Consider a police force that deploys an artificial intelligence system to generate a predictive "heat map" of where crime is expected, and an officer who relies on it to form the reasonable suspicion required for a stop, search or arrest. The algorithm belongs to a private vendor and its methodology is a trade secret. Can the arrested person be told the grounds of arrest, as Article 22(1) of the Constitution requires, when the operative ground is a model output nobody will explain? And can the accused examine and challenge that output at trial, as the Article 21 fair trial guarantee requires? Indian courts have not answered either directly, but the framework that would decide them is in place, and on the authorities reviewed it points consistently in one direction.
The Grounds of Arrest Must Exist and Must Be Communicable
Arrest without warrant is the exception in Indian criminal procedure, not the rule, and must rest on specific grounds capable of being articulated. Section 35 of the Bharatiya Nagarik Suraksha Sanhita 2023 (BNSS), replacing Section 41 CrPC 1973, permits warrantless arrest in cognizable offences. Section 47 BNSS then imposes the safeguard: any person arrested shall be informed of the grounds of arrest and of the right to bail, "as soon as may be" after the arrest. The BNSS also requires the investigating officer to record in the case diary the reasons for a search at a particular location.
Two obligations sit inside that structure and are easy to conflate: grounds must exist and be articulable, and they must be communicable in intelligible form. An algorithmic output can arguably satisfy the first while failing the second, and it is the second Article 22(1) protects:
"No person who is arrested shall be detained in custody without being informed, as soon as may be, of the grounds for such arrest nor shall he be denied the right to consult, and to be defended by, a legal practitioner of his choice."
In Vihaan Kumar v. State of Haryana, the Supreme Court held that failure to inform an arrested person of the grounds of arrest violates Article 22(1) and renders the arrest unconstitutional: "any infringement of this fundamental right would vitiate the process of arrest and remand". Grounds must be communicated in writing, and the requirement is mandatory rather than a formality.
That holding is the sharpest edge of the problem. A heat map does not produce grounds in the sense Article 22(1) contemplates; it produces a probability attached to a place. If the officer's reason for the arrest is that the model flagged a location, and the model's reasoning is proprietary, the written communication handed to the arrested person either omits the operative ground or restates it in terms conveying nothing. Neither discharges the duty.
Article 21: The Right to Examine the Case Against You
Article 21 guarantees life and personal liberty, and the Supreme Court has read it to encompass the right to a fair trial. In Maneka Gandhi v. Union of India, (1978) 1 SCC 248, the Court held that the procedure established by law must be "fair, just, and reasonable, and it cannot be arbitrary, oppressive, or unreasonable". In Zahira Habibulla H. Sheikh v. State of Gujarat, (2004) 4 SCC 158, it defined fair trial as "a trial before an impartial Judge, a fair prosecutor and atmosphere of judicial calm", one "in which bias or prejudice does not operate".
The content of the right includes the right to be heard and present evidence in one's defence, to examine and cross-examine evidence, to know the case against oneself, and protection against arbitrary state action. Each is engaged here. An accused cannot cross-examine a system whose logic, training data, error rate and biases are withheld as commercial confidences: there is no witness to test, and the document standing in for one cannot be interrogated.
The Disclosure Cases: Grounds Presuppose Explicability
Indian courts have long required the prosecution to disclose the basis of its actions. In Pranab Chatterjee v. State of Bihar, the Supreme Court held that Section 50 CrPC, now Section 47 BNSS, is mandatory: "[i]f particulars of offence are not communicated to an arrested person, his arrest and detention are illegal", though "[t]he grounds can be communicated orally or even impliedly by conduct".
The apparent generosity of that formulation is deceptive. Oral or implied communication is permitted precisely because grounds are assumed to be the kind of thing one person can convey and another grasp. The principle presupposes that grounds exist and can be communicated, and a heat map that cannot be explained fails at the presupposition, not the mode of delivery.
