Robodebt built debts from averaged income, then made recipients disprove them

A debt the recipient had to disprove

In July 2016 the Australian government's Department of Human Services began sending letters asserting a welfare debt and asking the recipient to log in and explain it. No one had checked the figure. It came from a computer that matched two records: what a person had told Centrelink they earned, fortnight by fortnight, against what their employers reported to the tax office for the whole year. Where the two disagreed, the system raised a debt and moved to recover it, garnishing tax refunds and handing accounts to debt collectors.¹

Reviewing overpayments was ordinary work. Welfare depends on income, and someone who earned more than they declared can owe the difference back. What was new in Robodebt was the arrangement: the government produced a number, called it a debt, and made the recipient prove they did not owe it.

How dividing a year made a fortnight

Australian welfare entitlement is assessed fortnightly: what you are owed in any two-week period depends on what you actually earned then. The tax office reports income the other way, as a single aggregate figure for the whole year, which cannot say what a person earned in any particular fortnight.²

Robodebt bridged that gap by division. It took the annual employer-reported total and, in the Royal Commission's words, “divided up evenly, and allocated into the number of fortnights included in the period being reviewed,” then treated that flat, invented figure as the person's income in every fortnight.¹

The arithmetic is on the record in one recipient's case. For the 2011–12 year, Deanna Amato's employer reported $24,811 in income to the tax office. Robodebt divided that by the 366 days in the period, reaching $67.79 a day, then multiplied by 14 for an “apportioned fortnightly income” of $949.05. That flat figure replaced what she had actually reported earning each fortnight. The sum assumed she had earned the year's income in equal fortnightly amounts, and her own reported earnings showed she had not. It produced a debt of $2,924.28, and the tax office was ordered to hand over her refund to cover it. A federal court later declared the demand “was not validly made”: there was “no probative material” for the assumption, and her own figures contradicted it.³

The department's own guidance had predicted this: averaging over 12 months “will not result in a correct result if the customer should have received a full rate at other times of the year.”² Averaging did not always invent the whole debt. The Commonwealth later accepted that Amato owed something, though “very substantially smaller” than the sum demanded. What averaging invented was the amount.³

The check that used to catch it

Income averaging did not arrive with the computer. The department had used it for years, but only as a last resort and often with the recipient's agreement, after asking for their actual records.² What Robodebt changed was who did the checking, and whether anyone did.

Before, a flagged discrepancy went to a compliance officer who examined the file, contacted the recipient, and could compel the employer to hand over the actual pay records for the period. About 20 000 files a year were checked this way.¹ That officer caught a wrong average before it became a debt. Robodebt removed the officer. Under the fully automated system, if the recipient did not go online and supply the evidence, the department would “no longer use its information gathering powers to obtain it directly from employers.”² The averaged figure stood, and disproving it fell on the person least able to make an old employer dig out years-old records.

Even staff who wanted to intervene often could not. A welfare advocate told a Senate committee that when a client's debt recovery was supposed to be paused, the department explained: “The computer is initiating these calls and they can't be overridden. The best we can do is reduce the amount that she has to repay.”⁴

The error rate per file may have been no worse than before. The scale was. In the manual program that ran just before full automation, reviews rose roughly fivefold in a single year, and about nine in ten of the debts it raised used averaging.¹ Removing the human check did not invent the error. It mass-produced it.

What it cost and how it ended

In the class action Prygodicz v Commonwealth, the Federal Court found the Commonwealth “unlawfully asserted such debts, totalling at least $1.763 billion against approximately 433 000 Australians.” It had already recovered about $751 million from roughly 381 000 of them.⁵ The Royal Commission put the refund program at $746 million reimbursed to some 381 000 people, with debts totalling $1.751 billion written off.¹ The two accounts measure different things at different times, not a disagreement.

The program had been sold as a saving. Budget papers projected about $2 billion over its first four years. Services Australia spent $606 million building it, and the savings never came.⁴

The human cost is on the record. The court recorded what group members reported: financial hardship, anxiety and distress, including suicidal ideation and in some cases suicide.⁵ The Royal Commission was aware that people with debts raised under the scheme had died by suicide, assigned no toll, and was explicit that causation is specific to each case.¹ Even the death most associated with Robodebt, that of Rhys Cauzzo, came from a debt raised through the earlier manual process, not the online system. Both used averaging.¹

Averaging is not unlawful in itself; using it as the sole basis to prove a debt is.¹ Amato was one case. In Prygodicz the court reached the same conclusion for the whole class: a debt decision resting solely on averaged tax-office data, with no other evidence of actual fortnightly income, “was not validly made.”⁵ The two bodies differ on what officials knew. Ruling in 2021 on the record before him, Murphy J declined to find that ministers and senior public servants actually knew the scheme was unlawful, calling it closer to “a stuff up” than “a conspiracy.”⁵ The Royal Commission, working from a wider base in 2023, found the department had legal advice as early as 2014 that averaging this way was unlawful, and that senior people knew of it.¹

Part of why the scheme survived was that an early review did not stop it. When the Commonwealth Ombudsman examined the system in April 2017, it did not reach lawfulness. It found the debts “accurate, based on the information which is available to DHS at the time the decision is made,” and said the fairness of the department's discretion “can only be answered by the courts.”² The department then relied on that report as evidence that Robodebt was legal, and later worked to keep the legality issue out of the Ombudsman's follow-up.¹

The retreat came in stages. In November 2019 the government announced it would stop raising debts on averaging alone,⁶ and days later the Federal Court's Amato consent orders held that averaged tax-office data could not by itself prove her debt.³ In May 2020 it announced it would refund debts raised wholly or partly on averaging, by then some 470 000 of them.⁶ The 2021 class action added a settlement of $112 million, inclusive of legal costs, on top of the refund program.⁵ A Senate committee majority, over a dissent from Government senators, recommended a royal commission. The government of the day did not adopt it, and the commission was established only after the 2022 change of government.⁴ When it released the Commission's report in July 2023, the government said it would consider the recommendations and respond in full.⁷ In September 2025 it agreed to a further $475 million to settle an appeal, Knox, still subject to Federal Court approval.⁸

Where the same bargain still runs

Robodebt stopped, but the system beneath it did not. Automated income data-matching between Centrelink and the tax office continued. The Ombudsman's 2021 report records that the online compliance system was replaced by later programs, one of which, Check and Update Past Income, “is still in operation.”⁶ What ended was the single move the courts struck down: using an average as the sole proof of a debt.

The weakness that carries beyond Robodebt is the bargain, not the software that ran it. A system takes a cheap signal, a data-match, a model's score, an average, treats it as proof of a fact about one person, then makes that person disprove it. The Dutch tax office made the same bargain with a fraud-risk model and drove tens of thousands of families into a dragnet they could not see or appeal. Only the signal changes. What stays is the shape: proof asserted cheaply, and the doubt made the citizen's to clear.