【Thriller Archives】

If the past few years have Thriller Archivestaught us anything, it's that algorithms should not be blindly trusted.

The latest math-induced headache comes from Australia, where an automated compliance system appears to be issuing incorrect notices to some of Australia's most vulnerable people, asking them to prove they were entitled to past welfare benefits.

Politicians and community advocates have called foul on the system, rolled out by Australia's social services provider, Centrelink.

Launched in July, the system was intended to streamline the detection of overpayments made to welfare recipients and automatically issue notices of any discrepancies.

The media and Reddit threads have since been inundated with complaints from people who say they are being accused of being "welfare cheats" without cause, thanks to faulty data.

The trouble lies with the algorithm's apparent difficulty accurately matching tax office data with Centrelink records, according to the Guardian, although department spokesperson Hank Jongen told Mashableit remains "confident" in the system.

"People have 21 days from the date of their letter to go online and update their information," he said. "The department is determined to ensure that people get what they are entitled to, nothing more, nothing less."

Independent politician Andrew Wilkie accused the "heavy-handed" system of terrifying the community.

"My office is still being inundated with calls and emails from all around the country telling stories of how people have been deemed guilty until proven innocent and sent to the debt collectors immediately," he said in a statement in early December.

The situation is upsetting albeit unsurprising. The siren call of big data has proved irresistible to governments globally, provoking a rush to automate and digitise.

What these politicians seem to like, above all, is that such algorithms promise speed and less man hours.

Alan Tudge, the minister for human services, proudly announcedthat Centrelink's system was issuing 20,000 "compliance interventions" a week in December, up from a previous 20,000 per year when the process was manual. Such a jump seems incredible, and perhaps dangerous.

As data scientist Cathy O'Neil lays out in her recent book Weapons of Math Destruction, the judgments made by algorithms governing everything from our credit scores to our pension payments can easily be wrong -- they were created by humans, after all.

The math-powered applications powering the data economy were based on choices made by fallible human beings. Some of these choices were no doubt made with the best intentions. Nevertheless, many of these models encoded human prejudice, misunderstanding and bias into the software systems that increasingly managed our lives. Like gods, these mathematical models were opaque, their working invisible to all but the highest priests in their domain: mathematicians and computer scientists.

These murky systems can inflict the greatest punishment on the most vulnerable.

Take, for example, a ProPublicareport that found an algorithm being used in American criminal sentencing to predict the accused's likelihood of committing a future crime was biased against black people. The corporation that produced the program, Northpointe, disputed the finding.

O'Neil also details in her book how predictive policing software can create "a pernicious feedback loop" in low income neighbourhoods. These computer programs may recommend areas be patrolled to counter low impact crimes like vagrancy, generating more arrests, and so creating the data that gets those neighbourhoods patrolled still more.

Even Google doesn't get it right. Troublingly, in 2015, a web developer spotted the company's algorithms automatically tagging two black people as "gorillas."

Former Kickstarter data scientist Fred Benenson has come up with a good term for this rose-coloured glasses view of what numbers can do: "Mathwashing."

"Mathwashing can be thought of using math terms (algorithm, model, etc.) to paper over a more subjective reality," he told Technical.lyin an interview. As he goes on to to describe, we often believe computer programs are able to achieve an objective truth out of reach for us humans -- we are wrong.

"Algorithm and data driven products will always reflect the design choices of the humans who built them, and it's irresponsible to assume otherwise," he said.

The point is, algorithms are only as good as we are. And we're not that good.