Introduction
The choice of whether to detain, sentence, or set free a defendant has long been regarded as one of the most serious acts of governmental authority against the individual, and thus the one most fortified by procedural protections. Notice, a reasoned decision, a hearing, and an appeal against the decision on its merits have always constituted the heart of due process in common law systems. However, during the last twenty years, courts, parole boards, and police agencies have increasingly turned over some of this reasoning to statistical and machine learning algorithms that rate defendants according to their risk of committing new crimes, absconding, or failing to comply.
This particular instance of delegation differs from previous incarnations of actuarial reasoning in criminal justice, whether through sentencing guidelines or parole matrices, along three main lines, which give sufficient reason to regard it as a novel object for due process scrutiny. To begin with, these models often belong to commercial entities which license them on a proprietary basis and, by virtue of trade secret protection, make it impossible for the accused, and sometimes for the court itself, to learn how the algorithm transformed each input into a particular score. In addition, these models were built on historical criminal justice records, which, rather than reflecting neutral observations of crime, reflect earlier enforcement decisions and thus encode any bias present at those stages of criminal justice processing. Lastly, and perhaps most importantly for due process analysis, these models produce a score that does not come with the explanation one would expect a human decision-maker to provide, so that even a transparent algorithm would fall short of the reasoned decision standard.
The thesis of this paper is that the delegation of bail, sentencing, and parole decisions to algorithmic software causes a due process crisis rather than a mere doctrinal difficulty. The crisis has two faces. In jurisdictions where the use of risk-score algorithms has been litigated, like the United States, the due process challenge yielded a meagre and mostly rhetorical due process right: the courts allowed the practice to continue so long as procedural caveats were issued, but left the lack of transparency, and therefore of reviewability, untouched. In jurisdictions where such tools have not yet been litigated, of which India is an important example, the absence of litigation signals not the absence of harm but the absence of a forum, of a disclosure requirement, or of a rights-bearing litigant group.
This study therefore investigates the hypothesis that existing due process jurisprudence in the United States and India is structurally incapable of overseeing the algorithmic risk scoring involved in criminal justice processes, since that jurisprudence is premised on an articulated, human chain of reasoning that the algorithms do not provide. The subsidiary hypothesis is that judicial oversight without an accompanying legal duty to disclose the workings of the algorithm and its potential errors will not actually help the defendant, and that the absence of litigation does not mean that no damage is being done.
The paper continues through five more sections. Section II lays down the basis of due process as it exists in the United States and India respectively, and provides the backdrop against which algorithmic decision-making is analysed. Section III analyses the American model of judicial review, focusing on State v. Loomis and on the pretrial risk assessment tools used in bail determinations. Section IV shifts focus to India, where the absence of risk assessment tools at sentencing is matched by an extensive and largely unregulated use of algorithms in investigation and pretrial detention. Section V compares the two systems and asks whether judicial review, as seen in the American model, actually delivers due process better than what obtains in India. Section VI concludes.
Due process: doctrinal foundations for algorithmic scrutiny
A. The American Due Process Clause and individualised justice
Under the Fifth and Fourteenth Amendments to the United States Constitution, no person may be deprived of life, liberty, or property without due process of law.1 In sentencing, due process requires, amongst other things, that the sentence be based upon accurate information and that the defendant have an opportunity to contest the facts upon which the sentence is based. This rests on the assumption that the rationale for a sentence, even an unfavourable one, can always be stated, analysed, and contested. Actuarial risk assessment methods undermine this very presumption, since the weighting that goes into the actuarial calculation is kept secret by the software vendor who sells the service to the state.
