LLexfolio

Algorithmic Bias in Predictive Policing and Criminal Profiling

Written by

in

Open access journal · Editorial and peer review

Lexfolio Journal of Law, Policy and Social Sciences

  • LJLPS
  • Published by Lexfolio, India
  • Bi-monthly, rolling publication
  • Online ISSN: applied for after Volume 1, Issue 1

Submit manuscript Author guidelines

Lexfolio/LJLPS/Algorithmic Bias in Predictive Policing and Criminal Profiling

Algorithmic Bias in Predictive Policing and Criminal Profiling

Online FirstEditorially reviewedOpen accessLJLPS-2026-0001
Research Article
Aman Sorout, Jagannath University, Jaipur
Dr. Alaknanda Rajawat, Faculty of Law, Jagannath University, Jaipur

Lexfolio Journal of Law, Policy and Social Sciences (LJLPS)

Received 3 November 2026 · Accepted 12 November 2026 · Published 12 November 2026

Abstract

Predictive policing refers to the use of statistical models, data analytics and artificial intelligence (AI) to estimate where crime may occur, identify patterns in criminal incidents, or assess the risk that an individual may be involved in future offending. Criminal profiling, in this context, includes the automated or semi-automated classification of persons as potential suspects or as presenting a particular risk. These technologies promise faster analysis, improved allocation of police resources and more informed investigation. Yet their outputs are only as reliable and fair as the data, assumptions, design choices and institutional practices on which they depend.This article examines how algorithmic bias can arise from historically unequal policing, under-reporting, proxy variables, incomplete datasets, opaque commercial systems, flawed model design and uncritical human reliance on automated scores. It distinguishes place-based crime forecasting from person-based profiling and explains why the latter presents especially serious risks to liberty and the presumption of innocence. The article analyses the Indian constitutional framework, including Articles 14, 19 and 21, the privacy doctrine in Justice K.S. Puttaswamy, procedural fairness under Maneka Gandhi, and the statutory setting created by criminal procedure and data protection legislation. It also considers the European Union Artificial Intelligence Act, the United Nations Committee on the Elimination of Racial Discrimination’s General Recommendation No. 36 (2020), and comparative cases such as State v. Loomis and R. v. Ewert.The central argument is that technology can assist lawful investigation, but an algorithmic prediction must never become a substitute for reasonable suspicion, admissible evidence, independent judgment or judicial scrutiny. India needs a clear statutory framework, mandatory impact assessments, independent audits, transparency, meaningful human review, accessible remedies and strict limits on the use of personal data. The article adopts a doctrinal and comparative research method and concludes that predictive policing should be governed by constitutional principles of legality, necessity, proportionality, equality, accountability and human dignity.

Keywords: Algorithmic bias; predictive policing; criminal profiling; artificial intelligence; equality; privacy; due process; facial recognition; criminal justice; India

1. Introduction

The criminal justice system has always relied on information. Police officers examine complaints, witness statements, crime records, forensic material, locations, relationships and patterns of conduct to investigate offences and prevent future harm. The digital transformation of society has greatly expanded the volume and speed of information available to law-enforcement agencies. Police databases, CCTV networks, mobile-phone records, online activity, vehicle registration systems, fingerprint repositories and digital case files can now be analysed through software capable of detecting correlations that may not be obvious to a human investigator.

Predictive policing is one consequence of this transformation. It generally describes the application of analytical techniques to estimate the likelihood of criminal events or identify persons, places or networks that may require police attention. Some systems forecast locations and times at which particular offences may occur. Other systems identify links between incidents, analyse social networks, or produce risk scores concerning individuals. Facial recognition and biometric matching systems may also contribute to investigative profiling by comparing an image or biometric feature against a database.

The attraction of predictive policing is understandable. Police resources are limited, crime patterns may be complex, and timely intervention can protect life and property. Data analysis can assist in identifying repeat burglary locations, patterns of vehicle theft, trafficking routes, cybercrime networks or links between otherwise disconnected incidents. Properly designed systems may reduce repetitive administrative work and help investigators organise large quantities of information. The Ministry of Home Affairs describes the Crime and Criminal Tracking Network and Systems (CCTNS) as a project intended to interlink police stations for investigation, data analytics, research, policy-making and citizen services. The Ministry also identifies the use of data analytics and AI/ML tools as part of the broader criminal justice information environment.1

However, the use of data does not automatically make policing objective. Police records are not neutral descriptions of all crime occurring in society. They are created through decisions about where police patrol, whom they stop, which complaints they register, what they investigate and which incidents result in arrest. A community that is more heavily policed may generate more recorded incidents even if its actual offending rate is not higher. If those records are used as training data, a model may reproduce the very deployment patterns that produced the records. The resulting prediction can send more officers to the same area, leading to more stops and arrests and still more data supporting the original prediction.

This phenomenon is commonly described as a feedback loop. The algorithm appears to confirm its prediction, but the confirmation may partly be the result of the policing activity prompted by the prediction itself. The United Nations Committee on the Elimination of Racial Discrimination has expressly warned that historical arrest data may reflect biased policing and that feeding such data into predictive systems can lead to over-policing and further arrests in the same neighbourhood.2

Algorithmic bias is not confined to race. In the Indian context, proxies such as locality, caste-linked residential segregation, religion-linked neighbourhood patterns, migration status, occupation, poverty, language, age, gender and prior contact with police may produce unequal effects. A model need not directly ask for caste or religion to reproduce discrimination. A postal code, school history, family association or frequency of police contact may operate as a proxy for a protected or vulnerable status. The danger is intensified where the system is opaque and the affected person cannot understand why they were flagged.

