The Robodebt Kill Chain: 12 Letters, 13 Phone Calls, and a Dead Man

Between 2015 and 2019 the Australian government’s Online Compliance Intervention (“Robodebt,” once the press got hold of it) issued roughly 433,000 unlawful debts worth about AUD$1.76 billion. A Royal Commission under former High Court Justice Catherine Holmes later found it had been illegal from the day it launched. The ministers who ran it were warned of that, in writing, while they ran it, and they kept going.

What the algorithm produced was a debt notice. What stood on the other end of the notice was a citizen now required to prove a negative. In writing, against a number generated by software the government refused to explain.

Six hundred and sixty-three people who were issued a debt under the scheme died before their case was closed. None of those deaths was the algorithm’s doing in any way a coroner could write down. That gap, between what the machine did and what happened to people, is the whole subject of this piece.

The System

Robodebt set the annual income a person reported to the Tax Office against the fortnightly income they reported to Centrelink. It took the yearly ATO figure, smeared it evenly across 26 fortnights, and called that average your real fortnightly pay. Any gap between that invented number and the Centrelink record became a debt.

The arithmetic could not have done anything else. Anyone whose income arrived unevenly (a seasonal worker, a contractor, a student picking up casual shifts between exams) would show “discrepancies” that were nothing of the kind. Averaging was the flaw at the heart of it, and the scheme ran on that flaw as though it were the design.

The old practice had a human in it: an officer who reconciled the claim against the actual fortnightly records before anyone was billed. Robodebt deleted that officer and flipped the burden of proof onto the citizen. You got the notice, and then you had to dig up years of documentation to disprove a finding a machine had already made. When the documentation was gone (payroll records shredded under the standard seven-year retention rule, bank statements no longer reachable, the casual employer long since folded) the debt simply stood.

The Scale

433,000 unlawful debts. AUD$1.76 billion claimed. It ran for four years before the Federal Court struck down the averaging method in the Amato case in November 2019, and even then the government’s first move was to settle that one case and keep the program alive. The full stop did not come until 2020, and the repayments dragged into 2022.

Rhys Cauzzo

Rhys Cauzzo was twenty-eight. His Robodebt notice was for about AUD$2,000. He wrote twelve letters and made thirteen phone calls trying to sort it out, and in January 2017 he died.

His mother, Jennifer Miller, became one of the most visible advocates for Robodebt’s victims, through the Royal Commission and the long years before it. Her account was that her son had tried, again and again, to engage the system, and that every call produced a different answer from a different officer, none of whom would concede the one thing that was true: that the debt notice itself was the mistake.

The other deaths trace the same outline. A notice arrives. The appeals process turns out to be unnavigable. The documentation the government will accept does not exist, and there is no money to hire anyone who could find a way through. What the survivors describe is the stress and the shame of it, and the particular cruelty of being ordered to produce paperwork the algorithm never asked for and the citizen had no reason to keep. SBS Australia documented 663 deaths among people who had been issued notices. Whether the scheme caused any single one of those deaths is the kind of finding that would need evidence no government and no investigator has assembled person by person; the 663 is the count of people with a debt on file who died before it was resolved. The Royal Commission heard the families.

The Royal Commission

Catherine Holmes, former Chief Justice of the Supreme Court of Queensland, was appointed in 2022 to lead the Royal Commission into the Robodebt Scheme. Her report, handed down in July 2023, was unsparing.

Senior public servants and ministers, it found, had been told the income-averaging method was unlawful and had pressed on regardless; internal legal advice flagged the problem at multiple points during design and operation, and was simply not acted upon. The Commission referred a handful of individuals to the National Anti-Corruption Commission and to professional disciplinary bodies, and sealed a chapter naming people for further consideration. No criminal charges followed.

The Commonwealth had already agreed to a class-action settlement of roughly AUD$1.8 billion (debt waivers, repayment to those who had paid the unlawful debts, and damages), approved by the Federal Court in 2021 and processed over the years after. The legal record on the program’s unlawfulness is now closed. The mechanism that produced it (automated decision, burden of proof on the citizen, sold to the public as efficiency, deployed over legal advice nobody followed) is still the template for the government services next door to it.

