AI in Criminal Sentencing: Constitutional Challenges to Predictive Risk Assessment Tools
Introduction
The judge who sentences a convicted person exercises one of the most consequential forms of state power. The sentence determines not merely the duration of incarceration but the trajectory of a human life, the separation of families, the foreclosure of opportunities, and in capital cases, the ending of life altogether. It is a decision that our constitutional order surrounds with elaborate procedural safeguards, precisely because its stakes are so high and its potential for arbitrariness so significant.
Predictive risk assessment tools, which purport to calculate the likelihood that an offender will reoffend, have been adopted by courts and correctional authorities in several jurisdictions as inputs into sentencing and parole decisions. These tools, built on historical criminal justice data, claim to provide an objective, evidence-based supplement to judicial intuition. Their critics argue that they encode racial and socioeconomic bias from historical data, that they penalise individuals for group characteristics rather than individual conduct, and that their opacity violates the right to a fair hearing.
In India, while the formal adoption of predictive risk assessment in sentencing has not occurred at scale, algorithmic tools are being piloted in police resource allocation, bail risk assessment in some district courts, and parole board processes in certain states. Understanding the constitutional challenges to these tools is urgently necessary before their deployment becomes entrenched.
Legal Framework
Article 14 guarantees equality before law and equal protection of the laws, prohibiting arbitrary and discriminatory state action. Article 20 has been interpreted to embody a broader principle of individual accountability: punishment must be for what an individual did, not for who they are or who statistical models suggest they may become.
Article 21 protects life and personal liberty against deprivation except according to procedure established by law, and the Supreme Court’s expansive interpretation in Maneka Gandhi v Union of India requires that such procedure be fair, just, and reasonable. A sentencing procedure that incorporates an opaque algorithmic assessment of future risk, without permitting the offender to understand, challenge, or rebut the assessment, falls short of the Maneka Gandhi standard on its face.
Judicial Developments
The Supreme Court’s 2017 decision in Puttaswamy (Privacy-9J) recognised informational privacy and decisional autonomy as constitutional rights and specifically warned against profiling that assigns individuals to categories based on group characteristics. In the context of bail, the Supreme Court in Satender Kumar Antil v CBI (2022) directed that bail decisions should be based on individual assessment rather than categorical assumptions about the accused.
The Wisconsin Supreme Court’s 2016 decision in State v Loomis held that reliance on COMPAS risk scores in sentencing did not violate due process so long as the score was not the determinative factor. Subsequent empirical research by ProPublica demonstrating that COMPAS falsely flagged Black defendants as high risk at nearly twice the rate of white defendants has substantially undermined its doctrinal authority. Germany’s Constitutional Court, in a 2025 preliminary injunction, held that a deprivation of liberty premised on future risk prediction must satisfy the strictest standard of justification.
Contemporary Issues and Analysis
The constitutional challenges cluster around three distinct arguments. The first, equal protection, targets the tool’s differential impact on constitutionally protected groups. If a risk assessment tool trained on historical criminal justice data produces scores that systematically rate members of historically over-policed communities as higher risk, and if those communities map to constitutionally protected characteristics such as caste, religion, or ethnicity in the Indian context, the tool may perpetuate unconstitutional discrimination through a laundered algorithmic mechanism.
The second argument, due process or natural justice, targets the tool’s opacity. An offender sentenced on the basis of an algorithmic risk score has no meaningful opportunity to contest the score if the algorithm’s methodology is proprietary, if the inputs are not disclosed, or if the reasoning that produced the score is not explicable. The audi alteram partem principle requires that an individual be given a meaningful opportunity to be heard before adverse state action is taken. This is not satisfied by access to a black-box output.
The third argument, individual accountability, is philosophically the most fundamental. Criminal punishment is justified by individual culpability for past conduct. Predictive risk assessment, by contrast, is about future probability. Punishing someone more harshly because a statistical model predicts they will commit future crimes conflates punishment for past offence with preventive detention based on predicted future danger.
Comparative and International Perspective
The European AI Act explicitly classifies AI systems used in criminal justice, including those used for risk assessment in sentencing, parole, and bail, as high-risk AI systems under Annex III, subjecting them to strict conformity requirements including bias testing, explainability obligations, and human oversight mandates.
The United States’ President’s Council of Advisors on Science and Technology 2024 report concluded that current predictive risk assessment tools do not meet the scientific standards required for reliable use in individual sentencing decisions, and recommended a moratorium on their use pending the development of validated, bias-tested, and explainable alternatives.
Practical and Policy Implications
For state governments and correctional departments in India, the adoption of algorithmic tools in any component of the criminal justice process creates constitutional exposure that existing procurement and deployment frameworks do not adequately address. A risk assessment tool purchased from a vendor without algorithmic audit, bias testing, or explainability documentation is a constitutional liability in waiting.
Suggestions and Reforms
India should enact a Judicial Decision Support Systems Act that specifically regulates the use of algorithmic tools in all criminal justice decision-making contexts. The Act should prohibit the use of any AI system as a determinative factor in sentencing, bail, or parole decisions, and should require that any advisory use of such systems be accompanied by full disclosure to the affected individual of the system’s methodology, the inputs used in their specific case, and the limitation and error rates of the tool.
The National Law University’s Centre for Criminology, in conjunction with IIIT Hyderabad, should be mandated to conduct an independent audit of any risk assessment tool before its deployment in Indian criminal justice, assessing bias, accuracy, explainability, and constitutional compliance.
Conclusion
The criminal justice system is where the state’s power over individuals is most concentrated and most consequential. Introducing AI decision-making into that system without rigorous constitutional analysis, independent validation, and transparent accountability is not a modernisation of justice but a delegation of judicial responsibility to a machine whose biases, errors, and reasoning cannot be scrutinised. Before Indian courts or correctional authorities adopt predictive risk assessment at scale, the constitutional case for rigorous regulation must be taken seriously at the institutional level, not merely at the academic one.