The distinction between predicting crime in a place and predicting criminality in a person is therefore essential. A place-based forecast may help allocate patrol resources, though it can still lead to discriminatory concentration of police attention. Person-based profiling can more directly affect an individual’s liberty, reputation, movement, privacy and access to a fair trial. A person should not be treated as a suspect merely because an algorithm identifies statistical similarity with persons previously arrested or because of their neighbourhood, associations or personal characteristics.

In India, this issue must be examined against the constitutional promise of equality, freedom and dignity. Article 14 prohibits arbitrary State action and guarantees equality before the law and equal protection of the laws. Article 19 protects specified freedoms, including speech, movement and association, subject to constitutionally permitted restrictions. Article 21 protects life and personal liberty and requires fair, just and reasonable procedure. The Supreme Court’s recognition of privacy as a fundamental right in Justice K.S. Puttaswamy (Retd.) v. Union of India3 provides a central constitutional foundation for scrutiny of mass data collection, biometric identification and automated profiling.

The subject is especially important as Indian policing institutions are increasingly adopting digital systems. In June 2026, a Press Information Bureau release described plans to use AI, machine learning and pattern analysis to identify repeat offenders and inter-State criminal networks, and referred to analysis of large digital record collections through CCTNS.4 This official development makes it necessary to examine not only the benefits of data-driven policing but also the safeguards required before such systems influence police attention, investigation or coercive decisions.

This article argues that predictive policing must be treated as a high-impact exercise of public power. It cannot be governed merely by procurement contracts, internal manuals or voluntary ethical principles. A constitutionally legitimate system requires a clear legal basis, a defined purpose, reliable and representative data, independent testing, disclosure of material limitations, meaningful human review, reasoned decisions and remedies for affected persons.

2. Research methodology, objectives and questions

2.1 Research methodology

The article follows a doctrinal and comparative legal research methodology. It analyses constitutional provisions, Indian statutes, judicial decisions, government publications, international human-rights instruments, regulatory materials and selected academic or investigative research. Comparative references are used to understand the legal treatment of algorithmic risk assessment in other jurisdictions; they are not treated as binding law in India. The research is qualitative and analytical. It does not claim to conduct an original field survey, algorithm audit or empirical assessment of any Indian police database.

2.2 Objectives of the study

To explain predictive policing, algorithmic bias and criminal profiling in accessible legal and technological terms.

To identify the principal sources through which discrimination may enter predictive systems.

To examine the constitutional implications of automated suspicion under Articles 14, 19 and 21.

To analyse Indian data-protection and criminal-procedure provisions relevant to police use of data and AI.

To study important Indian and comparative judicial decisions concerning privacy, fairness, profiling and risk assessment.

To evaluate international standards and propose safeguards for accountable use of predictive technologies.

2.3 Research questions

What is predictive policing, and how does it differ from conventional intelligence-led policing?

How do biased historical data, proxy variables and feedback loops produce discriminatory predictions?

Can an algorithmic risk score lawfully justify police surveillance, search, arrest, bail opposition or sentencing?

How do Indian constitutional rights and data-protection principles apply to algorithmic profiling?

What institutional and legal safeguards are necessary to ensure transparency, accountability and effective remedies?

2.4 Hypothesis

The working hypothesis is that, without enforceable safeguards, predictive policing systems may reproduce existing patterns of unequal policing and create new forms of arbitrary or discriminatory State action. Their use can be constitutionally acceptable only when the system is lawful, necessary, proportionate, independently audited and incapable of replacing evidence-based human judgment.

3. Meaning, nature and working of predictive policing

3.1 Meaning

Predictive policing is the use of data analysis to anticipate crime-related risks and inform the deployment of police resources. It may involve statistical forecasting, machine learning, network analysis, geospatial mapping, pattern recognition or automated classification. The term does not describe one single technology. It covers a range of systems with different purposes, data inputs and consequences.

3.2 Main categories

Place-based prediction: estimates locations or time periods in which specified offences may occur, often using past incident data and geographic variables.

Person-based prediction: assigns a risk score or flag to an individual based on personal data, past contacts, associations or behavioural indicators.

Network-based analysis: identifies possible relationships among persons, events, addresses, devices or organisations. A relationship may be relevant to investigation but does not itself prove criminal conduct.

Crime-linkage analysis: compares the modus operandi, timing, location or other features of separate incidents to identify possible links between offences.

Biometric and facial-recognition systems: compare faces, fingerprints or other biometric features to support identification. These systems may be used in investigations and can also feed into profiling.

3.3 Typical stages of an algorithmic system

A predictive system usually passes through several stages. First, an agency defines the problem it wants to address, such as burglary, vehicle theft or repeat offending. Second, it collects data from police records and sometimes other government or privately held sources. Third, developers clean and label the data, select variables and train a statistical model. Fourth, the model produces an output, such as a hotspot, alert, similarity score or risk category. Fifth, an officer or analyst interprets the output and decides whether to patrol, investigate, question, monitor or take another action. Finally, the system may be updated using new data.

Each stage involves human choices. The decision about which offences count, what period is examined, which variables are included, what counts as a successful prediction and how an officer should act on an alert can influence outcomes. Algorithmic systems therefore do not eliminate discretion; they redistribute it among police administrators, data custodians, vendors, programmers, analysts and frontline officers.

3.4 Predictive policing and traditional policing

Traditional policing often relies on local knowledge, complaints, witness information, intelligence and investigation. Predictive policing adds large-scale data processing and statistical inference. It can identify patterns across many records, but it may also conceal the assumptions behind those patterns. A human officer may be questioned about the basis of suspicion; a complex model may generate an apparently precise score without revealing how the score was produced. The correct approach is not to assume that human judgment is always fair or that software is always biased. Both can be wrong. The legal task is to ensure that neither is allowed to operate without reasons, evidence and accountability.

