Before Prediction Becomes Punishment: Governing AI in Bangladesh's Criminal Justice System

Imagine a world where a computer program helps decide who gets arrested, who stays in jail before their trial, and how long someone serves a sentence. This is not science fiction; it is the emerging reality of criminal justice in many places, including Bangladesh. The core idea behind these systems is "predictive policing" or "risk scoring." Instead of a police officer or judge relying solely on their memory or a single file, they use software that scans vast amounts of data—past arrests, where people live, their age, or their family history—to guess who might commit a crime in the future or who is likely to run away if released. The promise is efficiency: catching bad actors faster and treating everyone the same. But the danger is equally real. If the computer learns from past mistakes, like a police force that only patrolled poor neighborhoods, it will simply tell the police to patrol those same neighborhoods again, creating a loop where the poor get punished more often, not because they are worse, but because they were watched more closely. The question facing nations today is not whether to use these tools, but how to use them without turning historical unfairness into automated punishment.

A researcher named Shanjida Israt Jahan Efat, writing from Notre Dame University Bangladesh, has examined this exact moment for her country. Bangladesh is currently building the digital foundation for such a system. The government has introduced portable devices that can scan a person's fingerprint and instantly pull up their identity, passport, and criminal history. New laws now allow digital records to be used as evidence in court, and a draft national policy is being prepared to guide how artificial intelligence should be used from 2026 to 2030. However, the country does not yet have clear rules on when a computer's guess is strong enough to justify taking away a person's freedom. Efat's work is a roadmap designed to prevent the country from making a critical error: adopting technology that looks modern but actually deepens old inequalities.

The paper begins by mapping out exactly where these computer systems could enter the justice process, from the moment police plan their patrols to the day a prisoner is released. The researcher identifies seven key stages where things could go wrong. For instance, when police use software to predict where crimes might happen, the system might simply send more officers to areas that are already heavily policed. This creates a cycle where more officers find more minor offenses, which the computer then uses as proof that the area is dangerous, justifying even more officers. This is not a neutral observation of crime; it is a self-fulfilling prophecy. Similarly, when software is used to identify suspects through facial recognition or fingerprints, a mistake in the system can lead to a false arrest. The paper notes that these systems often struggle with accuracy for specific groups, such as women with darker skin or people wearing religious clothing, leading to a higher chance of error for those who are already vulnerable.

Efat argues that the current laws in Bangladesh are not enough to stop these problems. While there are new rules about protecting personal data, those rules have a large loophole that allows the police to bypass privacy protections when they claim they are investigating a crime. The law protects the data itself, but it does not explain how a judge should treat a computer's prediction when deciding whether to keep someone in jail. The researcher points out that simply having a human "in the loop"—meaning a human looks at the computer's answer before making a decision—is not enough. If a judge or police officer feels pressured to trust the machine, or if they do not understand how the machine reached its conclusion, their presence does not fix the problem. A human who just signs off on a score without understanding its limits is not providing real oversight; they are just a rubber stamp.

To solve this, the paper proposes a strict five-step test, or "five gates," that any new system must pass before it can be used to make decisions about people's liberty. The first gate asks for a clear law. The government must have a specific written rule that says exactly what the computer is allowed to do, who can use it, and what data it can see. A general desire to "fight crime" is not enough to justify mass surveillance. The second gate demands proof of necessity. The authorities must show that they have a real problem that cannot be solved by simpler, less intrusive methods, and that the benefits of the new system are worth the risks.

The third gate is perhaps the most critical: local validity. A system that works well in one country, or even in one city, might fail completely in another. The paper insists that any tool used in Bangladesh must be tested on real Bangladeshi people, in local conditions, with local languages and local data. You cannot simply buy a system from abroad and assume it will work. The fourth gate focuses on fairness. The system must be checked to see if it makes more mistakes for certain groups of people, such as the poor, women, or ethnic minorities. If the computer is more likely to falsely accuse a specific group, it cannot be used. The final gate is about the right to challenge. If a computer's output is used to arrest or sentence someone, that person must be told how the decision was made, given a chance to see the data, and allowed to ask a human to review the result. If the system is a secret "black box" that no one can question, it fails this gate.

The researcher is clear about what should not happen. The paper argues against using computers to predict a person's future behavior based solely on their personality or background. It also argues against using these tools to make the final decision on arrest, detention, or sentencing without a strong, independent human review. The paper does not claim that artificial intelligence has no place in justice. In fact, it suggests that these tools could be very helpful for non-coercive tasks, such as finding missing persons, detecting fraud, or organizing huge piles of evidence to find documents that prove someone is innocent. The goal is to use technology to help the system work better, not to let the technology take over the judgment of right and wrong.

The paper concludes with a practical plan for Bangladesh to move forward. It suggests that the government should first take an inventory of all the systems they already have or are planning to buy. They should freeze any use of technology that violates the five gates. Then, they should create a public list of all algorithms used by the police and courts, so citizens know what is being used. Finally, they need to set up an independent group of experts—including judges, lawyers, and community representatives—to oversee these tools and stop them if they start causing harm. The message is one of cautious optimism. Bangladesh has the chance to build a justice system that is both modern and fair, but only if it puts clear rules in place before the technology becomes too entrenched to change. The future of criminal justice should not be a choice between inefficient institutions and automated injustice; there is a third path where technology serves the law, rather than replacing it.