Services Australia is trialling machine learning to detect potential instances of identity theft affecting Centrelink customers, with the goal of stopping payments from being rerouted.
The agency was forced to defend its trials of machine learning technology late on Thursday night, after offering only a vague explanation in response to a news report by Information Age.
The report identified use cases in “debt prioritisation” and “fraud detection”.
Appearing before senate estimates late on Thursday night, officials sought to provide a more open view of the technology trials, while also saying they were “a long, long way from being able to deploy” anything to production.
The agency offered a twofold explanation for how machine learning is being applied in fraud-related use cases.
General manager of the fraud control and investigations Peter Timson said a key fraud-related use case relates to identity theft, and specifically to Services Australia wanting to alert customers when suspicious behavior associated with their account is detected.
“It’s being used in the investigative space where we can actually identify – without going into too much detail because scammers would be watching – what we identify as traits if they’ve taken your identity,” Timson said.
Deputy CEO for payments and integrity Chris Birrer said the “archetypal example here would be indicators that the person whose name the claim is submitted in, has not actually submitted it – that somebody else has – either by tricking that person or other forms of identity theft – taken their identity.”
Timson cited a SIM farm operation that harvested Centrelink customers details via malicious links sent in millions of text messages.
“People have clicked on that link and [the attackers have] harvested your name, and then someone’s starting to change your bank accounts,” he said.
“How do we [Services Australia] actually ‘forward lean’ to protect you because someone else is trying to get into your account and redirect payments.”
Timson said the technology is aimed at “someone who we suspect is not who they say they are”, rather than conducting broad-brush identity checks.
A second fraud-related check using machine learning is to aid “prepayment checks” around Australian government disaster relief payments.
“We aim to process those claims as fast as possible because people have been impacted by a disaster, and we want to get money into their bank accounts – as long as it’s not a fraudster’s bank account,” Birrer said.
“We have a system where we identify, through a number of potential checks, where a claim might potentially be fraudulent, and then staff look at it, and either release the claim because they don’t think it’s fraudulent, or they do something like follow up with the customer to check to see their identity or whether or not the bank account is indeed their bank account.
“What we’re looking at here is how to further refine that process … to help to predict certain anomalies which mean it’s more likely to be a fraudulent claim.”
Debt backlog reduction
The report in Information Age also revealed a “debt prioritisation” trial also involving machine learning technology.
This was characterised in the report as a triage exercise to aid efficiency.
The agency clarified at senate estimates that the trial is not about raising debts, but about detecting a subset of cases likely to be “finalised, no debt”, removing them from a “backlog” of debt decisions needing to be made.
General manager of payment assurance, program and appeals Robert Higgins said this was “not an insubstantial number”. Previous audits have put this at about seven percent of debt determinations.
Services Australia’s CEO David Hazlehurst characterised the trial as “a mechanism for us with a backlog of potential debts to say which of these are most likely to not result in a debt – and let’s get rid of those quickly.”
“The machine is identifying which ones are most likely to be that. A person still makes the decision about finalising that matter,” Hazlehurst said.
“It’s not taking the human out of the loop. It’s simply a process of helping us try to be more efficient in getting through the potential debt backlog.”
Birrer said the model could also help in internal allocation of potential debt cases, noting that not all staff were equipped to handle more complex cases, which could delay determinations from being made.
“One inefficiency we have is being able to allocate the right type of work of the right complexity to the skill of the staff member,” Birrer said.
“‘Finalised, no debt’ is amongst the easiest of the debt work, and so if there’s new staff within the payments and integrity group it’s a good thing for them to do as they’re on their skills ladder.
“If you’ve got a very large and complex potential debt that relates to something that happened historically where there might be different policy settings and they need to do a retrospective calculation there, that’s a much more complex task, it also helps.
“We do know that sometimes staff get allocated work they’re not skilled to do, they’ll do a bit of it, their time isn’t used productively, because they throw it back into the pool for somebody else to pick up.”
Entitlements decisions
Services Australia’s CEO David Hazlehurst said there are “no current plans to use AI” to make decisions about entitlements.
“We’ve got a long list of things that we would consider before we could do anything in relation to that,” he said.
Minister for government services Katy Gallagher said that “there would be some level of government decision-making involved in that as well.”
“A decision to move into that space would have to be elevated, I would think,” she said.
Agency officials repeatedly said the machine learning uses were only at a trial phase, and that “any further movement” towards production usage would require multiple gates and assessments.