the scheme, and the Commonwealth Ombudsman examined it. Anyone who wanted to see how the scheme worked could see it. What failed around the scheme was the legal advice and the absence of anyone holding both the authority and the incentive to stop it while it ran. I set that failure against the UK’s HMRC experience in The Same Failure, in Sequence, and in both cases accountability only activated once the damage had become public.

Research published on 20 July puts an academic frame around that distinction. A paper drawing lessons from major industrial disasters argues that AI risk cannot be understood by evaluating technical components on their own, because organisational and political factors are what convert a known hazard into a catastrophic failure (Unsafe at any AUC, 2026). Its recommendation is mandatory sociotechnical risk assessment covering the traceability of responsibilities and the organisational dynamics around a system. A technical audit of the Robodebt algorithm would have passed. Income averaging is arithmetically simple and it did what it was designed to do. The catastrophe was in the organisation wrapped around it.

The clearest available test of whether a safeguard is real comes from European case law. Following the Court of Justice’s SCHUFA ruling, Article 22 of the GDPR now operates as a default prohibition on solely automated decisions with significant effects, and nominal human oversight without genuine authority to deviate does not satisfy its safeguards. Public bodies using algorithmic tools for eligibility or resource allocation must ensure that reviewers hold real interpretive discretion (Another View on Article 22 GDPR, 2026). Set that against Australian practice and the implication is uncomfortable. Robodebt had humans in the loop at enormous scale. Compliance officers reviewed cases. Call centre staff spoke to affected people daily. What none of them had was documented authority to set aside the output, and a review function that cannot deviate is a processing step wearing the costume of a safeguard.

That distinction should drive how the Australian framework is drafted, and authority is not hard to write down. Name the officer who holds authority to suspend an automated decision process, and publish the threshold at which that authority is used. Monitor how often reviewers actually depart from the automated output, and treat a departure rate of zero as a finding requiring explanation, because a review function that never disagrees with the machine is exercising no authority at all. Log the legal advice obtained before deployment, with the date, so a later inquiry does not have to reconstruct who knew what and when. Then test the escalation path against a real case before go-live, the way an agency tests a business continuity plan, because authority nobody has ever used is authority nobody is sure they hold.

The Blake proceedings matter here for a reason beyond their own merits. Whistleblowing is the mechanism of last resort in every accountability system, the one that operates when the internal channels have already failed. Three years after the royal commission, a person who used that mechanism is in court over how she was treated for using it. Whatever those proceedings decide, every public servant weighing up the same mechanism is watching them, and what they learn will sit underneath the new automated decision-making framework whether the framework acknowledges it or not. A guardrail only works when somebody is willing to reach for it.

There is one more warning worth carrying into the drafting. Work published in mid-July on Norway’s pandemic response describes how emergency conditions produced a mode of governance in which rights and regulations were suspended so that siloed data could be integrated across health, education and defence. Infrastructure built under those conditions, the authors warn, can permanently reshape data rights afterwards (Digital Preparedness, 2026). The next Australian automated decision failure will most likely be built quickly, during something that feels like an emergency, by people with entirely defensible motives. That is the case the framework has to survive, and a framework tested only against calm conditions will not.

Australia sits in a better position than it did in 2016. The safeguards now have a commencement date and a department coordinating them, and they rest on findings nobody can pretend not to have read. What the safeguards have yet to prove is that they hand somebody the authority to halt a system that is working precisely as designed and hurting people anyway. Authority of that kind is what makes a safeguard a safeguard, and it is the easiest thing to leave out of a draft. December is four months away.

References