B. Article 21 and the journey from “procedure established by law” to substantive due process
Article 21 of the Indian Constitution provides that no person shall be deprived of his life or personal liberty except according to procedure established by law.2 While the text does not contain the words ‘due process’, the Supreme Court of India in Maneka Gandhi v. Union of India held that the procedure referred to in Article 21 must be just, fair and reasonable.3 This is reinforced by the Court’s recognition of informational privacy as a component of the right to life and personal liberty, and by its proportionality test for state action that infringes that right.4 These two principles provide the constitutional grounds in India on which a challenge to the opaque use of algorithms in decision-making can be mounted, even though, as Section IV demonstrates, no party has yet done so.
The United States: litigated risk, unresolved crisis
A. COMPAS and State v. Loomis
The litigation around the use of algorithmic risk scores in the criminal justice process has centred on one risk assessment system, the Correctional Offender Management Profiling for Alternative Sanctions, known as COMPAS, made by Northpointe.5 In State v. Loomis, the State alleged that the defendant had been the driver in a drive-by shooting; he denied involvement and pleaded guilty to two of the less severe charges, attempting to flee a traffic officer and operating a motor vehicle without the owner’s consent. The sentencing court referred to a COMPAS assessment that placed him at high risk of recidivism and imposed six years of initial confinement and five years of extended supervision. The defendant contended that the use of such a closed-source algorithm violated his due process rights to be sentenced on accurate information and to an individualised sentence, and that the tool impermissibly took gender into account.6
B. The limits of the Loomis safeguards
On the due process claim, the Supreme Court of Wisconsin found no violation in the use of COMPAS scores at sentencing, so long as the score was not the determinative factor in the sentence and the presentence investigation report carried a written advisement explaining the proprietary nature of the tool and its general limitations.7 The United States Supreme Court later denied certiorari, leaving the Wisconsin decision as the leading case on the issue.8 In effect, the Loomis safeguards make known the existence of the tool and its general limitations rather than its underlying logic. The defendant is told that a proprietary score has been taken into account, but is not told, and cannot work out, how the weights were applied, what the model’s error rates are for individuals like the defendant, or whether the specific data used to compute the score are reliable.
C. The accuracy and bias problem
Independent examination of COMPAS has cast further doubt on its ability to support the individualised and accurate sentencing that due process requires. In 2016, an investigative analysis found that the algorithm was more likely to classify black defendants who did not go on to reoffend as high risk, and to classify white defendants who did reoffend as low risk, although race is not among the factors the algorithm considers.9 The maker of the software contested the statistics behind these conclusions, stating that error rate parity is the wrong measure of fairness and that COMPAS satisfies predictive parity.10 This unresolved methodological dispute itself raises a due process issue: if experts in the field cannot agree about the fairness of the tool, a defendant who is denied access to the model has no realistic possibility of contesting his score.
D. Pretrial risk-assessment tools in bail decisions
The due process arguments posed against COMPAS apply with no less force to the risk assessment instruments employed at the bail stage in American courts, which assign a numeric score reflecting the correlation between static and dynamic factors in a defendant’s profile and the probability of failure to appear for trial or of pretrial rearrest. Because the bail decision deprives a person presumed innocent of his freedom, the problem of individualisation is even more acute here than in sentencing: the deprivation takes place before any guilt has been established and rests on a probabilistic estimate rather than on a proven fact. Yet despite the presence of the same flaws as in Loomis, namely vendor-claimed confidentiality of the algorithm, the use of group rather than individual data as input, and the lack of a reasoned output, the pretrial use of such instruments has attracted less attention from courts, in part because poor defendants at the bail stage lack the time and means for prolonged appeals.
India: unlitigated risk, invisible crisis
A. Predictive policing without a statute
India has not yet implemented anything similar to COMPAS for use in sentencing, but algorithmic risk assessment has entered its criminal justice system at an earlier and, to a certain degree, more consequential stage: police work and investigation before trial. Police forces have adopted predictive policing software that maps crime hotspots from past crime records, such as the Delhi Police’s Crime Mapping, Analytics and Predictive System (CMAPS), alongside other algorithmic tools such as Telangana’s real-time biometric matching application.11 Because such systems are built on historical data of arrests and complaints, and because Indian policing has long been criticised for the disproportionate surveillance of particular communities and localities, commentators have warned that these systems risk reproducing and reinforcing the very patterns of policing from which their training data were drawn.12 In contrast to COMPAS, these programs function largely without a statute governing their use, and outside the adversarial procedure through which COMPAS was challenged in Loomis.