4. Algorithmic bias: concept, sources and forms

4.1 Meaning of algorithmic bias

Algorithmic bias refers to systematic and unfair differences in the outputs or effects of an automated system for different individuals or groups. Bias may arise even when developers do not intend to discriminate and even when a model does not use an expressly protected characteristic. It may be caused by unrepresentative data, inaccurate labels, unsuitable variables, design objectives, unequal error rates, deployment practices or the social environment in which the system operates.

4.2 Historical data bias

Historical police data record police activity as well as criminal activity. Arrest records, stop records and complaints may reflect the priorities, resources and practices of law-enforcement agencies. If a locality has historically been subject to intensive patrols, more offences may be detected and recorded there. A model trained on those records may interpret the higher number of recorded incidents as proof that the locality is inherently more dangerous. It may then recommend further patrols. The UN Committee on the Elimination of Racial Discrimination has identified precisely this risk in General Recommendation No. 36 (2020), explaining how arrest data shaped by biased policing may reproduce over-policing through a feedback loop.5

4.3 Sampling and representation bias

A dataset may fail to represent the full population or the full range of criminal conduct. For example, offences that are reported frequently may be better represented than offences that are hidden, under-reported or difficult to detect. Crimes committed in public spaces may appear more often in police data than offences occurring in private settings. Communities with limited access to police services may have fewer recorded complaints. A model trained on incomplete information can therefore produce a distorted picture of risk.

4.4 Proxy discrimination

A proxy is a variable that indirectly reflects another characteristic. Residential location, educational history, employment pattern, language, travel route, social network or prior police contact may correlate with caste, religion, ethnicity, class, gender or migration status. Removing a protected characteristic from the dataset does not necessarily remove discrimination if proxies remain. A model may appear neutral because it does not expressly use caste or religion, while still producing outcomes that disproportionately burden communities associated with particular locations or social conditions.

4.5 Measurement and label bias

A model learns from the labels it is given. If an arrest is treated as proof that a person committed an offence, the model may confuse police suspicion with guilt. If a re-arrest is used as a measure of reoffending, the result may be influenced by differences in surveillance and detection. A person who is more frequently stopped has more opportunities to be arrested again. The label may therefore measure exposure to policing rather than the underlying conduct that the model claims to predict.

4.6 Design and objective-function bias

A system is often optimised for a particular measure of performance, such as overall accuracy, false-positive rate, false-negative rate or efficient allocation of resources. Different fairness measures may conflict, particularly where groups have different base rates or where the data are imperfect. The choice of metric is a normative and legal decision, not a purely technical one. A system that maximises aggregate accuracy may still impose a disproportionate burden on a smaller group.

4.7 Automation bias and human overreliance

Automation bias occurs when a human decision-maker gives excessive weight to a computer-generated recommendation. A police officer may treat an algorithmic flag as objective because it is expressed numerically or generated by sophisticated software. The officer may overlook contradictory witness information, fail to investigate alternative explanations or record the score as if it were independent evidence. Human review is meaningful only when the reviewer has the information, authority, training and time to reject the system’s output.

5. Criminal profiling and the risk of automated suspicion

Criminal profiling involves the use of characteristics, patterns or inferred relationships to identify persons who may be connected with criminal conduct. Profiling may be based on a description provided by a witness, forensic evidence, modus operandi, a verified link to an offence or other objective information. Such investigative profiling is not inherently unlawful. It becomes problematic when a person is singled out primarily because of broad social characteristics, statistical associations or opaque risk scores rather than objective facts connected to a particular offence.

It is important to distinguish a suspect profile from proof of guilt. A profile may generate an investigative lead, but it cannot establish that a person committed an offence. A person’s presence in a neighbourhood, association with another person, financial condition, travel history, online speech or previous police contact cannot, without more, establish criminal conduct. The presumption of innocence requires the State to prove the charge through legally admissible evidence; it does not permit punishment based on a forecast of future behaviour.

Profiling may also be collective. A system may classify a neighbourhood, a group of young people, a set of social-media users or a network of contacts as high risk. The risk is that the classification attaches suspicion to individuals who have not engaged in unlawful conduct. It may chill lawful association, religious practice, political participation and freedom of expression. The UN Committee’s General Recommendation No. 36 warns that algorithmic profiling may affect privacy, liberty, presumption of innocence, movement, assembly and access to an effective remedy.6

A lawful investigative lead should therefore satisfy a minimum standard: it must be connected to a defined offence or legitimate investigative purpose; based on relevant and verifiable facts; independently assessed by a trained officer; recorded with reasons; and open to challenge where it materially affects an individual. A prediction alone should never be sufficient for arrest, search, detention or adverse judicial findings.

6. Predictive policing in India: institutional and technological context

6.1 Digital criminal justice infrastructure

India’s criminal justice system has developed a substantial digital information infrastructure. The Ministry of Home Affairs describes CCTNS as a nationwide project intended to connect police stations through common application software and support investigation, data analytics, research, policy-making and citizen services. The Ministry’s description also refers to the Inter-Operable Criminal Justice System (ICJS), e-Prisons and the use of data analytics and AI/ML tools in investigations.7 These systems can improve the availability and exchange of information, but the ability to connect records also increases the need for accuracy, access controls, purpose limitation and audit trails.

