Senators call out government denial over IAT human oversight

Tanya Plibersek’s claim that the government would never let a computer make an abstract decision were called out in today’s Senate estimates.

---

Senators have questioned the federal government’s insistence that it would never let a computer algorithm make a decision without human oversight, pointing at the Support at Home algorithmic Integrated Assessment Tool as exhibit A.

The Community Affairs Legislation Committee in Canberra heard that minister for social services Tanya Plibersek made comments in Question Time last week that the government would not “just hand over decision making to an abstract computer program”.

Ms Plibersek’s comments referred to the Robodebt scheme, which pursued around 440,000 people for unverified Centrelink debts.

South Australian Liberal Senator Anne Ruston told the Budget Estimates inquiry this morning that Ms Plibersek had said:

“Those opposite came up with an automated decision-making program that was designed simply to raise revenue. It was used to attack the most vulnerable people in our community. There is certainly no way that our government would ever do such a thing.”

Senator Rustin saw the opportunity and took it.

“So, in light of that statement, can you explain how the government’s use of an Integrated Assessment Tool is consistent with what Minister Plibersek said?”

Queensland Labor Senator Nita Green, appearing on behalf of federal health minister Mark Butler and aged care minister Sam Rae, said trained assessors were involved “in every step of the process”.

But Senator Ruston said evidence had been provided “time and time and time again that there is no capacity for human override in the assessment outcome”.

“Which basically means the decision as to the level of [Support at Home] package you’re going to receive and the … prioritisation of when you receive that package has no human override. Do you dispute that?”

Senator Green said Ms Plibersek’s comments were not being asked in relation to the IAT.

“I think that it is a bit tricky to associate the comments she made as being directly relevant to the Integrated Assessment Tool,” she said.

“Each of the three stages [of SaH assessment] have expert human oversight involved, and if there’s more detailed information, then that data would be with the department, and they can answer that for you.”

Queensland Greens Senator Penny Allman-Payne saidhuman oversight was “different to override”.

“Certainly, what we heard at the inquiry into this, and the evidence from the department, is that a [assessment clinical] delegate cannot override the IAT classification, and there is no data kept on how often concerns are raised.

“So how can that stage be described as oversight if the human reviewer can’t actually change the outcome?”

Senator Green said the government had a system “to ensure that there’s efficient and accurate and fair distribution of these assessments, and the most important part is that a clinical assessment is made around the information that goes into the algorithm”.

But Senator Allman-Payne asked how that would explain why an older person with motor neurone disease – “which everybody knows is a disease that only goes in one direction” – was being assessed under the IAT as needing less care than they were originally assessed for.

In 2023, a Royal Commission ruled that the Robodebt scheme was unlawful, and a 2025 class action resulted in total compensation and refunded debts of more than $2.4 billion.

Ms Plibersek also said in parliament that whenever the government used AI to process information more quickly and efficiently, that only occurred “with strict guardrails and strict human oversight”.

“Issues around when and how the government uses automated decision-making are very important for us to grapple with as a parliament because AI is increasingly prevalent right across our society and right across our economy,” she said.

“Of course, we as a government have discussed in a great amount of detail the appropriate safeguards that we need when automated decision-making tools are used in any of our portfolios.

“Our approach is that in every case there needs to be human oversight of any use of AI. We don’t just hand over decision-making to abstract computer programs, as was the case under Robodebt, which caused such great suffering for 430,000 Australians. 

“Those opposite came up with an automated decision-making program that was designed simply to raise revenue. It was used to attack the most vulnerable people in our community.

“There is certainly no way that our government would ever do such a thing that we know resulted in the suicide of Australians who were impacted by this.”

DoHDA Secretary Blair Comley said inputs to the IAT algorithm were “determined by a human” which then determines an individual’s classification level.

“That is why we consistently are saying the human is in the loop, they are the ones entering clinical judgments, judges about circumstances that then get processed into a classification level.”

In response, Mr Pocock said “I think most Australians would probably see a key difference between human input and human oversight. Those are two very different, different things.”

Earlier, Senator Ruston told the inquiry that her office had been contacted by many older people who were told their assessment was not urgent.

“When you’re 93 and you’ve got incredibly high care needs, and then you’re told you’ve got to wait seven or eight months, quite frankly, people who are waiting seven or eight months may not be around to receive their package,” she said.

“And then when they do eventually receive their package, they receive a package at 60% of what they’ve been allocated … for potentially another three months, so quite reasonably they could be waiting nearly up to 12 months to actually receive the care that they need on a level seven or a level eight.

“So, I’d be very keen to understand how many and what is the rationale built into the algorithm that allows the algorithm to spit out a seven or an eight package with a basic prioritisation on it.

“What is the department’s response, or maybe what the minister’s response … when somebody is assessed as that, they go back through the process that takes 90 days, they get told actually there was a mistake, they should have been high priority.

“What is the department’s response to the incredible pain, suffering of that person for that period of time? And then they end up only getting 60% of their package.”

Senator Green said “the tool supports assessors, it doesn’t replace them, and the tool provides consistency to make sure that the right level of care is being given”.

The inquiry also heard questions around how the algorithm was assessed and whether any clinicians were involved during its development and finalisation.

Independent ACT Senator David Pocock asked the inquiry whether any aged care providers, advocacy groups or older people were consulted about the removal of the human override function of the IAT.

DoHDA first assistant secretary for access and home support, Greg Pugh said “in terms of if we ever consulted one individual provider or older person about the removal of the override, I believe the answer to that is no, but we would have to double check”.

Earlier, Senator Pocock told the inquiry he had been contacted by an experienced geriatrician who described the experience of an older patient with “significant visual impairment”.

“He could get to the toilet because he knew the layout of his house very well, but he couldn’t see anything. He couldn’t see if things were dirty, he couldn’t see what medications he was taking.

“The IAT and algorithm said that he was mobile and continent and only needed a level two, but any experienced clinician could see that the man needed a level four.

“So, his daughter has now moved in with him, and while he waits for reassessment, she’s using up her long service leave because he’s obviously now at significant risk of injury.

“So again, which group of clinicians has actually looked at the algorithm that’s currently in place and said this is fit for purpose?”

The inquiry heard that the current iteration of the IAT algorithm was not part of the trial in 2023.

“So what is the basis of your confidence in relation to the accuracy of the … algorithm component of the IAT?” Senator Rushton asked.

Mr Pugh said the outcomes of the IAT and the supporting algorithms were validated through 200,000 aged care assessments.

“Each time they were an output or an outcome of that trial was reviewed, those algorithms were then progressively improved to make sure that they were fit for purpose,” Mr Pugh said.

The inquiry also heard that as at 31 March, 1117 Support at Home assessments were up for review and 989 of those were reviews of IAT decisions.

Of those, 606 were finalised, 92 had been affirmed (with the original decision upheld), 215 had been withdrawn, and 102 had been found that the application for review had no legal standing.

A further 81 had been set aside and 51 had been varied, which were found to be DoHDA’s responsibility.

The average time to review completion was 72 days, with reviews of 24 cases taking longer than the maximum 90 days.