B. Facial recognition and pretrial detention: the Delhi riots investigations
The best-known example of the due process problems that machine learning algorithms create in the Indian criminal justice system arises not from a risk score but from facial recognition algorithms used to identify suspects. After the 2020 Northeast Delhi riots, investigators turned to facial recognition technology to identify suspects. Its reliability has been questioned: according to one policy commentary, over 80 per cent of the facial recognition cases that reached a verdict ended in acquittal or discharge, and in at least one documented case a man spent four and a half years in custody before securing bail, his detention having rested substantially on an 80 per cent similarity score generated from a CCTV frame.13 This case is critical for the due process analysis because it illustrates the same problem seen in the United States: an unreviewable algorithmic output used to make a decision that restricts liberty. Unlike in Loomis, no disclosure requirement obliged the authorities to disclose the error rate, training data, or confidence threshold of the facial recognition system.
C. Historical continuities: from “criminal tribes” to “history-sheeters”
Predictive policing and its procedural fairness dilemma in India cannot be dissociated from an older institutional practice of keeping surveillance registers for communities identified as “habitual criminals.” Communities such as the Pardhi, branded “hereditary criminals” under the colonial criminal tribes legislation that began with the Criminal Tribes Act, 1871, are reported still to be kept on record and under surveillance as “history-sheeters” despite the repeal of that legislation in 1952, and police records of this kind now feed the Crime and Criminal Tracking Network and Systems, the centralised digital records system that is promoted as enabling algorithm-based detection of crime hotspots and predictive policing.14 If predictive policing systems are trained on decades of such police data, the caste-based profiling embedded in that data will be encoded in the model as a seemingly neutral statistical correlation, so that the colonial practice of presuming criminality by birth into a community continues within the framework of risk calculation.
D. The regulatory vacuum
India’s principal data protection law is the Digital Personal Data Protection Act, 2023. Under section 17(1)(c), processing in the interest of the prevention, detection, investigation or prosecution of offences is exempted from most of the Act’s obligations, including its notice and consent requirements and the data principal’s rights; that section, like most of the Act’s substantive provisions, takes effect eighteen months after the commencement notification of November 2025.15 Consequently, analysts have described India as deploying facial recognition and algorithmic policing tools at scale without a governing framework.16 The implication for due process is that the Maneka Gandhi and Puttaswamy jurisprudence outlined above remains legally available to any litigant but has yet to be brought to bear on an algorithm used within the criminal justice system.
Comparative analysis and the case for reform
A. Does litigation lead to due process?
The comparison between the United States and India is meant to test a particular hypothesis: that a developed adversarial framework, in which the defendant has sufficient capability and standing to litigate, yields substantially better due process in algorithmic decision-making than a system in which no litigation has occurred. The American experience complicates this hypothesis. Although Loomis was a thoroughly litigated case that reached the highest state court and received extensive academic attention, the due process protection it offered, namely a written statement of the tool’s limitations, gives the defendant no means of testing the correctness of the score attributed to him.
In India, conversely, the lack of a lawsuit does not equate to the lack of injury. Four and a half years of pretrial custody resting substantially on a facial recognition match is at least as serious a due process concern as the one raised in Loomis, and probably more so, because it involved physical detention before trial rather than a somewhat harsher sentence. What distinguishes the two situations is not the seriousness of the underlying injury but the visibility of its cause. The Indian case came to light through reporting and policy research, not through a judicial opinion, which means that there is currently no legal standard that would serve to prevent future incidents or impose any duty of disclosure on the police.
