Wednesday, January 25, 2017

Centrelink debacle much worse than thought

It has become the most widely cited figure in the Centrelink robo-debt debate: that 20 per cent of the debts identified by its data-matching machine are wrong.

But the figure itself is wrong. The true number of mistakes is almost certainly higher, perhaps as high as 90 per cent.

Twenty per cent has become the accepted truth in part because the figure is big - big enough for critics to use to condemn the data-matching program and big enough for Centrelink to use to fob off requests for the truth.

Even Malcolm Turnbull's disenchanted former digital transformation chief Paul Shelter embraced it.

"All I can say is, if they were a commercial company, you would go out of business with a 20 per cent failure rate, a known 20 per cent failure rate, you would go out of business," he told The Guardian this month.

Labor frontbencher Anthony Albanese backed him up, wrongly saying that "on the government's own figures, 20 per cent of people who've been sent debt letters, often accompanied by threats of debt collection agencies being involved, have been sent them on a false basis".

The 20 per cent isn't the proportion of debt letters sent out that are false. We won't know that for a long time, if ever. Some people have been paying up even when the debt letters are wrong, sometimes because they don't have the records to argue otherwise, sometimes because they trust the government, and sometimes because they can't be bothered dealing with Centrelink.

A Centrelink whistleblower alleges that, disgracefully, staff have been ordered not to use information in Centrelink's possession to correct false debt notices. Another says that of hundreds of debt notices reviewed, only a few dozen (at a "generous estimate") turned out to be correct.

Here is where the 20 per cent figure comes from. Between July and December, Centrelink's computer sent out 232,000 letters asking people to log on to a website to confirm or update their income history. Around 169,000 did so. (An email to Fairfax Media from the office of Human Services Minister Alan Tudge implies that none of the 63,000 who did not log on have been issued with debt notices. Their cases are "are still active and in progress or require further review".)

When the 169,000 logged on and ticked a box or corrected the income information, they were presented with an instant, on-screen estimate of money owed. Despite weeks of requests, the minister's office has been unable to tell Fairfax Media the proportion who were told they owed money. It's likely to be extremely high. The software can't account for income that is exempted from the Centrelink tests and it averages annual income to often falsely conclude that the averaged income is received in the weeks when the recipients are on sickness or other benefits, among other flaws.

But the online estimate isn't a debt notice. That is sent later, after a human assesses the machine's conclusions. A whistleblower says staff are instructed to only lightly assess those conclusions, letting most of them through.

Eighty per cent is the proportion of the 169,000 who are eventually sent letters with demands to pay. Twenty per cent is the proportion who are not. There's no reason to think the proportions who have been wrongly assessed by the computer and wrongly issued debt notices are not much, much higher.

Fairfax Media asked the minister's office three weeks ago for the proportion of computer debt assessments that those who logged on objected to, but despite repeated reminders, still hasn't been told.

The error rate may be even higher than the objection rate. Given what's known about the design of the system, there's no reason to think it is not north of 90 per cent. Twenty per cent would be bad enough (as the Prime Minister's former digital chief says, enough to put a private sector firm out of business) but it's probably far short of the truth.

UPDATE: Department of Human Services spokesman Hank Jongen later responded and said the figure of 90 per cent was wrong. "There is no basis for suggesting the error rate is 90 per cent. This is a figure plucked out of the air that doesn't serve the public discussion well."

In The Age and Sydney Morning Herald