Beware of governments using algorithms to collect revenue

A great advantage of having children and grandchildren is that they can show you how to do things on the internet – or your phone – that you can’t for the life of you see how to do yourself. But a small advantage that oldies have over youngsters is that we can remember how much more clunky and inconvenient life used to be before the digital revolution.

When you had to get out of your chair to change the telly to one of the other three channels. When you spent lunchtimes walking to companies’ offices to pay bills. When you had to visit your own branch of a bank to get cash or deposit a cheque.

When you had to write and post letters, rather than dashing off an email or text message. When we were paid in notes and coins rather than via a direct credit. When buying something from a business overseas was too tricky to ever contemplate.

Computers connected by the internet are transforming our world, making many of the things we do easier, more convenient, better and often cheaper. Businesses are adopting new technology because they see it as a way to cut costs, compete more effectively, attract more customers and make more money.

Governments have been slower to take advantage of digitisation, websites, artificial intelligence and machine learning. But there’s a lot to be gained in reduced red tape, convenience for taxpayers, efficiency and cost saving, and now governments at all levels are stepping up their use of new technology – which most of us would be pleased to see.

But, as with so many things, new technology can be used for good or ill. The most egregious case of government misuse of technology was surely Centrelink’s “automated income compliance program” aka robo-debt.

Here, a dodgy algorithm was used to accuse people on benefits of understating their income over many years and to demand repayment. Despite assurances by the two ministers responsible – Christian Porter and Alan Tudge – that all was fair and above board, huge anxiety was caused to many unjustly accused poor people.

The Coalition government obfuscated for years before a court finally ruled the program unlawful and the government agreed to return $1.8 billion. For a scheme intended to cut costs, it was an immoral own goal.

But last week the NSW Ombudsman, Paul Miller, revealed that something similar had been going on in NSW. Between 2016 and 2019, the state’s debt-collection agency, Revenue NSW, unlawfully used an automated system to claw back unpaid fines from financially vulnerable people, in some cases emptying bank accounts and leaving them unable to buy food or pay rent.

The automated system created garnishee orders, requiring banks to remove money from debtors’ bank accounts, without the debtors even being informed. The computer program took no account of any hardship this would impose.

Between 2011 and 2019, the number of garnishee orders issued by Revenue NSW each year jumped from 6900 to more than 1.6 million. The Ombudsman began investigating after receiving “a spate of complaints from people, many of them financially vulnerable individuals, who had discovered their bank accounts had been stripped of funds, and sometimes completely emptied.

“Those people were not complaining to us about the use of automation. They didn’t even know about it.”

As usually happens, we find out about this long after the problem has been (apparently) rectified. Unlike the feds, the NSW Revenue office agreed to modify its process in 2019, so that garnishee orders are no longer fully automated.

People identified as vulnerable were excluded from garnishee orders. And a minimum amount of $523.10 (!) is now left in the garnished account.

But Revenue NSW didn’t seek advice on whether the original scheme was legal, so the Ombudsman did. It wasn’t. The lawyers say the law gives the right to extract money from people to an “authorised person” – not to a machine.

So, two isolated incidents – one federal, one state – and everything’s now fixed? Don’t be so sure. Miller says that, in NSW, other agencies are known to be using machine technologies for enforcement of fines, policing, child protection and driver licence suspensions.

How do we know it isn’t happening – or will happen – in other states?

NSW Finance Minister Damien Tudehope said garnishee orders were a last resort after a person had been contacted multiple times in writing and given options about overdue fines.

“For those who have chosen to ignore our notices and simply don’t want to pay, the community has an expectation we take action to recover what is owed,” he said.

Yes, but few of us want governments riding roughshod over people who can’t pay because they’re on poverty-level social security benefits.

I leave the last word to Anoulack Chanthivong, NSW opposition finance spokesman: “The pursuit of economic efficiency through artificial intelligence should never come at the expense of treating our more vulnerable citizens with dignity and decency.

“Governments at all levels are meant to serve our community and make their lives better, not find unlawful ways to make them worse.”

Read more >>

How the threat from robots was exaggerated

You'd have to have been hiding under a rock not to know that 40 per cent of jobs in Australia – about 5 million of them – are likely to be automated in the next 10 to 15 years.

