Use of Artificial Intelligence in the Courtroom: An Exploration of Digital Justice

This Blog is Written by Harshit Singh, 3rd Year, B.A. LL.B. (Hons.), Faculty of Law, Banaras Hindu University (BHU), Varanasi.

Blog 09 | Edition VII

Introduction

Artificial intelligence systems have evolved from being peripheral elements of court administration into indispensable systems. Systems that have the capability to predict the results of cases, evaluate flight risks of accused persons, summarize the file of the case, and generate the draft of the order are common nowadays. The SUPACE project of the Supreme Court, AI assisted legal research platforms, and the use of AI in transcription and translation of trials in trial courts in India show the change has already occurred. This needs much deeper consideration than has been devoted to the topic so far. Efficiency is just one factor in a case where the machine generated score may determine whether an individual is granted bail, the length of his sentence, and which case will get a priority among others. The right to fair trial, equality before the law, and the constitutional principle of the exercise of judicial power only by the independent mind make place for statistical predictors rather narrow.

Where the Technology Already Exists

Predictive analytics systems already influence the process of litigation planning, using thousands of judgments to predict the decision of a given judge or bench. Much more worrying is the risk assessment -- the software predicts whether the accused person is likely to recommit the offense or skip the trial, and it influences the decisions on bail. The use of the COMPAS tool in the State v. Loomis case in front of the Wisconsin Supreme Court showed what could go wrong when the risk score entered the courtroom while the accused person was unable to contest the reasoning behind it. The tools for document review and legal research are more solid ground. Technology assisted review has been outperforming manual review in the process of discovery for more than a decade now, and modern case-law search tools are able to identify legal concepts instead of keywords because of keeping the decision-making power at the discretion of the lawyer/judge. Contested ground lies in decision-making itself, and even more so in the replacement thereof. An AI system for small claims below 7,000 euros has been introduced in Estonia; however, the final decision-making authority rests with a human judge. Neither Indian nor other common law court allows a machine to produce an authoritative judgement, yet such tools as bail and sentencing ones already influence judicial decisions.

Opacity and the Bias it Hides

The key complaint in this regard is opacity. Predictive algorithms used for risk assessment are proprietary and statistically opaque. This creates a conflict with a defendant's right to know and contest the reasons of his/her unfavourable decision. Bias aggravates the problem of transparency. A predictive algorithm learns from historic data, and the historic crime data carries with it the imprint of decades of biased policing and sentencing, which it is able to reproduce under a veil of neutral statistics. Proving discriminatory intent is becoming almost impossible since there is no individual bias but the diffusion of biases in the training data set. The independence of judiciary is compromised as well. A judge who consistently defers to the algorithmic output is no longer exercising independent judgement but merely endorsing someone else's decision, and more often than not -- of a private corporation that seeks profit rather than performs a public duty.

Artificial Intelligence and Access to Justice

There is a strong case of access to justice too. AI-powered case management is able to cut down the time between filing and hearings, and automatic translation removes language barriers. All of it is true, yet once an efficient system ceases to care about fairness, it becomes two-tiered to the benefit of the well-off only. This is not to say that AI need never have a role in courtrooms, which are already overstretched. The argument is rather that guardrails be established prior to reliance, with emphasis on those most directly impacted by such use: undertrials unable to rebut the score, and litigants facing artificially prolonged delays.

Four Guardrails

Four guardrails would accomplish much. First, all output from AI which significantly informs a decision of the court must contain an easily comprehensible explanation of the key considerations, which a party could challenge. Second, there must be regular independent audits of courtroom technology whose results are publicly disclosed, thus allowing correction or withdrawal of a discriminatory algorithm. Third, decision-making must remain within the remit of the judge; maintaining records of cases where judges rejected an AI recommendation would reinforce this principle. Finally, procurement must be considered; reliance on proprietary algorithms means losing control of the process to private vendors. There must be a higher bar of transparency in the criminal process. In light of what is at stake when the freedom of an individual is involved, any algorithmic output which does not provide an adequate explanation cannot be admitted into evidence and defence counsel must be able to gain access to such output.

Purpose of Courts

On an even more fundamental level, this is a discussion of the purpose of courts. Judicial decision-making must be public reason-giving: the judge explains the reasons behind his/her decision, and the reason-giving becomes part of development of the law. Statistical predictions, no matter how precise, do not reason; they merely correlate. Reason-giving is replaced with prediction; the process becomes faster, but it is not more just. Efficiency which is achieved at the cost of accountability is not the type which the justice system can afford. Using AI with proper explanation, audit, and human decision-making in place, courts can accomplish what they were established to do.

(Write to the author at harshitsingh97727@gmail.com.)