B. Institutional capacity and trade secrecy as common obstacles
Another aspect of the comparison lies in the technical capability of the reviewing body itself, irrespective of the jurisdiction in which it sits. The Wisconsin Supreme Court in Loomis had no such technical capability, nor did it attempt to acquire it. Its concern was whether there had been adequate disclosure rather than whether the disclosed method was in itself reliable. Indian courts facing a similar challenge would confront an identical dilemma under the proportionality test in Puttaswamy, whereby a court must determine whether a measure is a suitable, necessary, and proportionate means of accomplishing the aim of the state. Such a determination is impossible without either compelling disclosure and expert testimony or developing sufficient technical capability, a task beyond the capacity of most of the magistrates’ courts that decide the vast majority of bail applications.
Arguments of trade secrecy arise in both jurisdictions as the legal device that preserves the opacity of algorithms even after a dispute has reached the courts. In Loomis, the Wisconsin Supreme Court accepted that the developer treats COMPAS as a proprietary instrument and a trade secret and does not disclose how the factors are weighed, and it held that due process was nonetheless satisfied because the defendant could verify the inputs and challenge the resulting scores.17 In India, predictive policing software and facial recognition technology have usually been procured from vendors, and the terms of these contracts do not provide for proactive disclosure. In the absence of a dedicated framework for algorithmic policing tools, there is currently no statutory basis to require even the kind of low-level warning imposed in Loomis. The critical point is that trade secrecy operates as a general-purpose block on the due process review of algorithms whichever due process regime governs, since the basic economic structure of the transaction is identical in either case: the state obtains a licence to use the product of the algorithm without a licence to disclose its internal logic to those subjected to it.
C. Toward reform
Reform proposals in this area therefore tend toward ex ante legislation rather than ex post constitutional challenges on the model of Loomis, on the view that a person should not have to take a case all the way to a state supreme court to obtain the disclosure that should have accompanied the use of the technology. At the very least, this would involve: first, mandatory disclosure to the individual concerned, and to the court or magistrate, of the particular inputs used to generate the risk score or identity match, as opposed to the general disclaimer accepted in Loomis; second, independent periodic audits of error rates disaggregated by race, gender and, in the Indian context, caste and community; third, a requirement of meaningful human review with power to depart from the algorithm’s results, supported by a written rationale for following or departing from the score; and fourth, in jurisdictions like India, where law-enforcement data processing is exempt from data protection requirements,18 a carve-back guaranteeing at least a right to notification and a right to challenge the decision-making process underlying any decision leading to detention.
A notable comparative point is the European response to the same type of issue, arising there in welfare administration rather than criminal justice, which suggests that a judicial remedy can be found whether or not there is a law specific to the algorithm: the District Court of The Hague held that the legislation underpinning SyRI, the Dutch welfare fraud detection system, was insufficiently transparent and verifiable and contrary to the right to respect for private life under Article 8 of the European Convention on Human Rights, and declared it to have no binding effect.19
This shows that, following the proportionality framework in Puttaswamy, Indian courts have room to reach a similar decision in the realm of criminal justice; what is needed is not legal power but an interested party, and a disclosure system strong enough to give that party a reason to litigate.20
Conclusion
This paper has asked whether due process principles in the United States and India can adequately oversee the use of algorithmic scoring in decisions relating to bail, sentencing, and parole. The evidence from both jurisdictions confirms the hypothesis that they cannot, at least not without further statutory backing. In the United States, the litigation culminating in State v. Loomis produced a disclosure that left unsolved the core due process problem of a non-reviewable basis for a liberty-affecting decision. In India, the absence of such litigation has allowed predictive policing and facial recognition technologies to be used with even less oversight, resulting in at least one known instance of prolonged pretrial detention resting substantially on an algorithmic match, and risking the incorporation of historical patterns of discriminatory classification into models that appear statistically neutral. The crisis identified in this paper is therefore not confined to the case law of a particular jurisdiction; it is inherent in applying a due process framework designed for reasoned and individualised decisions by human agents to an apparatus that lacks those qualities.