In Joginder Kumar v. State of Uttar Pradesh, (1994) 4 SCC 367, the Supreme Court placed the burden on the officer to satisfy the court that he had reasonable grounds of suspicion before a warrantless arrest, emphasising that the police must be "quick to see the possibility of crime, but they ought to be anxious to avoid mistaking the innocent for the guilty". A model output, standing alone, cannot discharge that burden: the officer cannot explain what the reasonable ground was, only report what the machine said.
The principle of natural justice, audi alteram partem, is embedded in Indian criminal procedure and is not in doubt, though it is commonly grounded in Daryao v. State of UP, AIR 1961 SC 574, on a secondary source rather than the report itself, so that attribution needs checking.
The Evidence Act Was Not Built for This
The Indian Evidence Act 1872, and its successor the Bharatiya Sakshya Adhiniyam 2023, contain no provisions specific to AI-generated evidence. Section 65B requires that electronic records be accompanied by a certificate attesting to their integrity and provenance. In Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473, the Supreme Court held that electronic evidence must adhere to Section 65B certification, and in Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 1 SCC 1, it reaffirmed that digital evidence is inadmissible without it.
Both decisions address electronic evidence generally. Section 65B was drafted for conventional digital formats, emails, files and photographs, and presumes human authorship: the question it answers is whether the record is what it purports to be. With an algorithmic output that authenticity question is the easy one; the hard question is the reasoning that produced the output, which is frequently opaque, and Section 65B does not ask it.
A 2025 analysis in the Indian Journal of Integrated Research in Law makes the practical point: where the algorithm is proprietary and controlled by a private corporation, the Section 65B certificate becomes impractical, because the methodology cannot be disclosed. A certificate that cannot attest to reliability, validation or known biases certifies the container, not the contents. Its conclusion is that "[t]he present framework is inadequate" and that "[l]aw enforcement agencies are adopting AI tools without any statutory oversight".
What Indian Courts Have Said About Algorithmic Policing So Far
No Supreme Court ruling exists on proprietary algorithmic heat maps used for arrest; what exists is adjacent material pointing the same way. A 2022 paper in the International Review of Information Ethics, "Building the case for restricted use of predictive policing tools in India", documents the use of predictive policing across Indian states and identifies the structural defect:
"Inferior quality datasets are being used to train algorithms and citizens are unable to contest inaccurate algorithmic outcomes that could lead to their preventive detentions, often to the detriment of criminal justice norms and constitutional fundamental rights of citizens."
The paper identifies the risk of preventive detention under the CrPC's Section 151 power on the basis of algorithmic predictions. (That power is elsewhere mapped to a BNSS provision, but the mapping does not appear reliable and is not reproduced.) On accuracy, it is reported that the Delhi High Court, dealing with facial recognition technology used by the Delhi Police, noted the software had only 2% accuracy and directed an upgrade. That figure comes through a secondary source and should be verified against the order, but the posture it reflects, scepticism toward opaque tools deployed against individuals, is the signal.
Trade Secrecy Is Not a Privilege Against Disclosure
Indian law does not recognise a blanket trade secret privilege permitting evidence to be withheld in criminal proceedings. India has no dedicated trade secret statute; prosecutors proceed under general criminal law on theft, dishonest misappropriation, criminal breach of trust and cheating, and under the Information Technology Act 2000, none of which creates a privilege. The principle has old roots in English common law adopted by Indian courts: when the East India Company refused to produce records at a public trial because they contained secrets, the court held that company papers containing material evidence must be produced, because "[h]umanity requires it should be produced, when in favour of a criminal, justice when against him".
The consequence is direct. A law enforcement agency cannot invoke a vendor's commercial interest to withhold a heat map's methodology from the accused or the court: the vendor's contract is not a source of privilege, and an agency cannot acquire by procurement an immunity the law does not grant.
