Joint letter on automated decision making legislative framework

Dear Attorney-General,

Re: Framework for automated decision making in Government

We welcome the work you are leading to develop a framework for automated decision making in Government (ADM framework bill), which would fulfill the Albanese Government’s commitment to implement a key recommendation of the Robodebt Royal Commission.¹ We are writing to raise some crucial safeguards that we wish to see included in the forthcoming framework.

When properly implemented, ADM can increase the efficiency of decision making, particularly for high-volume, routine decisions – especially decisions that don’t involve discretion or evaluative judgment. However, the consequences of using ADM without sufficient safeguards can be catastrophic, and in some cases deadly, as illustrated by the experience of Robodebt. The risk is especially high for vulnerable and marginalised people, who are more likely to rely on government services to meet their basic needs.

By way of background, some of the signatories to this letter have expressed concern about the Government’s use of ADM in aged care and social security, particularly the Integrated Assessment Tool and the Targeted Compliance Framework.² There are also strong concerns about the inadequacy of safeguards for ADM in the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Act 2026.

The ADM framework bill is an important opportunity to set the rules to enable safe, effective ADM across government. We consider the following safeguards as vital for inclusion in the ADM
framework bill. Each is an essential protection against another Robodebt-style disaster.

1. Parliament should determine which types of decision may be automated

Where ADM is used to make decisions that affect people’s most basic human rights, the risk can be very high. Hence, Parliament should be responsible for determining which categories or types of government decision are suitable to be made via ADM. In exceptional circumstances,
where a Minister is responsible for determining which types of decision may be automated, this ministerial power should be subject to strict criteria to uphold the human rights of affected people.

2. Protections against automation of discretionary or evaluative decisions

Automation can be appropriate for routine decisions, especially where discretion does not need
to be exercised. However, discretionary and evaluative decisions should not be subject to automation as a general rule.. Discretionary decisions tend to be where government can consider a person’s individual circumstances. Discretion also enables the law to be applied fairly, and as intended, in complex or unique cases – including where decisions involve vulnerable and marginalised people.

This principle is well-understood: several Government authorities including the Commonwealth
Ombudsman and the Australian Human Rights Commission have cautioned against automating
discretionary decisions.³

3. People should be able to seek review of automated decisions
The law should enable a person who wishes to challenge an automated decision to do so using accessible and appropriate review processes.

This should include:

• internal review by a more senior official, which is the most common, cost-effective and accessible pathway to address inaccurate or unfair decisions⁴
• extend to external merits (eg, by a tribunal)
• judicial review.

The use of ADM should never have the effect of denying or reducing review rights for people affected by government decisions. If a decision would be reviewable if made by a human, it should also be reviewable if made through automation. The impact on a person’s rights is the same, regardless of how the decision is made.

4. Affected people should be given reasons for automated decisions

When a decision is made affecting a person’s legal or similar rights, they should be provided with a plain language explanation of the legal and factual basis for the decision. This is a key lesson from Robodebt, where Australians were told they had a debt, but weren’t given an explanation of the basis of the debt. They therefore had no practical ability to challenge inaccurate debts.

5. Proactive consideration of human rights, including impacts on vulnerable and marginalised communities, and the law

Before adopting ADM, government should be required to proactively assess the human rights of those affected, especially for people experiencing vulnerability or marginalisation. ADM systems that unjustifiably restrict or limit human rights should not be implemented.

Government should also be required to ensure that any ADM complies with the law so that there is no repeat of Robodebt or more recent examples like unlawful payment cancellations under the Targeted Compliance Framework. Government should test and be assured that the ADM is lawful before it goes live.

We urge you to ensure that the five principles above are reflected in the ADM framework bill. This is essential for enabling safe and effective ADM, and preserving trust in government itself.
We would be pleased to expand on these issues with you or your Department.

Yours sincerely,
Your Signature Here

1. Hon Michelle Rowland MP, ‘AI Consumer safety priorities’ (Media Release, 20 July 2026); Robodebt Royal Commission, Final Report (July 2023) Recommendation 17.1
   ↩
2. See e.g. Melissa Davey, ‘Labor’s controversial algorithm tool for aged care under investigation by Ombudsman’ The Guardian, 15
   April 2026; Joseph Brookes, ‘Automated welfare compliance saga hangs over budget’ Innovation Aus, 15 May 2026
   ↩
3. Commonwealth Ombudsman, Automated decision-making: Better practice guide (2019) 10 ; Australian Human Rights
   Commission, Human Rights and Technology (Final Report, March 2021) 56.↩
4. Rights to internal review are especially important since most people will not seek external review – e.g. only five out of every 10,000 Robodebts were reviewed by the AAT: ‘Unravelling Robodebt: Legal Failures, Impact on Vulnerable Communities and Future Reforms’ University of Sydney (Web Page) 13 December 2023 ↩