How Robodebt went wrong.

The short version.

From 2015 to 2020 the Australian Government ran an automated program that raised welfare debts by comparing a person's annual income on tax-office records with the fortnightly income they had reported to Centrelink. Where the person did not supply payslips, the system spread the annual figure evenly across the year and treated the difference as an overpayment. About 794,000 debts were raised against about 526,000 people.¹

The method had no basis in law, and government lawyers had said so before it started. The Royal Commission that examined it called it "a crude and cruel mechanism, neither fair nor legal".¹ It was budgeted to save about A$4.8 billion and ended as a net cost of about A$565 million, before two class action settlements. A$746 million was refunded and A$1.751 billion of debts was written off in total.¹

Robodebt did not use artificial intelligence. The Commission is explicit that it used automation.¹ Its lessons apply to AI because the failures were in how decisions about the system were made: the legal basis was never settled, the human check was designed out, the burden of proof moved to the public, and five years of warnings went unheeded.

How the scheme began.

The idea was to use data the government already held. The tax office receives each employer's annual pay figures. Centrelink receives the income a person reports every fortnight. Where the two did not match, there might be an overpayment. Comparing them was not new. What was new was the plan to let the annual figure stand in for the fortnightly ones.

In late 2014 the Department of Social Services' own lawyers advised that averaging "might not be consistent with the legislative framework", and policy advice warned that debts calculated this way might be wrong. In February 2015 the Social Services Minister, Scott Morrison, signed an executive minute that noted the department's view that legislative change would be needed. The New Policy Proposal that went to Cabinet in March 2015 said legislation was not required and did not mention averaging. The Commission called that statement false, and found that Mr Morrison allowed Cabinet to be misled because he did not make the obvious inquiry.¹

The measure was funded in the 2015–16 Budget under the heading of strengthening the integrity of welfare payments, and extended in later budget updates.¹ A 2015 pilot showed that only about 40 per cent of people contacted made contact back. The pilot was not properly evaluated, and the online screens were never tested with users.¹

How averaging produced debts.

Social security entitlement depends on what a person earned in each fortnight. Someone who works for part of the year and claims benefits for the rest can owe nothing, because they earned nothing while on benefits. Averaging divides the year's income by 26 and assumes the same amount was earned in every fortnight, including those spent on benefits.²⁷ For anyone with casual, seasonal or broken work, that produces a debt for income that was never earned.

The Commission identified five changes from the earlier, manual practice. Taken together they turned a lead for an officer to investigate into a debt notice generated by the system.¹

Process maps put before the Commission show how much was automated. A person who did not respond was treated as accepting the tax-office figure. Entitlement, the debt and a 10 per cent penalty were calculated automatically, and the debt notice was generated automatically.¹ The letters left out a compliance phone number by design, to push people online. The Commission described the penalty as a "euphemistically called" recovery fee and found no evidence that anyone had considered its legal basis.¹

The scheme at full scale.

The Online Compliance Intervention began in July 2016 and was fully rolled out in September, at 20,000 notices a week. Under the manual process there had been about 20,000 interventions a year. In November 2016 the Minister for Human Services announced that new technology was helping raise $4.5 million in welfare debts a day.¹

By December the news was full of people who had received debts they did not owe, and the Australian Council of Social Service wrote to the Minister. By 27 January 2017, 76 per cent of people in the online phase, 99,404 people, had debts raised from averaged earnings.¹ People were asked for payslips going back years, while the department's own website told them to keep payslips for six months.¹ Debts were passed to private collectors without checking that the person had received the notice. The collectors were paid about A$11.6 million between 2015 and 2021.¹

Seven warnings.

The scheme ran for three more years after the first public alarm. Each warning had an answer that let it continue.

Two of these matter most for anyone governing an automated system. The first is the Ombudsman. When it investigated in 2017, the departments gave it a 2017 internal view that averaging could be used "as a last resort", which cited no law, rather than the 2014 advice. The Commission found that officers' explanation to the Ombudsman was untrue. The Ombudsman did not examine legality, and ministers then cited its report as evidence the scheme was lawful.¹ The oversight body was used as cover.