6.2 Facial recognition and biometric identification

Facial recognition systems compare an image of a person with images stored in a database. They may assist in identifying a missing person or comparing a suspect image with a lawfully maintained record. However, a match is probabilistic, not conclusive. Image quality, camera angle, lighting, age, demographic representation and the quality of the reference database can affect accuracy. A false match can expose an innocent person to questioning, surveillance or arrest. A match must therefore be treated as an investigative lead requiring independent corroboration, not as proof of identity or guilt.

The Bureau of Police Research and Development’s 2025 Indian Police Journal discussion of AI in policing identifies the National Automated Facial Recognition System and describes its integration with databases such as CCTNS and ICJS as part of India’s technological policing environment.8 The existence of such infrastructure makes it important to establish clear rules about the purposes for which images may be collected, the legal basis for database access, retention periods, accuracy testing, independent review and correction of false matches.

6.3 Emerging AI-led crime analysis

A Press Information Bureau release dated 19 June 2026 reported that the Government was developing a framework using AI, machine learning and pattern analysis to identify repeat offenders and inter-State criminal networks. It referred to large collections of fingerprint, narcotics and human-trafficking records and to AI-based analysis of digital records available through CCTNS.9 This is an official description of a developing policy direction. It should not be read as proof that every proposed analytical system is already deployed in every State or that a specific algorithm has been independently validated.

6.4 Federalism and police accountability

Police and public order are primarily State subjects under the Seventh Schedule of the Constitution, while the Union has important roles in national coordination, central agencies, criminal law, telecommunications and digital infrastructure. This distribution creates a challenge for algorithmic governance. A national database or common technical platform may be used by multiple State police forces, each with different operational rules and data quality. A sound framework must therefore establish minimum national standards while respecting constitutional federalism and ensuring that State-level use is reviewable.

6.5 The Indian evidence gap

Publicly available material establishes that India is developing and using digital policing infrastructure, but there is limited public, independent evidence about the accuracy, demographic error rates, operational protocols and real-world impact of specific predictive policing models. This article therefore avoids claiming that a particular Indian system has been proven discriminatory. The legal concern is prospective and structural: where systems are deployed without published validation, independent audits and accessible remedies, the public cannot determine whether they are fair, accurate or lawful.

7. Constitutional and statutory framework in India

7.1 Article 14: equality and non-arbitrariness

Article 14 guarantees equality before the law and equal protection of the laws. State action must not be arbitrary, irrational or discriminatory. In E.P. Royappa v. State of Tamil Nadu10, the Supreme Court connected arbitrariness with the constitutional guarantee of equality. An algorithmic system used by police is an exercise of State power and is therefore subject to Article 14. A system that disproportionately targets a community through unreliable proxies, or treats a statistical correlation as an adequate reason for coercive action, may fail the requirement of non-arbitrariness.

Equality analysis must examine both the design and the effects of a system. It is not sufficient for an agency to say that the model does not use caste, religion or gender. The agency should assess whether variables such as location, police contact, income or social networks create indirect discrimination. It should also examine false-positive and false-negative rates across relevant groups, while collecting and handling sensitive audit information under appropriate safeguards.

7.2 Article 21: life, liberty and dignity

Article 21 provides that no person shall be deprived of life or personal liberty except according to procedure established by law. In Maneka Gandhi v. Union of India11, the Supreme Court held that the procedure must be fair, just and reasonable and not arbitrary or oppressive. Algorithmic flags can affect liberty when they lead to surveillance, questioning, search, arrest, detention, bail decisions or sentencing. Any such use must have a lawful basis and must satisfy procedural fairness.

A risk score cannot replace the legal grounds required for arrest or detention. Under the criminal procedure framework, coercive action must be justified by the applicable statutory conditions and constitutional safeguards. The Bharatiya Nagarik Suraksha Sanhita, 2023 now governs criminal procedure, replacing the Code of Criminal Procedure, 1973, subject to its commencement and transitional provisions. Its arrest and investigation powers must be exercised consistently with Articles 21 and 22. An algorithm may at most provide an investigative lead; the officer must independently assess the facts and record legally sufficient reasons.

7.3 Article 19: speech, movement and association

Algorithmic monitoring of online activity, public gatherings or social networks can chill freedoms protected by Article 19. A person may avoid lawful political speech, protest, journalism or association if they believe that their posts, contacts or attendance at a public event will cause them to be classified as a security risk. Restrictions on Article 19 freedoms must fall within the constitutionally specified grounds and satisfy proportionality. Generalised surveillance based on political views or lawful association is not a legitimate substitute for evidence of criminal conduct.

7.4 Privacy as a fundamental right

In Justice K.S. Puttaswamy (Retd.) v. Union of India (2017), a nine-judge Bench recognised privacy as a fundamental right protected by the Constitution. Privacy includes bodily privacy, informational privacy, decisional autonomy and the ability to control aspects of personal life. The judgment requires State interference with privacy to satisfy legality, a legitimate State aim and proportionality, including safeguards against abuse. Predictive policing may involve extensive processing of location, biometric, communication, association and behavioural data. The more intrusive the collection and inference, the stronger the justification and safeguards required.

The Aadhaar judgment, Justice K.S. Puttaswamy (Retd.) v. Union of India (2019)12, further considered proportionality in the context of identity infrastructure and personal data. It demonstrates that technological efficiency and administrative convenience do not, by themselves, answer the constitutional question. The State must establish a lawful objective and show that the interference is proportionate and appropriately safeguarded.

7.5 Article 22 and safeguards against arbitrary arrest

Article 22 protects persons arrested or detained by requiring communication of grounds of arrest, access to legal representation and production before a magistrate within the constitutionally prescribed period, subject to its terms and exceptions. A predictive score cannot dilute these safeguards. If a police officer relies on an algorithmic alert, the officer must still disclose the legally relevant grounds and comply with statutory arrest procedures. The magistrate must not treat the existence of a computer-generated risk score as a substitute for independent judicial scrutiny.