*****
Footnotes
1. U.S. Const. amend. V; U.S. Const. amend. XIV, § 1.
2. India Const. art. 21.
3. Maneka Gandhi v. Union of India, AIR 1978 SC 597 (India).
4. Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1 (India).
5. Julia Angwin, Jeff Larson, Surya Mattu & Lauren Kirchner, Machine Bias, ProPublica (May 23, 2016), https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing.
6. State v. Loomis, 2016 WI 68, ¶¶ 11–16, 19, 22 & n.18, 34, 371 Wis. 2d 235, 881 N.W.2d 749, cert. denied sub nom. Loomis v. Wisconsin, 137 S. Ct. 2290 (2017).
7. Loomis, 2016 WI 68, ¶¶ 8–9, 98–104; see also Recent Case, State v. Loomis, 130 Harv. L. Rev. 1530 (2017).
8. Loomis v. Wisconsin, 137 S. Ct. 2290 (2017) (mem.) (denying certiorari).
9. Angwin et al., supra note 5; Julia Dressel & Hany Farid, The Accuracy, Fairness, and Limits of Predicting Recidivism, 4 Sci. Advances eaao5580 (2018), https://doi.org/10.1126/sciadv.aao5580 (finding COMPAS no more accurate than predictions made by people with little or no criminal justice expertise).
10. William Dieterich, Christina Mendoza & Tim Brennan, COMPAS Risk Scales: Demonstrating Accuracy Equity and Predictive Parity, Northpointe Inc. Res. Dep’t (July 8, 2016), https://web.archive.org/web/20240624143309/http://go.volarisgroup.com/rs/430-MBX-989/images/ProPublica_Commentary_Final_070616.pdf.
11. Tracing the Rise of Predictive Policing in India, Software Freedom L. Ctr. India (Feb. 26, 2026), https://sflc.in/tracing-the-rise-of-predictive-policing-in-india/ (describing the Delhi Police’s Crime Mapping, Analytics and Predictive System); see also Purushraj Patnaik, Algorithmic Policing in India: The Case for a Governing Framework, Observer Res. Found. (May 29, 2026), https://www.orfonline.org/research/algorithmic-policing-in-india-the-case-for-a-governing-framework (describing Telangana’s TSCOP application).
12. Tracing the Rise of Predictive Policing in India, supra note 11.
13. Patnaik, supra note 11.
14. Ameya Bokil, Nikita Sonavane, Srujana Bej, Avaneendra Khare & Vaishali Janarthanan, Settled Habits, New Tricks: Casteist Policing Meets Big Tech in India, Transnat’l Inst. (May 19, 2021), https://www.tni.org/en/article/settled-habits-new-tricks.
15. The Digital Personal Data Protection Act, 2023, No. 22, Acts of Parliament, 2023 (India), § 17(1)(c) [hereinafter DPDP Act]; see Ministry of Electronics and Information Technology, Notification G.S.R. 843(E) (Nov. 13, 2025) (appointing eighteen months from its publication in the Official Gazette as the date on which §§ 3–5, 6(1)–(8), (10), 7–10 and 11–17 come into force).
16. Patnaik, supra note 11.
17. Loomis, 2016 WI 68, ¶¶ 51–56.
18. DPDP Act § 17(1)(c).
19. NJCM v. Netherlands (SyRI), C/09/550982/HA ZA 18-388, ECLI:NL:RBDHA:2020:1878, ¶¶ 6.7, 7.2 (District Ct. The Hague, Feb. 5, 2020) (Neth.); see Convention for the Protection of Human Rights and Fundamental Freedoms art. 8, Nov. 4, 1950, 213 U.N.T.S. 221.
20. See Puttaswamy, (2017) 10 SCC 1.