The second is the tribunal. By the end of May 2017 about 29 first-tier tribunal decisions had rejected averaging.² The department fixed each case that was appealed and otherwise carried on.¹ An external error report that arrives one case at a time can be absorbed indefinitely unless someone is required to read the pattern.

The legal collapse.

The legal position, known inside government since 2014, came out in court. The Australian Government Solicitor advised in March 2019 that averaging could not establish a debt. In September 2019 the Solicitor-General advised that averaged tax-office data could not support a conclusion about whether a debt existed or how much it was without further evidence.¹ On 19 November 2019 the government stopped raising debts on averaged data alone, and a class action was filed the same day. Eight days later, in the test case Amato v Commonwealth, the Commonwealth consented to a declaration that the averaged debt had not been validly made.²

The scheme closed on 30 June 2020 and refunds began that July.¹ The class action settled on the day set for trial, in November 2020. Approving the settlement in June 2021, Justice Murphy recorded the Commonwealth's admission that it "did not have a proper legal basis to raise, demand or recover" debts based on averaging.²

The human cost.

This section records deaths by suicide. If it raises something for you or someone you know, in Australia call Lifeline on 13 11 14, in Singapore call the Samaritans of Singapore on 1767, or find a local helpline at findahelpline.com.

The Commission heard from people who had received debts they did not owe and could not disprove, and from families of people who had died. It records that Rhys Cauzzo, aged 28, died in January 2017 after debts raised by averaging in the manual phase that preceded the online scheme had been referred to a debt collector. It records that Jarrad Madgwick, aged 22, died in May 2019, and that his mother identified a debt notice as a precipitating factor; the Commission notes that his debt does not appear to have come from averaging.¹ The Commission found that by July 2017 the Minister for Human Services, Alan Tudge, knew that at least two people had died by suicide, and that no comprehensive review followed.¹

Its summary of what the design left out is one line: "Empathy could not be programmed into the Scheme."¹

What it cost.

The scheme was budgeted to save A$4.772 billion. It achieved about A$406 million in savings at a total cost of A$971 million, a net cost that the Commonwealth accepts of about A$565 million.¹ A$746 million was refunded to some 381,000 people, and A$1.751 billion of debts was written off in total, a figure that includes the refunds.¹ The 2021 class action settlement added A$112 million.² A second action, appealing that settlement on the ground of misfeasance, settled in September 2025 for A$475 million in compensation, about A$548.5 million with administration and legal costs. Justice Beach of the Federal Court approved it on 23 June 2026, with about 125,000 people registered to share in it.²⁴²⁶²⁹ The Scheme Administrator expects to make fixed payments from October 2026 to February 2027 and to pay individually assessed claims in mid to late 2028, about eight years after the scheme closed.²⁴

The reckoning.

A Royal Commission, led by Catherine Holmes AC SC, was established in August 2022 and reported on 7 July 2023. It found the scheme "a costly failure of public administration, in both human and economic terms", and made adverse findings about ministers and senior officials. A sealed chapter referred individuals to the Australian Public Service Commission, the National Anti-Corruption Commission, the Australian Federal Police and a legal regulator.¹ The report closes on "a myriad of ways that things can go wrong through venality, incompetence and cowardice".¹

The accountability process took five more years and did not run smoothly. The Anti-Corruption Commission first declined to investigate. Its own Inspector then found officer misconduct in that decision, and an independent delegate reopened it. In March 2026 the Commission found that two former senior officials had engaged in serious corrupt conduct, one by misleading officers preparing the 2015 Cabinet submission and one by misleading the Ombudsman in 2017. It made no referral for prosecution, and found that four others, including Mr Morrison, had not acted corruptly.⁵⁶⁹¹⁰ The reason for no referral was evidential. The admissions that could prove an offence had been given under compulsion, and compelled evidence cannot be used in a criminal trial.³⁰ The Royal Commission's finding about Mr Morrison and the anti-corruption finding apply different legal tests, and both stand.

What the Royal Commission recommended.