7.6 Digital Personal Data Protection Act, 2023

The Digital Personal Data Protection Act, 2023 (DPDP Act) establishes a framework for processing digital personal data. Its stated purpose is to recognise both the individual’s interest in protecting personal data and the need to process such data for lawful purposes.13 The Act contains provisions concerning notice, consent, obligations of data fiduciaries, rights of data principals, security safeguards and exemptions. However, the Act must be read with its commencement notifications. A notification dated 13 November 2025 brought specified institutional and enabling provisions into force immediately, while many substantive provisions, including sections 3 to 17, were scheduled to commence eighteen months after publication, i.e. 13 May 2027.14 Accordingly, as at 1 November 2026, the substantive data-processing regime should not be described as fully operational.

Section 17(2)(c) provides an exemption for processing necessary for the prevention, detection, investigation or prosecution of offences or contraventions of law, subject to the statutory scheme and commencement. Such an exemption may be important for law enforcement, but it should not be interpreted as a blanket constitutional licence for unrestricted surveillance or discriminatory profiling. Constitutional rights, applicable procedural laws, judicial review and other legal duties continue to constrain State action. The future operation of the DPDP Act should be harmonised with robust police-specific safeguards, rather than treated as a complete answer to algorithmic bias.

7.7 Information Technology Act, 2000 and other laws

The Information Technology Act, 2000 regulates specified aspects of electronic records, cyber offences and intermediary obligations. Section 69 and section 69B contain powers relating to interception, monitoring or decryption and traffic-data monitoring, subject to statutory conditions and rules. These provisions do not create a general power to deploy any predictive policing model. Any access to communications or digital records must be authorised under the relevant law and used for the purpose for which access is permitted.

The Bharatiya Sakshya Adhiniyam, 2023 governs admissibility and proof of electronic and digital records. An algorithmic output should not be treated as self-authenticating proof. Where such an output is relied upon in criminal proceedings, questions arise about the source data, method, integrity, validation, chain of custody, expert evidence and the ability of the accused to test the material. The Bharatiya Nyaya Sanhita, 2023 defines offences but does not convert a prediction of future conduct into an offence. Criminal liability must be based on conduct satisfying the elements of an offence.

7.8 Criminal procedure and the threshold for coercive action

The distinction between an investigative lead and a lawful ground for coercion is fundamental. A prediction may justify further verification where the underlying information is lawfully obtained and relevant. It cannot, by itself, satisfy the statutory threshold for search, arrest or detention. The police must identify objective facts connecting the person to a specific offence and must comply with the BNSS and constitutional safeguards. The same principle applies to bail and sentencing: a score cannot replace the court’s duty to assess the individual facts, the applicable legal test and the reliability of the material.

8. International human rights and comparative legal framework

8.1 United Nations CERD General Recommendation No. 36 (2020)

The Committee on the Elimination of Racial Discrimination adopted General Recommendation No. 36 in 2020 on preventing and combating racial profiling by law-enforcement officials. It recognises that big data, automated decision-making and AI can improve efficiency but can also reproduce or reinforce bias. It identifies several sources of bias, including protected characteristics, proxy variables, biased datasets and poor-quality data. It specifically warns that predictive policing based on historical records may create feedback loops and over-policing.15

The Committee recommends prior human-rights impact assessment, transparency, independent oversight, continuing monitoring, public disclosure of the use and functioning of systems, community participation, independent audits and effective remedies. It also calls for States to ensure that private companies involved in developing or supplying law-enforcement algorithms respect human rights. These recommendations provide a valuable international benchmark for India, even though their precise legal force must be understood in the context of the relevant treaty obligations and India’s international commitments.

8.2 International Covenant on Civil and Political Rights

The International Covenant on Civil and Political Rights (ICCPR) protects equality before the law, privacy, freedom of expression, peaceful assembly and liberty and security of person. Article 17 prohibits unlawful or arbitrary interference with privacy, family, home or correspondence. Article 9 protects liberty and security and prohibits arbitrary arrest or detention. Article 14 guarantees equality before courts and a fair hearing. Article 26 guarantees equality before the law and equal protection without discrimination. These provisions are directly relevant when algorithmic systems are used to monitor individuals or influence criminal justice decisions.

8.3 European Union Artificial Intelligence Act

Regulation (EU) 2024/1689, known as the Artificial Intelligence Act, adopts a risk-based approach to AI regulation. Article 5 prohibits certain practices, including AI systems used to assess or predict the risk that a natural person will commit a criminal offence where the assessment is based solely on profiling or personality traits and characteristics, without objective verifiable facts and human assessment. The Act distinguishes this from permitted risk analytics that do not profile individuals, such as analysis of trafficking routes.16

The Act also classifies specified law-enforcement uses as high-risk, including certain systems for profiling natural persons in the detection, investigation or prosecution of criminal offences. High-risk systems are subject to requirements relating to risk management, data governance, documentation, record-keeping, transparency, human oversight, accuracy, robustness and cybersecurity. The EU model is useful because it differentiates between a prohibited form of individual prediction and tightly regulated investigative tools. It demonstrates that not every use of AI in policing must be banned, but uses affecting fundamental rights require stronger safeguards.

