That old idea that justice is blind operates with considerable difficulty in practice. Various studies show that in the United States, individuals belonging to minorities are treated unfavourably at different stages of the judicial system where, for example, judges impose on African American defendants harsher bail conditions than on white ones.
It would not be strange to find similar discrimination in some Latin American countries. Although we do not yet have a systematic review for the region, it is not hard to believe that the low levels of trust in the judicial system are related to less fair treatment of certain groups of the population. This, moreover, may be connected to one of the most worrying justice phenomena in the region: the large number of people held in pretrial detention. Prisons are packed with people awaiting trial, many of them poor. 36% of the prison population in the region has not been sentenced, a figure that exceeds 50% in several countries, including Bolivia, Guatemala, Uruguay and Venezuela.
Disparities in the criminal justice system
An important question is whether it would be possible to eliminate unjust discrimination in the judicial system and ensure that judges administer justice consistently in order to avoid some of these problems. There is a heated debate over whether artificial intelligence and the use of algorithms could achieve this.
Proving the existence of bias against a particular group is not straightforward. However, recent research has credibly demonstrated that minorities in the United States receive unfair treatment in areas as diverse as speeding fines, or the fact that, under similar circumstances, some groups have a higher probability of incarceration. Surprisingly, there is even evidence of political bias in sentencing.
On the other hand, there are significant consistency problems that undermine confidence in the judicial system. Like other human beings, judges can be influenced by environmental factors, such as “decision fatigue,” which makes them stricter when granting parole if they have not had the chance to take a lunch break. Or emotional stress when, for example, they hand down longer sentences after their favourite football team unexpectedly loses a match.
Do algorithms work as a solution against discrimination?
In this context, something as seemingly objective as artificial intelligence appears very promising. After all, most people have witnessed the power of algorithms to recognise images and process enormous amounts of data in tenths of a second. Furthermore, one might think that algorithms offer a transparent procedure that can be subject to continuous scrutiny, at least by a group of experts. The question, then, is whether it is possible to harness this technology to improve the impartiality of the judicial system.
A fundamental limitation of algorithms is their capacity to simply reproduce the social prejudices observed in the data. As Solon Baracas and Andrew Selbst point out in a recent paper, an “algorithm is only as good as the data it works with”. In other words, the data can cause algorithms to “inherit the prejudices of prior decision makers” or “simply reflect the widespread biases that persist in society at large”.
In part, these limitations stem from the technical difficulties that are intrinsic to the use of algorithms in the social sciences. One of them has to do with sample sizes. This means that, for example, majority groups of the population are represented more accurately in algorithms because more data exists about them than about other minority groups. Likewise, cultural differences play an important role in the goodness of fit of the models, since patterns that are valid for the general population may not be valid for minorities. Moreover, eliminating these differences is complicated because specifically labelling minority individuals in order to improve the algorithm may be objectionable in itself.
An interesting study in New York
Despite these intrinsic limitations, a recent study offers an interesting perspective on how to combine the use of algorithms and big data with judges’ decisions. This study compares how algorithms would have performed in the New York City judicial system against the decisions actually made by judges when determining which suspects should be released and which should be jailed before trial.
The study featured an ingenious design. Given the impossibility of observing the number of crimes that might have been committed by those who were jailed but could have been released, the authors construct a counterfactual scenario in which they compare suspects with very similar profiles who faced slightly less strict judges and who, given identical records, decided to release them while awaiting trial. By observing the post-release record of those freed by lenient judges, the researchers estimate how the group of people jailed by strict judges would have behaved had they been released. In this way, they compare judges’ decisions (whether or not to order pretrial detention) with those generated by an artificial intelligence algorithm in similar populations.
The study, rather than claiming that bias can be eliminated, shows that algorithms can offer a valuable contribution. In fact, its results suggest that, had judges’ decisions been replaced by the ruling of an algorithm, the total number of crimes would have been reduced by 24.7% with no change in incarceration rates — or, conversely, that incarceration rates could have been reduced by 41.9% without an increase in crime rates. Moreover, those results could have been achieved even as the algorithms reduced the disparities that negatively affect African Americans and Latinos.
Judges face difficult choices when making decisions about people’s liberty. On the one hand, if they release suspects they run the risk that they will commit more crimes, flee or threaten witnesses. On the other hand, if they detain them they may jeopardise the suspects’ jobs, create financial and psychological trauma for them and their families, and even increase the likelihood of their being convicted, especially through guilty pleas.
Latin America probably does not have sufficiently complex data or large enough samples to start using artificial intelligence in court decisions immediately. But it can begin to collect higher-quality data at the individual case level, develop models for artificial intelligence and test their performance against judges in simulations similar to those carried out in the New York study.
An interesting case is under way in Colombia, where several of the results observed in New York have already been replicated. The estimates are now being used to guide prosecutors in promoting a more efficient use of prisons, detaining before trial only those who pose a greater risk to society. While many questions remain open, we must stay attentive to how these or other tools may allow us to reverse the costly and counterproductive practice of holding large numbers of people in pretrial detention.
Originally published by the IDB, Patricio Domínguez
Research economist in the Research Department of the Inter-American Development Bank
