The use of AI in the US criminal justice system has become pervasive across all areas. AI assists in identifying crimes through predictive policing and audio sensors for gunshot alerts. It helps to solve cases through the probabilistic software used by forensic experts to aid in the identification of fingerprints, faces, and DNA. Finally, the use of risk assessment instruments support decisions such as setting bail, how long to incarcerate someone, and whether to grant them parole.
Many of these AI programs have been created with the best intentions: to reduce human bias, improve accuracy, and make the system more effective… Those in favor of risk assessment instruments suggest that by more accurately predicting recidivism, incarceration rates can be reduced. Unfortunately, the data used to train these systems can often contain biases which result in the system itself becoming biased, all under “a veneer of scientific objectivity”. This problem is further aggravated by the fact that these biases are hidden within the inner workings of these algorithms, which due to their privately owned status and claims on trade secrets, remain inaccessible to challenge. As a result, there are private companies which essentially determine “what neighborhoods to police, whom to incarcerate and, for how long”.
COMPAS
The Correctional Offender Management Profiling for Alternative Solutions (‘COMPAS’) software uses an algorithm to create a risk scale for general recidivism, violent recidivism, and pretrial misconduct to aid the decision of judges. The software is used across various federal courts in the US, such as New York, Wisconsin, California, Florida’s Broward County, etc. It was created as an attempt to deal with the human biases involved in decision-making, which are so prevalent that one study demonstrated that judges were more likely to make a favorable decision in a parole hearing straight after lunch.
The problem with the system is that due to its proprietary and black-box nature, it is difficult for individuals to challenge an inconsistent result. Glenn Rodriguez was denied parole due to his high COMPAS score. Yet, Rodriguez had not had a single disciplinary infraction in the last decade and was generally a model of rehabilitation. The board was not able to give him an explanation for his high-risk score and the way that COMPAS functioned was considered a trade secret.
Without being able to challenge his score Rodriguez would have been stuck in prison if it weren’t for his perseverance and ingenuity. Rodriguez asked to see the COMPAS score of his fellow inmates and in doing so worked out that his offender rehabilitator coordinator had checked the wrong answer to question 19, a question he later found out asked: “Does this person appear to have notable disciplinary issues?” The change in the response dropped his risk evaluation score from an 8/10 to a 1/10.
One of the basic tenets of the criminal justice system is that individuals should be able to challenge the evidence presented against them. The secret nature of the COMPAS algorithm prevents access to justice by withholding this evidence from defendants. While there are good reasons to protect confidential commercial information, the value of these laws must be balanced against the damage of withholding information from the defense in criminal proceedings.
Another concern presented by the inscrutability of predictive algorithms such as COMPAS is that people assume that these algorithms are far more advanced and accurate than they are. Dartmouth professor Dr. Hany Farid reverse-engineered the COMPAS system and ran studies to test its accuracy. He found that when random people on the internet were given 7 classifiers they were just as accurate as predicting recidivism rates COMPAS. Farid also found that he could create software using only two classifiers (age and prior crimes) with the same accuracy as COMPAS. This suggests that COMPAS is far simpler and less accurate than one might imagine after hearing about an algorithm created using big data and machine learning. As a result, judges may overestimate the sophistication of the software and rely more heavily on a COMPAS scoring system than they should.
TrueAllele
TrueAllele is a program created by Cybergenetics and used in forensic identification to analyze and identify DNA using probabilistic genotyping. This means that when a crime is committed and there is incomplete or an almost inscrutable amount of DNA, it is run through the TrueAllele program and it calculates how likely the DNA is to have come from a particular person.
For many years the courts have rejected all requests to access the source code of TrueAllele. At times leaving defendants, such as in the case of Michael Robinson, wrongfully accused based on misleading “proprietary” DNA software that he could not access or challenge. TrueAllele’s source code remained inaccessible for many years, even as its competitors (such as STRmix) shared their source code and continued to perform just fine in the market. Finally, in 2021 two judges (one in New Jersey and one in Pennsylvania) ordered prosecutors to hand over the source code for TrueAllele.
In cases with such high stakes for individuals, it is not just inappropriate but downright irresponsible to avoid scrutinizing algorithms which produce evidence used in criminal hearings. Particularly, when this evidence is held to such high standards.
PREDPOL
PredPol is described as an: “algorithmic, place-based predictive policing system designed to forecast the times and places — mapped to a 500-foot by 500-foot city-wide grid of neighborhood cells or hot spots—where property crimes, such as car thefts and burglaries, might occur.”
In response to public pressure, in 2015 PredPol published a description of its algorithm that enabled researchers to examine it. Tests found that the application of the algorithm to police records meant that even when crimes occurred evenly throughout a city, PredPol would concentrate on areas that were over-represented in the police database. Thus, resulting in over-policing in those areas, which consequently led to greater crime being detected, justifying the initial actions.
The PredPol program has also been criticized for not being particularly sophisticated. While some secrecy is needed when it comes to policing: “If anyone could predict IRS audits or airport security screenings, fraudsters and terrorists could avoid getting caught”. There must be a balance between attaining enough transparency to identify flaws in systems and programs like PredPol.
If we are to continue using AI in justice systems, we must be honest with ourselves and accept that most of these programs are just in their infancy. It’s important to be aware that these programs are far less accurate than we give them credit for. Perhaps most importantly, they are going to need extensive testing and research, which means algorithms created by private companies for criminal justice use need to be more transparent. Quite simply, it’s highly unethical to prioritize trade secrets for novel technologies over the fundamental values and rights inherent to the criminal justice system.
Written by Celene Sandiford, smartR AI