8.4 Comparative case: State v. Loomis

In State v. Loomis, 881 N.W.2d 749 (Wis. 2016), the Wisconsin Supreme Court considered the use of the COMPAS risk assessment tool during sentencing. The Court permitted the sentencing court to consider the assessment as one factor, but emphasised important limitations. The score could not be determinative, and courts had to recognise concerns relating to proprietary methodology, validation, group-based scoring and the proper use of the tool. The judgment is significant because it illustrates that a risk score may influence liberty even when it is formally described as advisory.17

The decision should not be read as blanket approval of algorithmic sentencing. Its cautions show that transparency and meaningful judicial reasoning are essential. In India, where constitutional protections and sentencing principles apply, a proprietary risk score should never become a substitute for individualised judicial assessment.

8.5 Comparative case: R. v. Ewert

In R. v. Ewert, 2018 SCC 30, the Supreme Court of Canada considered the reliability of actuarial risk assessment tools when used in relation to Indigenous offenders. The Court addressed the statutory duty of correctional authorities to take reasonable steps to ensure that assessment tools are accurate and appropriate for Indigenous offenders. The case is important for its recognition that a tool validated on one population cannot automatically be assumed to be valid for another. Local validation and attention to the characteristics of affected groups are essential.

9. Important judicial decisions

9.1 Justice K.S. Puttaswamy (Retd.) v. Union of India (2017)

The nine-judge Bench in Puttaswamy unanimously recognised privacy as a fundamental right. The judgment located privacy within liberty, dignity and the guarantees of Part III of the Constitution. It rejected the idea that privacy is a narrow or isolated interest and recognised informational privacy as an important dimension. For predictive policing, the case requires the State to justify collection and analysis of personal data through legality, legitimate purpose and proportionality. A database cannot be expanded or linked merely because technology makes it possible.

9.2 Maneka Gandhi v. Union of India (1978)

In Maneka Gandhi, the Supreme Court interpreted Article 21 to require a procedure that is fair, just and reasonable. The decision connected Articles 14, 19 and 21 and rejected arbitrary procedures. Its relevance is direct: if an algorithmic alert results in coercive action, the procedure must be legally authorised and fair. A person should not suffer a deprivation of liberty on the basis of an undisclosed or unchallengeable score.

9.3 E.P. Royappa v. State of Tamil Nadu (1974)

E.P. Royappa developed the principle that arbitrariness is antithetical to equality. An automated system that applies opaque or irrational criteria, or produces unequal outcomes without adequate justification, may violate Article 14. The case supports judicial scrutiny of the logic, purpose and effects of algorithm-assisted police decisions.

9.4 Shreya Singhal v. Union of India (2015)

In Shreya Singhal18, the Supreme Court struck down section 66A of the Information Technology Act as unconstitutional because of its vagueness and chilling effect on free speech. Although the case did not concern predictive policing, it is relevant to AI systems that monitor online speech or classify persons based on posts, keywords or associations. Vague standards can encourage arbitrary enforcement and deter lawful expression. Any digital monitoring that burdens Article 19 freedoms must be clearly authorised and narrowly tailored.

9.5 State v. Loomis (2016)

The Wisconsin Supreme Court’s decision in Loomis is a comparative authority on the use of risk scores at sentencing. It recognised the possible utility of COMPAS but imposed cautions against determinative reliance. It also highlighted the difficulty caused by proprietary methods and the need for courts to understand the tool’s limitations. Its central lesson is that an algorithmic score must remain subordinate to independent judicial reasoning.

9.6 R. v. Ewert (2018)

The Supreme Court of Canada’s decision in Ewert highlights the importance of validating assessment tools for the population on which they are used. A tool that performs acceptably in one setting may produce unreliable outcomes in another. For India, this is particularly significant because imported tools may not reflect local languages, social structures, policing patterns, demographic diversity or criminal justice conditions.

10. Challenges in regulation and accountability

10.1 Absence of a dedicated statutory framework

India does not yet have a comprehensive, police-specific statute setting out the permissible purposes, approval process, audit standards, disclosure duties and remedies for predictive policing systems. General constitutional and statutory rules remain applicable, but they may not answer technical questions about model validation, error rates, training data, procurement and algorithmic impact assessment. A dedicated legal framework is needed to translate constitutional principles into operational duties.

10.2 Data quality and interoperability

Police databases may contain duplicate records, inconsistent spellings, incomplete fields, outdated addresses, incorrect classifications or unverified allegations. When datasets are linked across agencies, an error in one system may spread to others. Interoperability should therefore be accompanied by data correction procedures, source tracing, access logs and clear responsibility for rectifying inaccurate records.

10.3 Vendor dependence and trade secrecy

Public agencies may depend on private vendors for software, updates and technical support. If the contract does not require audit access, documentation and independent testing, the State may be unable to explain a decision made using the tool. Procurement contracts should preserve public-law accountability and should not permit trade secrecy to obstruct fair-trial rights or independent oversight.

10.4 Measuring fairness

There is no single statistical measure that captures every aspect of fairness. Overall accuracy may conceal group-specific errors. A model may have a high false-positive rate for one group, or may perform poorly for people whose data are underrepresented. Agencies need context-sensitive testing, disaggregated evaluation where lawful and safe, and independent assessment of both statistical performance and real-world consequences.

10.5 Human oversight in practice

A nominal human-in-the-loop requirement is inadequate if officers are trained to follow the score, lack access to underlying information or face operational pressure to accept automated recommendations. Human reviewers must be able to reject the output without penalty, record reasons and seek technical assistance. The decision-maker should be responsible for the final action and should not attribute responsibility to the software.

10.6 Remedies and access to information

An individual may not know that an algorithm contributed to police attention or an adverse decision. Without notice, access to relevant information and a complaint mechanism, a person cannot effectively challenge a false match or discriminatory profile. At the same time, disclosure must be balanced with legitimate investigative confidentiality. A workable system can provide delayed notice, judicially supervised disclosure or independent review where immediate disclosure would genuinely prejudice an investigation.

