Why a pastor who abused children served less prison time than a low-level cannabis dealer
Monday, 12 August 2019
The justice system is increasingly utilising artificial intelligence and computer algorithms in sentencing decisions. Are those algorithms biased? Joel MacManus reports.
Philip David Wallace was a small-time Dunedin drug dealer. When he was asked for a few tinnies of cannabis and some ritalin, he was more than happy to oblige. He soon learned his mistake: he had just sold to an undercover police officer.
Neil Rischbieter was an Auckland pastor who, over the course of several years, repeatedly sexually abused two girls aged between 12 and 16. In one case, he took advantage of the young girl's emotional vulnerability while her mother was dying of cancer.
They received almost identical sentences. For Wallace, it was two years and three months in prison. For Rischbieter, slightly less: two years and two months.
**READ MORE:
* Independent watchdog needed to probe Government's use of AI: experts
* Algorithms are everywhere but the public sector seems scared to use them
* Potential law change could make services explain data decisions**
When Wallace came up for parole, it was declined almost instantly. The board ruling was about as definitive as it can get: 'The man can and will reoffend.'
When Rischbieter came up for his second parole hearing, the result was different. He was released , barely halfway through his sentence. The parole board ruling said he was 'considered to be at low risk'.
Speaking to Stuff, the families of his victims said they felt let down by the justice system.
'She has been scarred for life and I absolutely fear for her future. I have nothing but contempt for this man,' said the stepmother of one girl.
It's hard to say exactly why Rischbieter got such a short sentence and was paroled early, while Wallace wasn't. The intention of this reporting is not to cast judgement on the sentences and subsequent parole board decisions.
And the justice system is complicated and there are a multitude of factors at play. But let's consider the small number that appears beside their names in court documents.
Phillip David Wallace … 0.45346. Neil Rischbieter … 0.0608. That number is their Roc*Roi score (pronounced 'rock roy').
It stands for Risk of Conviction and Risk of Reimprisonment and it's a mathematical algorithm that the justice system uses to help determine how they treat offenders.
It's used by parole boards to help determine whether it's safe to release someone into the community, and in prison to determine what kind of rehabilitation programmes are available.
Roc*Roi plays a part during initial sentencing too. Unlike parole boards, judges aren't shown an offender's exact score, but they are given a rough estimate of their score as low/medium/high risk.
The computer calculated that Phillip Wallace has a 45 per cent chance of reoffending. Neil Rischbieter was deemed to have just a 6 per cent chance.
It's not possible to offer a judgment on the sentencing of the two men from the outside. They committed very different offences and had different backgrounds. Wallace also had several previous offences, including some for violent assault, which counted against his score.
But what these numbers show is just how much their lives were influenced by a computer that spat out a tiny number that they may not have known even existed.
CALCULATED RISK
Roc*Roi is just one example of an increasing trend in New Zealand and around the world of government agencies using algorithms and artificial intelligence to solve complex problems.
The appeal is obvious. Human decision-making is riddled with cognitive bias and logical fallacies. Machines are faster and more efficient than us, they are capable of taking into account far more factors, and they can't be swayed by gut feeling.
There's a tendency to think that algorithms are a way to eliminate bias. But Otago University law professor and artificial intelligence expert Colin Gavaghan says that's a misconception.
'We might wonder 'how can a computer be biased or prejudiced?' In a sense, a computer can't. But it's only as good as the information that's fed into it, and if that info is tainted by historical bias, that's going to be baked into the new system,' Gavaghan said.
Algorithms are statistical tools. They aggregate data. But all that data is formed by human decisions, and if those decisions are flawed, the output will be flawed.
It's like baking a cake. You could have the best oven in the world, but if your eggs have gone bad, it's not going to be a very good cake.
Here are the 16 pieces of information used to calculate a Roc*Roi score:
Gender
Age
Age at first offence
Frequency of convictions
Number of court appearances
Number of convictions
Total years in prison
Number of prison sentences
Whether the most recent crime resulted in a prison sentence
Maximum prison sentence
Seriousness rating of all previous crimes
Weighted past seriousness measure
Most serious past crime
Average seriousness of past crimes
Offence category
Number of convictions
All of those data points (apart from gender and age) are a result of human decisions. Police have to choose to stop someone, then choose to arrest them, and choose whether to give a diversion or warning. Prosecutors have to decide whether to charge, judges have to decide whether they are guilty, and how long to sentence them for.
Any bias in the way police, courts, or judges act toward a certain group will be reflected in their scores. And we know that our justice system has some significant issues.
From 2005-2014, police apprehended roughly the same number of Māori and Pākehā, despite the Pākehā population being three times higher.
Of those arrested, Māori were less likely to be let off with a warning and more likely to face prosecution. Of those prosecuted, Māori were more likely to be given a prison sentence.
When you compare young Māori and non-Māori with the same self-reported offending history and social background, Māori are between 1.6 to 2.4 times more likely to have a criminal conviction.
All of those things are factors taken into account by Roc*Roi, and all of them are weighted against Māori.
Every step along the way, a Māori offender is more likely to face punishment, and that in turn increases their Roc*Roi score, which increases the chance they face harsher punishment, which increases their Roc*Roi score, and so on.
Similar feedback loops have been seen time and time again internationally.