Ask a young person what they know about the future of work and that's it. Which may help explain why so many of them seem angry and depressed about the economic future they're inheriting.

This information is widely known because it's the key finding of a major report, Australia's future workforce?, published in 2015 by the Committee for Economic Development of Australia, a well-regarded business think tank, derived from modelling it commissioned.

It's the sort of proposition you see many references to on social media, particularly because it chimes with a similar widely known prediction made in 2013 that 47 per cent of American jobs could be automated in the next 20 years.

Neither figure is a fact, of course, just a prediction about the distant future based on "modelling".

Why is it that if a prediction is big enough and gloomy enough, everyone keeps repeating it and no one thinks to question it? Why do we accept such frightening claims without asking for further particulars?

Why doesn't anyone ask the obvious question: how – would – they – know?

Because the prediction is based on "modelling"? That if it came out of a computer, it must be true?

Because the modelling for Australia reached similar results to the modelling for America? Sorry, it's actually the same model applied to different figures for each country.

Fortunately, not everyone is as easily convinced that the sky is falling. Two economists from Melbourne University, Professor Jeff Borland and Dr Michael Coelli, have taken a very hard look at the modelling undertaken for the committee by Professor Hugh Durrant-Whyte, of Sydney University, and other engineers at National Information and Communication Technology Australia.

Durrant-Whyte's modelling simply applies to Australia modelling of US occupations by Carl Frey, an economic historian, and Dr Michael Osborne, an engineer, of the Oxford Martin School for a sustainable future at Oxford University.

Frey and Osborne provided some colleagues with descriptions of 70 US occupations and asked them to judge whether they were "automatable" or not. This sample was then analysed and used to classify all 702 US occupations according to their likelihood of being automated.

Any occupation with a predicted probability of automation of more than 70 per cent was classed as being at "high risk" of automation.

Borland and Coelli make some obvious criticisms of this methodology. First, the colleagues found that 37 of the sample of 70 occupations were at risk of automation. Should these subjective assessments prove wrong, the whole exercise is wrong.

For instance, the colleagues judged that surveyors, accountants, tax agents and marketing specialists were automatable occupations, whereas Australian employment in these has grown strongly in the past five years.

Frey and Osborne say the need for dexterous fingers is an impediment to automation, but their method predicts there is an automation probability of 98 per cent for watch repairers.

Second, Frey and Osborne's modelling makes the extreme assumption that if an occupation is automated then all jobs in that occupation are destroyed. The advent of driverless vehicles, for instance, is assumed to eliminate all taxi drivers and chauffeurs, truck drivers, couriers and more.

Third, their modelling assumes that if it's technically feasible to automate a job it will be, without any need for employers to decide it would be profitable to do so. Similarly, it assumes there will be no shortage of the skilled workers needed to set up and use the automated technology.

More broadly, their modelling involves no attempt to take account of the jobs created, directly and indirectly, by the process of automation.

No one gets a job selling, installing or servicing all the new robots. Competition between the newly robotised firms doesn't oblige them to lower their prices, meaning their customers don't have more to spend – and hence create jobs – in other parts of the economy.

All that happens, apparently, is that employment collapses and profits soar. But if it happens like that it will be the first time in 200 years of mechanisation and 40 years of computerisation.

In 2016, the Organisation for Economic Co-operation and Development commissioned Professor Melanie Arntz and colleagues at the Centre for European Economic Research to offer a second opinion on Frey and Osborne's modelling.

Arntz and co noted that occupations categorised as at high risk of automation often still contain a substantial share of tasks that are hard to automate.

So they made one big change: rather than assuming whole occupations are automated, they assumed that particular tasks would be automated, meaning employment in particular occupations would fall, but not be eliminated.

They found that, on average across 21 OECD countries, the proportion of jobs that are automatable is not 40 per cent, but 9 per cent.

Those countries didn't include Oz, so Borland and Coelli did the figuring – "modelling" if you find that word more impressive – and found that "around 9 per cent of Australian workers are at high risk of their jobs being automated".

Why are we so prone to believing those whose claims are the most outlandish?

Read more >>