Should We Trust AI to Sentence Criminals?
When judges, correctional authorities, and parole boards are making sentencing, supervision, and releasing decisions, they're predominantly trying to peer into the offender’s future to discern the person’s potential for recidivism*. To help guide these determinations - which are no doubt influenced by our contemporary infatuation with AI - authorities are progressively turning towards risk assessment instruments (RAIs) on the speculation that their AI can accurately identify those likely to be repeat offenders.
A new study in Science Advances more rigorously confirms that algorithmic judgements may be more accurate than humans. What’s concerning, though, is given the stakes involved - future crimes, a defendant’s freedom or continued incarceration - may still not be reliable enough to ensure that justice is well and truly deserved and that tragic faults can be avoided.

U.S. State and Federal Incarceration Relative to All Reported Crimes (1970-2014). Punishment rate is calculated based on the number of people incarcerated per year rather than convictions in a given year. Image Credit: National Archive of Criminal Justice Data, Bureau of Justice Statistics.
The new study, led by computational social scientist Sharad Goel of Stanford University, is hypothetically a reply to a recent work by programming expert Julia Dressel and digital image specialist Harry Farid. In earlier research, participants attempted to predict whether or not any of the 50 individuals would reoffend of any kind within the next two years, based on a short summary of their case histories. (No images or racial/ethnic information was provided to participants to avoid a skewing of results due to related biases.) The average accuracy rate participants achieved was 62%.
The same criminals and case histories were also processed through a widely used RAI called COMPAS, which stands for “Correctional Offender Management Profiling for Alternative Sanctions.” The accuracy of its predictions was similar: 65%, leading Dressel and Farid to conclude that COMPAS is no more accurate than predictions made from people with little or no expertise in the criminal justice system.

Human v COMPAS algorithmic predictions from 1000 defendants. Image Credit: Science Advances
But what are the repercussions of using an algorithm to predict the trajectory of someone’s life?
The US imprisons more people than any other country in the world. At the end of 2016, approximately 2.2 million adults were being held in prisons or jails, and an additional 4.5 million were in other correctional facilities. Put another way, 1 in 38 adult Americans were under some form of correctional supervision.
Under immense pressure to reduce prison numbers without increasing the rate of crime, courtrooms across the US have turned to automated tools in attempts to move defendants through the legal system as efficiently and safely as possible.
AI is used in almost every task imaginable. Police departments use algorithms to devise where they should send their ranks. Law enforcement agencies use facial recognition systems in order to certify their suspects. These practices have accumulated well-deserved scrutiny for whether they improve safety within our society or maintain existing inequities. Researchers and civil advocates, for example, have repeatedly demonstrated that facial recognition systems can fail dramatically, particularly those individuals who have darker skin - even mistaking members of Congress to be convicted criminals.
But the most controversial tool? Criminal risk assessment algorithms.
Risk assessment tools are designed to do one thing: take in the details of a defendant’s profile and spue out a recidivism score. A judge then factors that score into other countless decisions that can determine what type of rehabilitation services particular defendants should receive, whether they should be held in jail before their trial, and the severity of their punishments. A low score is associated with a kinder fate. A high score is exactly the opposite.

These histograms show that scores for white defendants were skewed toward lower-risk categories, while black defendants were evenly distributed across scores. In the two-year sample, there’s were 3,175 black defendants and 2,103 white defendants, with 1,175 female defendants and 4,997 male defendants. There were 2,809 defendants who recidivated within tow years in this sample. Image Credit: Pro Publica
The logic for using such algorithmic tools is that if you can accurately predict criminal behaviour, you can allocate the resources accordingly such as severity of punishments and rehabilitation services. In theory, it’s perceived that it reduces any bias influencing the process as judges are making conscious decisions based on data rather than their gut.
But here’s the problem.
Modern-day risk assessment tools are often driven by algorithms which have been fed historical crime data.
Machine-learning algorithms use statistics to find patterns in data. So, if you feed it historical crime data, it will recognise and pick out the patterns associated with crime. But these patterns are statistical correlations - not the same as causations. For instance, if low income was correlated with high recidivism, it would leave you confused about whether low income actually caused crime. This is undoubtedly what risk assessment tools do: turn correlative insights into causal scoring mechanisms.
Now onto those populations that have historically been disproportionately targeted by law enforcement are being discriminated against as they are at a higher risk of being imposed with high recidivism scores. As a result, the algorithm could amplify and perpetuate embedded biases and give rise to more bias-tainted data to feed a vicious cycle. As most risk assessment algorithms are proprietary, it’s also impossible to hold them accountable or cross-examine their decisions.
The debate over these tools is still raging on. In July 2018, more than 100 civil rights and community-based organisations, including the ACLU and the NAACP, signed a statement admonishing against the use of risk assessments. Simultaneously, more and more jurisdictions and states, including California, have turned them in a prayer to fix their overburdened jails and prisons.
Data-driven risk assessment is a way to sanitize and legitimise oppressive systems. It’s a way to draw attention away from the actual problems affecting low-income and minority communities, like defunded schools and inadequate access to health care.
We are not risks. We are needs.
* The likelihood a convicted criminal will reoffend
Article Thumbnail Credit: https://theconversation.com/why-using-ai-to-sentence-criminals-is-a-dangerous-idea-77734