A policing tool called PredPol, used in the US and UK, touted itself as using advanced analytics to identify high-crime areas, right down to a 46-square-metre box.
For the Los Angeles Police Department, there was some hope that an objective computer system could quell concerns of police profiling of black residents. In reality, the reverse happened.
PredPol processed data from previous arrest to make its decisions. But those previous arrests were tainted. Police were already spending more time in black neighbourhoods and making more arrests of black people.
That meant PredPol identified those neighbourhoods as high-risk areas, which meant police spent even more time there, which meant police made even more arrests, often for very minor crimes like loitering and tagging.
An artificial intelligence tool built by Amazon specifically to remove bias in the hiring process was scrapped last year after it turned out the program had taught itself to penalise women.
The Amazon tool used a neural network, a form of machine learning meant to replicate the way the human brain processes information.
The program was trained to identify good candidates based on the resumes of employees hired at the company in the past 10 years. But because the company had historically hired so many men, the AI believed that a good candidate meant a male candidate.
The machine taught itself to penalise graduates of women's colleges, members of women's clubs, or any other resume which included the word 'woman' or 'women'.
The problem is that algorithms, with their exact, definitive outcomes, have a tendency to seem like all-knowing, objective truth.
Gavaghan said this kind of over-reliance is one of the most concerning aspects of artificial intelligence.
'Imagine your company has just spent millions on a fancy new system and you're told to speak up if you think it's wrong - that's quite a challenging thing to do,' he said.
MAKING THE NUMBERS WORK
Roc*Roi has been used in New Zealand for 20 years now.
Corrections research and analysis manager Peter Johnston said he had 'high confidence in its accuracy, given repeated exercises, which show a high degree of correlation between predicted rates of reimprisonment, and actual rates of reimprisonment.
'Our researchers are well-informed of limitations, including the fact that accuracy can diminish over time, and are carrying out work to recalibrate the algorithm to ensure it is as accurate as it can be.'
There have been many tweaks made to Roc*Roi over the years. The original iteration, rolled out in 1999, featured one major difference to the version used today: it took 17 factors into account, one more than the 16 it includes today.
That 17th factor was ethnicity.
Māori offenders were rated more likely to reoffend than Pākehā, simply because of their race.
Among people with a relatively short criminal history, that had an especially strong effect. Ethnicity could account for as much as 20 per cent of their Roc*Roi score.
In 2002, Tom Hemopo, a Māori probation officer, took a claim to the Waitangi Tribunal arguing Roc*Roi was prejudiced.
Internally, staff were saying the same thing, as revealed by one email exchange between two of the algorithm's creators:
'We are also having a lot of heat applied from some staff about the fact that Roc*Roi is 'racist' because it penalises Maori. This is an issue that may require us to modify the algorithm to remove ethnicity. Can you confirm what reduction we would get in accuracy if we did remove this variable? Are there any other variables that are highly correlated with ethnicity that could be used instead?'
Ethnicity was quietly removed from the algorithm before the claim came before the tribunal. The official word was that it was removed for 'reasons of statistical goodness of fit'.
A similar issue arose with an algorithm police use to analyse the risk profile of youth offenders, the Young Offender Risk Screening Tool (YORST).
YORST originally used the decile rating of the nearest primary school to the offender's home as a factor. It meant that people from poorer communities were systematically overscored, and therefore more likely to wind up in Youth Court.
That was removed some time in the early 2010s.
Those changes were only made possible because the New Zealand Government commissioned regular reviews into the systems, and made the analysis publicly available.
It's a very different story in the United States. An algorithm called COMPAS is used in many states, including New York and California, to predict reoffending. It's similar to Roc*Roi in a way, but far more complex, and includes a psychological assessment as part of the score.
Unlike Roc*Roi, COMPAS wasn't built by the government. It was built by a private software company and licensed to the states.
That means the public can't actually see under the hood and check the maths, because it's protected by intellectual property laws.
Wisconsin man Eric Loomis, who pleaded guilty to eluding an officer and was sentenced to six years in prison, argued that because he could not see the inner workings of the algorithm, he had been denied due process.
He couldn't dispute the score, because he didn't know how it had been calculated.
A report by the New Zealand Law Foundation, authored by Gavaghan and his Otago University colleagues James MacLaurin and Alistair Knott, recommended establishing an independent watchdog organisation to oversee the use of algorithms and artificial intelligence by government, especially when the scores the systems produce can have a serious impact on someone's life.
They want to see regulations requiring that all algorithms be open and transparent, so everyone can see the back end of the equations and understand why a certain decision has been made.
The latest figures show there are 32 algorithms being used across 14 government agencies to automate outcomes for New Zealanders. They affect everything from visa applications, to ACC claims, to how we schedule buses to get kids to school.
Moving forward, those systems are only going to get more advanced, and that brings with it more complexities. But it also brings a huge amount of opportunity to provide better, fairer, more informed outcomes.
Gavaghan just wants us all to know how they work.
'If you're turned down for parole, or you're told you're not getting your children back that have been taken off you, that might be the right decision, or it might be the wrong one, but you can only challenge it if you know how it was made.'
Correction: This story has been updated to clarify that the cannabis dealer served more time than the pastor but not twice as much. The headline initially read: Why a pastor who abused children served half as much prison time as a low-level cannabis dealer. It has also been updated to clarify that Rischbieter was paroled at his second hearing.