11. Recommendations and safeguards

11.1 Enact a dedicated legal framework

Parliament and State legislatures, within their respective constitutional fields, should establish clear rules for AI and predictive analytics in law enforcement. The law should define predictive policing, criminal profiling, biometric identification and high-risk automated decision systems. It should specify permissible purposes, prohibited uses, approval procedures, retention limits, disclosure duties, audit obligations and penalties for misuse.

11.2 Prohibit prediction based solely on personal traits

No person should be classified as likely to commit an offence solely on the basis of personality traits, social background, protected characteristics, family association, neighbourhood or other broad profiling factors. Individualised assessment must be grounded in objective and verifiable facts connected to a legitimate investigation. This approach is consistent with the prohibition contained in Article 5 of the EU AI Act for specified forms of individual predictive policing.19

11.3 Conduct algorithmic and human-rights impact assessments

Before procurement or deployment, an agency should conduct an assessment identifying the purpose, legal basis, data sources, affected groups, possible errors, privacy impacts, discriminatory effects and available remedies. The assessment should be reviewed by an independent body and updated whenever the model, data or intended use changes. Community consultation should be included where a system may affect particular neighbourhoods or vulnerable groups.

11.4 Independent audits and local validation

Every high-impact system should undergo pre-deployment and periodic independent audits. Audits should test accuracy, false-positive and false-negative rates, demographic disparities, data quality, security, robustness and the effects of deployment. Tools developed or validated in another country should not be used in India without local validation. Audit results should be published in a form that protects personal information and legitimate security interests.

11.5 Transparency and explainability

Authorities should publicly disclose the existence, purpose, broad operation and legal basis of predictive systems. For any material adverse decision, the affected person should receive an intelligible explanation of the role played by the algorithm, subject to narrowly tailored and reviewable limits needed to protect an investigation. A numerical score without an explanation is not an adequate reason for State action.

12. Conclusion

Predictive policing and criminal profiling represent a significant change in the way law-enforcement agencies collect, interpret and act upon information. Data analytics can help identify patterns, link incidents, allocate resources and support investigations. In a large and diverse country such as India, digital systems may improve coordination between police stations and criminal justice institutions. Yet the promise of efficiency cannot be separated from the constitutional limits on State power.

The principal danger is that algorithmic systems may reproduce historical inequalities while presenting their outputs as neutral scientific facts. Police records reflect institutional choices and unequal patterns of reporting, detection and enforcement. If those records are used without careful assessment, the model may direct greater attention towards the same communities, generate more recorded incidents and reinforce its own predictions. Proxy variables, poor-quality data, unequal error rates, opacity and automation bias can further magnify the problem.

The Indian Constitution provides a strong foundation for regulating these risks. Article 14 requires equality and non-arbitrary State action. Article 19 protects freedoms that may be chilled by indiscriminate digital monitoring. Article 21 protects life, liberty, dignity and privacy, while Article 22 provides safeguards in cases of arrest and detention. Puttaswamy, Maneka Gandhi, Royappa and Shreya Singhal establish principles that apply to technological policing even though they did not all concern predictive algorithms directly.

International developments reinforce the need for a rights-based approach. CERD General Recommendation No. 36 identifies the feedback-loop danger and calls for transparency, impact assessment, independent oversight and effective remedies. The EU AI Act prohibits specified forms of individual prediction based solely on profiling and subjects other law-enforcement applications to a risk-based regulatory framework. State v. Loomis and R. v. Ewert show why risk scores must be limited, explained and validated for the population in which they are used.

India’s emerging digital policing infrastructure and the policy direction towards AI-based analysis make it urgent to adopt enforceable safeguards. The DPDP Act, 2023 is relevant to digital personal data, but its commencement is phased and its law-enforcement exemptions mean that it cannot alone resolve every issue raised by predictive policing. A police-specific framework should require legality, necessity, proportionality, data minimisation, independent audits, local validation, transparency, human review and effective remedies.

The most important principle is that an algorithm may assist an investigation but must not become the investigator, prosecutor and judge. A prediction is not proof; a correlation is not guilt; and a risk score is not a lawful ground for arrest. The legitimacy of modern policing must be measured not only by its capacity to detect crime but also by its ability to respect the rights and dignity of every person. Technology should strengthen the rule of law, not automate suspicion.

Footnotes

  1. Ministry of Home Affairs, Government of India, ‘ICJS/NCRB Administration’ <https://www.mha.gov.in/women-safety-iv-section> accessed 1 November 2026. ↩
  2. Committee on the Elimination of Racial Discrimination, ‘General Recommendation No 36 (2020) on Preventing and Combating Racial Profiling by Law Enforcement Officials’ (17 December 2020) UN Doc CERD/C/GC/36. ↩
  3. Justice KS Puttaswamy (Retd) v Union of India (2017) 10 SCC 1. ↩
  4. Press Information Bureau, ‘Union Home Minister and Minister of Cooperation Shri Amit Shah addresses the inaugural session of the 26th All India Fingerprint Conference 2026 in New Delhi as Chief Guest’ (19 June 2026) <https://www.pib.gov.in/PressReleasePage.aspx?PRID=2275306> accessed 12 November 2026. ↩
  5. CERD, General Recommendation No 36 (n 2). ↩
  6. CERD, General Recommendation No 36 (n 2). ↩
  7. Ministry of Home Affairs (n 1). ↩
  8. Bureau of Police Research and Development, ‘Artificial Intelligence in Policing’, The Indian Police Journal (January–June 2025). ↩
  9. Press Information Bureau (n 4). ↩
  10. EP Royappa v State of Tamil Nadu (1974) 4 SCC 3. ↩
  11. Maneka Gandhi v Union of India (1978) 1 SCC 248. ↩
  12. Justice KS Puttaswamy (Retd) v Union of India (2019) 1 SCC 1. ↩
  13. Digital Personal Data Protection Act 2023, Preamble. ↩
  14. Ministry of Electronics and Information Technology, Notification GSR 843(E) (13 November 2025). ↩
  15. CERD, General Recommendation No 36 (n 2). ↩
  16. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), art 5. ↩
  17. State v Loomis 881 NW 2d 749 (Wis 2016). ↩
  18. Shreya Singhal v Union of India (2015) 5 SCC 1. ↩
  19. Regulation (EU) 2024/1689 (n 16). ↩

Bibliography

A. Constitutional and statutory materials

The Constitution of India, 1950.

