Taking the AI out of fairness: the ineffectiveness of algorithmic sentencing as a remedy for political bias in the legal system
“Politicisation” in the law may conjure up images of Supreme Court nominations and constitutional battles, but it is equally present in the most common interaction people will have with the justice system: sentencing. The two dominant political parties have increasingly split views on the sentencing of certain crimes. It is perhaps predictable that these political views trickle into judges’ sentencing decisions. For instance, after the murder of George Floyd, the Democrats defined themselves largely in favour of the Black Lives Matter movement, whereas many Republican politicians decried the movement’s methods and message. Consequently, it is interesting to note that sentences for violent crimes are 92.3% longer for African-American offenders than white criminals in Republican-leaning counties (Durante, 2021) and that Republican-identifying judges typically “sentence blacks to 1.3 months longer compared to similar white defendants” (Cohen and Yang, 2019). In this sense, the sanctions levied on guilty criminals are directly affected by the personal convictions and ideologies of the judge present. While this politicisation may pose a threat to the legal system, a potential solution of algorithmic sentencing would constitute a dissatisfactory response.
Is politicisation a problem within the law? After all, one could argue that law inherently reflects societal values and sentencing should therefore be based on those values. The flaw herein is that equality before the law is constitutionally enshrined, yet the values that determine sentencing are not universal for all judges. The Fourteenth Amendment calls for “the equal protection of the law” (National Archives, 2021), which would be disrupted by the variation of sentences based purely on moralistic beliefs. Allowing different judges to arbitrarily sentence on different levels would be ethically and legally unjust.
Attempts to reduce this politicisation have been difficult due to judicial immunity. In effect, as long as a judge acts within their jurisdiction, they are legally immune, even if their actions are malicious (Block, 1980). This seems to be a dissatisfactory system of bringing political judges to account. In fact, a Reuters investigation finds that nine out of ten judges who broke the law in the past twelve years remain on the bench (Berens and Shiffman, 2020), so it is unsurprising that political biases are not given much weight either.
One way of addressing this problem would be to implement a system of algorithmic sentencing. This would replace a judge with an algorithm or an artificial intelligence that matches a sentence to a defendant based on the facts of the case. No such system exists yet – the closest is COMPAS, an algorithm that calculates the projected recidivism of a criminal, but the Supreme Court of Wisconsin ruled that it should not be substituted for sentencing (State v. Loomis, 2016). However, if full algorithm sentencing technology were possible, it could have unique benefits. The algorithm would remove human political influences from sentences, thereby providing a more consistent approach to the law. This would mitigate the effect of overly punitive sentences on ethnic minorities. Moreover, the United States faces a backlog of cases, with New Jersey judges handling over 2,700 cases each (Tully, 2021) – an algorithm could automate and simplify the job, making the justice system more efficient. Lawmakers and politicians may also find their work easier when automated sentencing carries through all the latest sentencing reforms, rather than relying on judges to implement the changes in the law. Furthermore, sentences would be harder to manipulate, and therefore, individual cases would be far harder to politicise. Sentences for police brutality would no longer be swayed by partisan interests, and the decisions would have to be accepted. Automated sentencing may therefore stall the wider politicisation of the law by removing individual biases from the system.
However, there are two fallacies within this argument. First, we can highlight the flaws within the algorithms used. Currently, automated sentencing cannot be accurate or nonpartisan. The most commonly used algorithms, such as COMPAS, have error rates of up to 40% because they are based on imperfect sociological theories of recidivism (Forrest, 2021). Similarly, many of these systems are based on historic sentencing data, which itself has been politicised to some degree. Attempts to filter out the politicised data will lend themselves to further politicisation, as it is unclear to what extent each individual case has been politicised. Thus, distinctions between which cases to include and exclude in the algorithm’s training set are themselves political.
Second, the use of algorithms within sentencing does not allow for the development of law. It is clear that the status quo leans increasingly heavily towards politicisation and fragmentation, but that does not mean that this trend is permanent. Increasingly, activists from both parties agree on certain issues; the conviction of Ahmaud Arbery’s murderers indicates progress. Trust in the criminal justice system is low: a Gallup poll shows that only 20% of Americans had “a great deal” or “quite a lot of confidence” in the criminal justice system, with relative consistency across both parties (Brenan, 2021). Legal practitioners are aware of the sentiment, providing an impetus to make the criminal justice system a fairer and less politicised sphere of public life. Setting the law in stone by automating a vital part of the legal system would prevent further developments in criminal law. While algorithmic sentencing may provide a temporarily better alternative to the status quo, it is certainly less desirable than the reforms that I believe are likely in the long run.
Ultimately, we must accept that the current system of criminal sentencing is imperfect. Political parties from both sides have vested interests in legal outcomes and judges seem to be influenced by their beliefs. However, algorithmic sentencing is by no means a panacea. While the theoretical benefits seem appealing, they rarely play out in reality due to the inconsistencies of their programming. Moreover, systems are only as good as the people who build them. A reformed criminal justice system holds more hope than that of today. Coding our current politicisation into the future would only deny the United States a chance to improve.
Bibliography:
Berens, M. and Shiffman, J. (2020). ‘Thousands of U.S. judges who broke laws or oaths remained on the bench’. Reuters. Available at: https://www.reuters.com/investigates/special-report/usa-judges-misconduct/ (Accessed 28th November 2021)
Block, J.R. (1980). Stump v. Sparkman and the history of judicial immunity. Duke Law Journal, 1980(5), 879-925
Brenan, M. (2021). ‘Americans’ Confidence in Major U.S. Institutions Dips’. Gallup International. Available at: https://news.gallup.com/poll/352316/americans-confidence-major-institutions-dips.aspx (Accessed 28th November 2021)
Cohen, A. and Yang, C.S. (2019). Judicial Politics and Sentencing Decisions. American Economic Journal, 11(1), 160-191.
Durante, K.A. (2021). County-Level Context and Sentence Lengths for Black, Latinx, and White Individuals Sentenced to Prison: A Multi-Level Assessment. Criminal Justice Policy Review, 32(9), 915-937.
Forrest, K.B. (2021) When Machines Can Be Judge, Jury, and Executioner. Singapore, World Scientific.
National Archives, (2021) The Constitution: Amendments 11-27. National Archives. Available at: https://www.archives.gov/founding-docs/amendments-11-27 (Accessed 28th November 2021)
State v. Loomis (2016). Supreme Court of Wisconsin. 881 N.W.2d 749
Tully, T. (2021). ‘Judges Juggle Over 2,700 Cases Each as Families Wait for Day in Court’. The New York Times. Available at: https://www.nytimes.com/2021/03/17/nyregion/federal-court-nj-judges.html (Accessed 28th November 2021)