Algorithmic Accountability Under the DPDP Act
The Digital Personal Data Protection Act 2023 came into effect on 13 November 2025 with the notification of the DPDP Rules 2025. The Act defines "processing" to include any "wholly or partly automated operation or set of operations performed on digital personal data", and the Rules impose an expectation of algorithmic accountability on significant data fiduciaries, a category that would capture law enforcement agencies processing personal data at scale:
"Significant data fiduciaries must observe due diligence to verify that the algorithmic software deployed by it for hosting, displaying, uploading... or sharing of personal data processed by it are not likely to pose a risk to the rights of Data Principals."
The Act applies uniform obligations across all personal data rather than carving out a "sensitive" category. The duty is one of verification, and verification is what a black box defeats: an agency that cannot inspect its vendor's model cannot observe the due diligence the Rules require. The Information Technology Act 2000 adds nothing here, penalising breach of confidentiality under Section 72 and inadequate security practices under Section 43A, but containing no transparency provisions.
As of July 2026 India has no dedicated legislation on AI in courts or law enforcement. Two developments mark official thinking. The Kerala High Court's AI policy for the District Judiciary (July 2025), the first of its kind in India, prohibits AI as a substitute for decision-making or legal reasoning, permitting approved tools only as assistive aids under human supervision; and the Supreme Court's White Paper on Artificial Intelligence and Judiciary (November 2025) would confine AI to administrative and assistive tasks rather than adjudication. Both draw the same line: assistance to a human decision-maker, yes; substitution for the decision, no. Transposed to policing, that is the line between a heat map directing patrol attention and one supplying the ground for an arrest.
Comparative Signals
Indian courts consider comparative jurisprudence when interpreting constitutional rights, as the Supreme Court did in Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1, recognising privacy as a fundamental right under Article 21. In State v. Loomis (2016), the Wisconsin Supreme Court rejected a challenge to COMPAS, a proprietary sentencing tool that had rated the defendant a "high risk of recidivism", reasoning in part that no due process violation arose because the judge's own access to the secrets was equally limited. That reasoning has been heavily criticised and, on the authorities reviewed, is not representative of emerging United States jurisprudence, where the Daubert standard requires validation studies, reliability testing and scrutiny of known biases. In Bridges v. South Wales Police (2020), the UK Court of Appeal held police use of facial recognition unlawful for want of a Data Protection Impact Assessment and adequate transparency. And the EU AI Act 2024, with the GDPR, classifies law enforcement systems as "high risk", subject to testing, validation and disclosure, illustrating that certification can give litigants and courts material information about a tool even where it stays proprietary. The choice is not binary between source code disclosure and total opacity.
Where This Leaves the Hypothetical
A proprietary heat map used as the basis for reasonable suspicion is exposed on five fronts. It cannot supply grounds communicable under Article 22(1), and on Vihaan Kumar that alone vitiates the arrest and remand. It cannot be examined or cross-examined, which offends the Article 21 fair trial right as developed from Maneka Gandhi to Zahira Habibulla. It cannot be certified in any meaningful sense under Section 65B, now Section 63 of the Bharatiya Sakshya Adhiniyam 2023, because the certificate cannot attest to reliability, validation or bias. It cannot be verified as the DPDP Rules require. And it cannot discharge the Joginder Kumar burden, a black box being no communicable ground.
Further consequences are projected if an arrest is invalidated: exclusion of the evidence obtained through it, a claim in damages for wrongful arrest, and DPDP penalties of up to INR 250 crores for failure to implement reasonable security safeguards. These are projections, not holdings: no Indian court has excluded evidence on this basis or awarded damages on these facts, and the exclusionary consequence is asserted without authority. Two asymmetries will shape any litigation. Challenging algorithmic evidence needs expert testimony not available to every accused; and a prosecution with access to the model facing a defence without it is an inequality of arms that is itself a fair trial problem, whatever the model's accuracy.
Four questions remain open: whether a heat map can play any role at all in establishing reasonable suspicion; what disclosure is required; whether trade secret protection can ever operate in criminal proceedings; and whether the reliability standard should be Daubert imported or built for Indian conditions. All of this is extrapolation rather than settled law. The DPDP Act and Rules are recent and untested against law enforcement, reported Indian cases are few, and comparative jurisprudence does not bind.