The Bharatiya Nagarik Suraksha Sanhita, 2023.

The Bharatiya Nyaya Sanhita, 2023.

The Bharatiya Sakshya Adhiniyam, 2023.

The Information Technology Act, 2000.

The Digital Personal Data Protection Act, 2023.

The International Covenant on Civil and Political Rights, 1966.

Regulation (EU) 2024/1689 (Artificial Intelligence Act).

B. Cases

Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1.

Justice K.S. Puttaswamy (Retd.) v. Union of India, (2019) 1 SCC 1 (Aadhaar judgment).

Maneka Gandhi v. Union of India, (1978) 1 SCC 248.

E.P. Royappa v. State of Tamil Nadu, (1974) 4 SCC 3.

Shreya Singhal v. Union of India, (2015) 5 SCC 1.

State v. Loomis, 881 N.W.2d 749 (Wis. 2016).

R. v. Ewert, 2018 SCC 30, [2018] 2 SCR 165.

C. Reports, articles and institutional materials

Committee on the Elimination of Racial Discrimination, General Recommendation No. 36 (2020) on Preventing and Combating Racial Profiling by Law Enforcement Officials, CERD/C/GC/36 (17 December 2020).

Ministry of Home Affairs, Government of India, ‘ICJS/NCRB Administration’ (official website), https://www.mha.gov.in/women-safety-iv-section (accessed 1 November 2026).

Press Information Bureau, Government of India, ‘AI/ML and pattern analysis framework for identifying repeat offenders and inter-State criminal networks’, 19 June 2026, https://www.pib.gov.in/PressReleasePage.aspx?PRID=2275306 (accessed 1 November 2026).

Bureau of Police Research and Development, ‘Artificial Intelligence in Policing’, The Indian Police Journal, January–June 2025, https://bprd.nic.in/uploads/pdf/The%20Indian%20Police%20Journal%20Jan-June-2025.pdf.

European Commission, AI Act Service Desk, ‘Article 5: Prohibited AI Practices’, https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-5 (accessed 1 November 2026).

European Commission, AI Act Service Desk, ‘Law Enforcement’, https://ai-act-service-desk.ec.europa.eu/en/law-enforcement (accessed 1 November 2026).

Council of Europe, Handbook on Human Rights and Artificial Intelligence, ‘Relevant human rights and principles’, https://www.coe.int/en/web/cddh-handbook-on-ai-and-hr/relevant-human-rights-and-principles1 (accessed 1 November 2026).

Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner, ‘How We Analyzed the COMPAS Recidivism Algorithm’, ProPublica (23 May 2016), https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm.

Julia Angwin and Jeff Larson, ‘ProPublica Responds to Company’s Critique of Machine Bias Story’, ProPublica (29 July 2016), https://www.propublica.org/article/propublica-responds-to-companys-critique-of-machine-bias-story.

U.S. Department of Justice, Office of Legal Policy, Artificial Intelligence and Criminal Justice, Final Report (3 December 2024), https://www.justice.gov/olp/media/1381796/dl.

India Code, Digital Personal Data Protection Act, 2023, https://www.indiacode.nic.in/indiacode/handle/123456789/22037?view_type=browse.

Government of India, Gazette Notification G.S.R. 843(E), 13 November 2025, commencement notification under the Digital Personal Data Protection Act, 2023, https://egazette.gov.in/WriteReadData/2025/267647.pdf.

Article information and declarations

Article history
Received 3 November 2026 · Accepted 12 November 2026 · Published online 12 November 2026
Review
Editorial review by the Editor-in-Chief, including originality, citation and copyright checks. Handling editor: Shivanshu K. Srivastava, Editor-in-Chief.
Conflict of interest
The author(s) declare no conflict of interest.
Funding
The author(s) received no specific funding for this work.
Originality
Checked for similarity and citation accuracy before publication. The author(s) warrant that the work is original and not published elsewhere.
Licence
© The author(s). Open access under CC BY 4.0.
Corrections
Any correction or retraction will be published as a linked notice on this page.

How to cite (OSCOLA): Aman Sorout and Dr. Alaknanda Rajawat, 'Algorithmic Bias in Predictive Policing and Criminal Profiling' (2026) Lexfolio Journal of Law, Policy and Social Sciences

How to cite (Bluebook): Aman Sorout and Dr. Alaknanda Rajawat, Algorithmic Bias in Predictive Policing and Criminal Profiling, Lexfolio J.L. Pol’y & Soc. Sci. (2026)

© The author(s). Published by Lexfolio under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. Opinions are the author’s own. Submit your paper to LJLPS.

Lexfolio Journal of Law, Policy and Social Sciences

Published by Lexfolio, Lucknow, Uttar Pradesh, India. Editorial office: support@lexfolio.in. Except where otherwise noted, content is licensed under CC BY 4.0.