The Pattern

Robodebt is one instance of a class: automated decision-making against individuals, where a false positive costs the individual everything and correcting it costs the individual everything again.

COMPAS, the Correctional Offender Management Profiling for Alternative Sanctions tool, scores defendants in American courtrooms on the odds they will reoffend. ProPublica’s 2016 investigation in Broward County, Florida found it flagged Black defendants as higher-risk-who-didn’t-reoffend at nearly twice the rate of white defendants, while white defendants more often got the low-risk score and went on to reoffend anyway. It is still in use.

The Allegheny Family Screening Tool scores child-welfare referrals in Pittsburgh. It has been audited, litigated, and adopted in something like twenty-six states by mid-2026, per ACLU tracking. The DOJ Civil Rights Division opened a Title II ADA probe in November 2022 that remains open. When their own tools failed audit, the London Borough of Hackney and the Oregon Department of Human Services dropped them; Allegheny kept its.

Michigan’s MiDAS, an automated unemployment-fraud detector, accused tens of thousands of people of fraud between 2013 and 2015 at an error rate later put above 90 percent, and the state was eventually ordered to refund around $20 million. The architecture was Robodebt’s exactly: the machine decides, the citizen disproves, and the citizen has no leverage against a reversed burden of proof.

The furthest end of the same architecture is the Lavender system the Israel Defense Forces used during the Gaza campaign. As documented by +972 Magazine and Local Call, it generated kill-list targets with reported human-review windows of about twenty seconds each. Robodebt is the welfare-state edition of that choice: the algorithm generates the output, the human signs off in a fraction of the time real verification would take, and the system keeps moving. The speed is the institution’s; the cost of the speed is shipped downstream to whoever the output lands on.

The Twenty-Second Rubber Stamp

“Twenty-second rubber stamp” is the operational signature of the whole class, and it needs no malice to work. The operator does not have to believe the algorithm is wrong. He only has to be handed more cases than verification allows, with the verification step being precisely the bottleneck the system was bought to remove.

Centrelink’s staff were not the villains here; they were working to throughput targets that made real review impossible, and the Australian Public Service had its own mid-level officers who flagged the legal problem and were overruled. Some were named in the Commission’s sealed chapter. The villains, where there are any, sit higher up. The referrals went to ministers and senior public servants, the criminal question to the National Anti-Corruption Commission and the prosecutors, the disciplinary question to the Public Service Commission. The one question that went nowhere is the one that matters most: how do you design automated decision-making so the cost of its errors does not fall entirely on the person least able to carry them.

What the Settlement Was

The AUD$1.8 billion bought back debt waivers, returned wrongly paid money, and added damages. What it could not buy back was the interval. The years between the notice and the resolution, during which the debt was real to the person carrying it whether or not it was real in law. The money returns the dollars. It does not return the years.

The Commission’s report and its sealed chapter are the legal record now. The civil settlement is processed, the recommendations are on the shelf, and the architecture itself (automated decision, reversed burden of proof, throughput-optimised review, no recourse the citizen can actually afford) has been retired nowhere, in Australia or anywhere else. They retired the word “Robodebt.” The mechanism came back under new acronyms.

433,000 unlawful debts. 663 dead. AUD$1.8 billion in settlement, one Royal Commission, one sealed chapter, no criminal charges.

The mother who lost her son fought the system for eight years before a Royal Commission finally confirmed what she had been saying since 2017. Issuing the debt took milliseconds; establishing that the debt had been unlawful took eight years and a Royal Commission. That asymmetry (instant to accuse, near-impossible to undo) is not a flaw in the system. It is the system.

The algorithm did not kill anyone. The system did.

The receipts (free, on this site): the digital-ID stack

This research appears in The Secret Life of Evil Robots, Chapter 11.