Predictive Policing in India: From CMAPS and CCTNS to State-Level AI Risk and Surveillance Systems

Predictive Policing in India: From CMAPS and CCTNS to State-Level AI Risk and Surveillance Systems

Research cutoff: August 2, 2026

Executive findings, scope, and standard of verification

India does not presently have one national “predictive policing system.” It has a layered and uneven digital-policing ecosystem in which national record infrastructure—principally the Crime and Criminal Tracking Network & Systems, or CCTNS—is increasingly connected to courts, prisons, prosecution offices, forensic laboratories, fingerprint repositories, emergency-response systems, sexual-offender and narcotics databases, CCTV networks, facial-recognition tools, and state or district analytics applications. The Inter-Operable Criminal Justice System, or ICJS, is the principal integration architecture: the Ministry of Home Affairs describes it as joining CCTNS, e-Courts, e-Prisons, e-Prosecution, and e-Forensics through a Supreme Court e-Committee-approved Data Sharing Matrix, with NCRB as nodal agency and NIC as technology partner. The ministry expressly identifies analytics and AI/ML-assisted investigation as an ICJS 2.0 objective, but that statement does not establish that every connected application predicts crime or future conduct. citeturn15view3turn14search4

The strongest finding of this investigation is therefore terminological. In Indian official publicity and technology journalism, “AI,” “predictive,” “crime analytics,” “facial recognition,” “criminal intelligence,” “hotspot mapping,” “decision support,” “network analysis,” and ordinary database search are frequently blended together. Technically, they are different. Searching an accused person’s name across FIRs is record retrieval. Linking a person to associates, vehicles, telephone records, or cases is entity resolution or network analysis. Matching a face against a gallery is biometric identification. Mapping past FIR locations is descriptive geospatial analysis. Estimating where offenses may occur during a future time interval is place-based prediction. Assigning an individual a probability or risk category for future offending is person-based predictive policing. Only the last two are predictive in the relevant sense, and person-based prediction is the most constitutionally intrusive.

No examined Indian police organization had, by the cutoff date, publicly released the documentation needed to validate a mature person-based predictive-policing model: a complete feature list, target definition, training and validation periods, false-positive and false-negative rates, subgroup performance, calibration, decision threshold, human-override rules, deployment logs, outcome evaluation, and an independent audit. Some systems may contain unpublished risk-ranking or recommendation functions. The evidentiary point is that their existence, design, and performance cannot be independently established from public material. Uncertainty here is not a research gap to be filled with inference; it is itself a governance finding.

Delhi’s Crime Mapping, Analytics and Predictive System, or CMAPS, unquestionably existed, but the best independent public evidence—the Comptroller and Auditor General’s audit—shows a major gap between planned capability and achieved implementation. CMAPS was conceived with crime analytics, security threat-rating, news geotagging and clustering, social-network analysis, text annotation, criminal profiling, and a mobile application. By September 2019, however, it largely drew from the Delhi Police PA-100 system, lacked full CCTNS integration, and had not implemented the more advanced security and open-source modules. Records reviewed by the CAG showed no substantive Delhi Police–ISRO/ADRIN communication after March 2018 concerning the stalled incremental work. The verified operational core was consequently much closer to crime mapping and descriptive decision support than a validated predictive engine. citeturn11view1turn11view2

A separate Delhi Police-associated research project published in 2026 describes a spatiotemporal kernel-density hotspot method and a real-world evaluation involving patrol-vehicle allocation. That work is more recognizably predictive than the audited CMAPS deployment because it estimates future spatial concentration from prior events. It should not, however, be silently treated as a proven new CMAPS production version: the public research does not establish that the algorithm replaced CMAPS, operates continuously citywide, or has undergone an independent impact evaluation measuring crime reduction, displacement, unequal patrol burden, or false alarms. citeturn0search21turn0academia29

