India’s criminal justice system is under pressure and goodwill has not been much of a balm. There are more than 5 crore cases pending in courts at all levels. According to the latest prison statistics, about three out of every four persons in Indian prisons are not convicted of any crime. There are no investigators in police stations, no examiners in forensic laboratories and legal aid, promised under article 39A of the Constitution, comes too late to too few. The State has been increasingly enthusiastic about artificial intelligence here. Facial recognition for policing, crime-mapping software, machine translation of judgments, transcription tools, and research assistants for judges are no longer laboratory curiosities. These are housed within the Crime and Criminal Tracking Network and Systems (CCTNS), the eCourts Mission Mode Project and the pilot projects of a few High Courts and have largely come in without a statute to govern them. The paper poses a question that enthusiasm tends to overlook. What is the constitutional basis for use of such technologies in the criminal process in India? It begins with the assumption that the question is not simply technical or administrative. A criminal trial is one place where the State can take away a person's liberty by a procedure. And that procedure itself must be right, just and fair. Maneka Gandhi Any tool that affects who is investigated, who is arrested, who is released and whose evidence is believed thus crosses into constitutional territory – regardless of the intentions of its designers. The study is doctrinal and comparative in nature. It measures the Bharatiya Nagarik Suraksha Sanhita 2023, the Bharatiya Sakshya Adhiniyam 2023 and the Digital Personal Data Protection Act 2023 against the line of Supreme Court authority from Maneka Gandhi through Selvi, Puttaswamy and Anuradha Bhasin, and it tests the Indian position against experience elsewhere: the Wisconsin decision in Loomis, the English Court of Appeal’s ruling in Bridges on facial recognition, the Ethical Charter of the Council of Europe’s CEPEJ, and the European Union’s Artificial Intelligence Act. It also draws on empirical work on algorithmic bias and automation bias, because an argument about legal safeguards is only as strong as its account of how these systems actually fail. The first finding is that the opportunities are real and should not be dismissed in the name of caution. Technology that can triage forensic backlogs, flag undertrial prisoners who have served the period which section 479 of the BNSS deems sufficient for release, translate judgments into regional languages, or help an overstretched legal aid lawyer organize a file, serves values that the Constitution itself protects: speedy trial, equal access to justice and reasoned adjudication. International studies of bail prediction indicate that well-constructed tools can in principle reduce unnecessary detention and avoidable offending. For an Indian prisoner waiting years for a hearing, there is little solace in theoretical objections to a machine that could have advanced his case. The second is that the risks are of a particular kind and existing law does a bad job of handling them. Opaque proprietary models can’t be interrogated. Training data based on Indian policing and prison records will reflect the overrepresentation of some castes, communities and economic groups in the system and a model trained on that data may entrench that as a prediction. Officials and judges that receive a score tend to defer to it. Facial recognition error rates are unevenly distributed, and its deployment in public spaces without a statutory footing sits uneasily with the requirements of legality, legitimate aim and proportionality laid down in Puttaswamy. Meanwhile, the Digital Personal Data Protection Act excludes processing by State agencies for the investigation of offenses in broad enough terms to put much of this activity beyond its protections. Third, the paper asserts that the constitutional framework is sufficient in principle but lacking in practice. Article 14 prohibits manifest arbitrariness and therefore provides a basis for challenging an unexplained outcome of an algorithm. Article 21 demands a fair procedure, which on any reasonable reading, demands disclosure of the material on which liberty is curtailed and a real opportunity to contest it. Article 20(3) and the reasoning in Selvi circumscribe the technologies which elicit mental or physiological evidence from an accused. There is no lack of principle, but there is a lack of legislation, procurement discipline and judicial guidance. Currently deployment is largely through tenders, executive notifications and administrative circulars, none of which meets the need of a restriction on a fundamental right being authorized by law. The paper ends by proposing a graduated framework, rather than a blanket ban or open licence. Uses that support administration, such as transcription, translation and scheduling, would require only transparency and periodic audit. Uses that inform decisions about liberty, such as risk scoring in bail or flagging suspects, would require statutory authority, disclosure to the accused, independent validation on Indian data and a duty on the decision-maker to give reasons beyond simply repeating the output. Prohibited are uses that supplant a machine’s adjudicative judgment of guilt, sentence, or credibility, and uses that infer mental states from physical signals. Judicial adoption policies such as the one issued by the Kerala High Court in 2025 are a useful starting point, but a policy addressed to judges cannot discipline the police or private vendors. The gap has to be closed by Parliament or failing that by guidelines of the Supreme Court in the manner of Vishaka and D K Basu. At the center is a modest but firm claim. Artificial intelligence can assist India’s criminal justice administration without violating its constitutional commitments, but only if the assistance is answerable to human beings who can be questioned, reversed and, when necessary, blamed. Efficiency is a proper constitutional issue. It was never meant to be bought fairly.