The report's text refers to 57 recommendations; its numbered list and the government's response contain 56. In November 2023 the government accepted all of them, in full or in principle.¹³ Two deal directly with automation. The Commission's finding behind them is short: "The automation used in the Scheme at its outset, removing the human element, was a key factor in the harm it did."¹

Recommendation 17.1 asks for a consistent legal framework for automation in government services; 17.2 for a body to monitor and audit it. The right to an explanation comes from the chapter's discussion behind 17.1. Royal Commission report, pp. 486 to 488.

Recommendation 17.1 asks government to consider legislating a consistent legal framework for automation, with a clear path to review, plain-language notice on agency websites that automated decision-making is used, and business rules and algorithms "made available, to enable independent expert scrutiny". Recommendation 17.2 asks for a body to monitor and audit automated decision-making for technical soundness, fairness, bias and usability.¹ The government accepted both.³

Progress since has been slow. The Attorney-General's Department consulted on a framework from November 2024 to January 2025 and published a summary in August 2025.²⁰ The Ombudsman updated its better practice guide in March 2025.¹⁶ Amendments to the Privacy Act require organisations to disclose certain automated decisions from 10 December 2026.²² In July 2026 the government listed a framework for automated decision-making in federal agencies among five AI safety priorities and asked the Attorney-General to lead it.³¹ As at 12 October 2026 no government bill has been introduced. A private member's bill to implement recommendations 17.1 and 17.2 was introduced on 7 September 2026 and is still before the House of Representatives.²¹³²

Why it matters for AI.

Every failure in Robodebt has a direct counterpart in AI deployment. A rules engine that averages income and a model that predicts it both produce an estimate that someone has to decide whether to act on. In Robodebt the estimate went straight to a debt notice. The same questions apply to any AI system that decides or recommends something about a person: is there a legal basis for acting on this output, who checks it before it reaches the person, who carries the burden of correcting it, and who is reading the pattern of complaints and appeals?

Two features make AI riskier on these points. A probabilistic model is harder to explain than a formula, so the "business rules made available" that the Commission asked for are harder to produce. And the speed and scale that made the online scheme attractive are greater with AI. The Commission's recommendations were written for automation. They read as a minimum for AI.

The same pattern is appearing again in Australian services. In 2026 the Commonwealth Ombudsman confirmed it was investigating complaints about an aged-care assessment tool that classified people's funding levels and that assessors could not override, and by July the Senate had passed a private bill to require that such tools support professional judgement rather than replace it.³¹

What boards and agencies can take from it.

1. Settle the legal basis before you automate. The lawyers' caveat was in the 2015 minute and missing from the Cabinet proposal. Any automated or AI decision about a person needs written confirmation that the method is lawful, and that confirmation should travel with the funding decision.
2. Keep a human check where an error reaches a person. The Commission found that removing the human element was a key factor in the harm. Decide which outputs a person must review before they take effect, and fund the reviewers when the volume rises. The oversight ceiling calculator estimates how many you need.
3. Do not move the burden of proof to the public. Asking people to disprove a machine's figure with records they were told they could discard turns an estimate into a verdict.
4. Read appeals and complaints as a pattern. Twenty-nine tribunal decisions were handled one at a time. Someone senior should be required to report adverse external decisions in aggregate, with the error rate they imply.
5. Protect the independence of legal advice. Advice that averaging was unlawful existed in 2014, 2018 and 2019. Legal advice on an automated program should go to the board or minister directly, not through the team that owns the program.
6. Measure harm as well as savings. The scheme was judged on dollars recovered while the deaths, complaints and tribunal losses went uncounted. Put harm indicators beside the financial ones from the first month.
7. Publish the rules and give reasons. The notices gave no reasons. The Commission asks for business rules and algorithms to be open to expert scrutiny. For AI, that means documenting what the system does and does not consider, and explaining each decision to the person affected.
8. Do not lean on an oversight body that was given a partial picture. The Ombudsman's report was used as proof of legality on a question it had not examined. A clean review covers only what the reviewer was asked to look at.

To check a proposed automated or AI decision against these points, place it on the agent risk assessment matrix, which sets the controls by autonomy and consequence.

Full timeline.

Common questions about Robodebt.