Practical Takeaways
- An algorithm cannot be the sole or primary ground for an arrest, stop or search. It may be one factor among several, but only where independent, communicable grounds exist and are recorded, and the Article 22(1) communication must stand on those alone.
- Agencies should build the record before deployment, not after a challenge: validation studies, accuracy testing, documented biases, methodology available to courts and defence counsel under appropriate confidentiality protections rather than blanket trade secret claims, and DPDP compliance. Procurement is no shield, and an agency that cannot inspect its own tool cannot perform the DPDP verification duty.
- For the defence, the first attack is on the grounds, not the technology. If the operative reason for the arrest was an unexplainable model output, the arrest is challengeable on Vihaan Kumar without litigating the model's accuracy, and the Joginder Kumar burden sits on the officer.
- Courts have a gatekeeping role available to them, Daubert-like in substance: validation, reliability testing and disclosure of known biases before admission, distinguishing algorithmic assistance to human judgment from replacement of it.
- The legislative gap is the real problem. The Bharatiya Sakshya Adhiniyam says nothing on AI-generated evidence, no statute defines or requires validation of high-risk AI systems, and none clarifies that trade secrecy does not travel into criminal proceedings.
Key Authorities
- Vihaan Kumar v. State of Haryana — failure to communicate grounds of arrest violates Article 22(1) and vitiates the arrest and remand; grounds must be in writing. (Reported as 2024 without a citation; date not reproduced.) Source
- Maneka Gandhi v. Union of India, (1978) 1 SCC 248 — procedure established by law must be fair, just and reasonable. Source
- Zahira Habibulla H. Sheikh v. State of Gujarat, (2004) 4 SCC 158 — definition of fair trial. Source
- Joginder Kumar v. State of Uttar Pradesh, (1994) 4 SCC 367 — burden on the officer to satisfy the court that reasonable grounds of suspicion existed. Source
- Pranab Chatterjee v. State of Bihar — Section 50 CrPC, now Section 47 BNSS, is mandatory. (Reported without a citation.)
- Anvar P.V. v. P.K. Basheer, (2014) 10 SCC 473, and Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal, (2020) 1 SCC 1 — electronic evidence inadmissible without Section 65B certification.
- Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1 — privacy under Article 21; the Court's use of comparative jurisprudence. Source
- Constitution of India, Articles 21 and 22(1) — fair trial; grounds of arrest and right to counsel. Source
- Bharatiya Nagarik Suraksha Sanhita 2023, Sections 35 and 47 — warrantless arrest; duty to communicate grounds of arrest and the right to bail. Source
- Indian Evidence Act 1872, Section 65B, now Section 63 of the Bharatiya Sakshya Adhiniyam 2023 — certification of electronic records; presumes human authorship.
- Digital Personal Data Protection Act 2023 and Rules 2025 (in effect 13 November 2025) — algorithmic due diligence for significant data fiduciaries. Information Technology Act 2000, Sections 72 and 43A — confidentiality and security only; no transparency provisions.
- State v. Loomis (2016), Wisconsin — challenge to the proprietary COMPAS tool rejected; heavily criticised. Bridges v. South Wales Police (2020), UK Court of Appeal — police facial recognition unlawful for want of a Data Protection Impact Assessment and transparency. EU AI Act 2024 and GDPR — law enforcement AI is high risk, subject to testing, validation and disclosure. (Loomis and the East India Company principle via "Life, Liberty, and Trade Secrets", Stanford Law Review, Vol. 70, 2018. Source)
- Kerala High Court AI policy for the District Judiciary (July 2025) and Supreme Court White Paper on Artificial Intelligence and Judiciary (November 2025) — AI as assistive tool only.
This analysis reflects the law as at July 2026. It is published for general information and does not constitute legal advice.