CCTNS and ICJS are best understood as enabling infrastructure, not prediction systems. CCTNS digitizes and searches FIRs, investigations, charge sheets and related entities, while ICJS makes records from the criminal-justice “pillars” interoperable. In June 2025, the government reported CCTNS implementation in 17,712 police stations, alongside ICJS-linked coverage of 1,373 prisons, 117 forensic laboratories, 751 prosecution districts and 3,637 court complexes. Those figures indicate vast potential data reach, but coverage counts do not establish completeness, correctness, timely updating, or lawful reuse. citeturn0search13turn15view3

Telangana’s TSCOP, Uttar Pradesh’s Trinetra and YAKSH, Maharashtra’s MARVEL organization and MahaCrimeOS products, Ghazipur’s AI-SPS, and Washim’s Smart Prahari all have evidence of existence, but not the same quality of evidence. TSCOP has been described in reporting as a mobile operational platform with criminal-record search and face-matching capabilities, and its reality is additionally corroborated by reporting on a serious data exposure; no current public technical specification or independent algorithmic evaluation was located. MARVEL is formally documented as a Maharashtra government-backed special-purpose entity, while MahaCrimeOS and related tools are documented principally through government announcements and partner or vendor case studies. Trinetra 2.0 and YAKSH are strongly associated in public reporting with Staqu’s CrimeGPT/JARVIS technology but lack a public government technical design and independent validation. AI-SPS has an unusually informative official district privacy page showing that its implemented core is police attendance, GPS duty and patrol tracking, and operational monitoring—not the broad “predictive policing” functionality attributed to it in launch publicity. Smart Prahari appears to be a genuine district hotspot and patrol-routing pilot, but its accuracy and effects remain undocumented. citeturn1search9turn1search10turn2search4turn2search6turn3search2turn3search11turn22search0turn4search9

The applicable constitutional standard is more demanding than “useful technology.” Articles 14, 19, and 21 require non-arbitrary state action, protection against unjustified chilling of speech and association, and a lawful, necessary and proportionate basis for privacy intrusions. The nine-judge privacy judgment in K.S. Puttaswamy recognized privacy as a fundamental right; subsequent Supreme Court formulations emphasize legality, a legitimate state aim, proportionality and safeguards against abuse. Predictive systems therefore require not merely executive authorization to buy software, but a sufficiently precise legal basis for collection, linkage, inference, watchlisting, retention, access and consequential use. citeturn21search0turn21search8turn21search12

As of August 2, 2026, India’s Digital Personal Data Protection Act, 2023 did not yet supply a fully operational rights framework for these systems. The November 13, 2025 commencement notification brought institutional and rulemaking provisions into force immediately, but scheduled core processing obligations, data-principal rights, major exemptions and enforcement provisions for eighteen months later—May 13, 2027. Even when operative, the Act contains special treatment for state processing, law-enforcement-related non-disclosure, and power to exempt notified state instrumentalities on national-security and public-order grounds. citeturn18view0turn18view1turn19view0

The result is an accountability mismatch. Police databases are becoming interoperable and analytically reusable faster than India is creating enforceable, police-specific rules for data quality, notice, correction, watchlist review, algorithmic testing, deletion, explanation, procurement transparency, and remedies. The proposed answer is not a general prohibition on analytics. It is a dedicated statutory framework that differentiates low-risk aggregate analysis from high-risk identification, association scoring, hotspot deployment, and person-based risk assessment; imposes pre-deployment necessity and equality review; protects acquitted and uncharged persons; and makes auditability a condition of procurement and continued operation.

National and state system inventory and verification table

The following inventory uses five evidentiary labels. Official technical means an operative government specification, privacy document, manual, audit, tender or detailed administrative order. Official publicity means a government announcement that confirms political adoption but not technical performance. Vendor marketing means a supplier or partner description that may accurately identify features but is not independent evidence. Journalistic documentation means reporting based on police or supplier interviews. Independent evaluation means an audit or research assessment not controlled by the deploying police or vendor. A system is classified as “deployed” only where evidence indicates real users or operational use, not merely a launch event.

Several national systems are analytically relevant even though they should not be called predictive policing. NAFIS creates a centralized criminal-fingerprint repository and permits cross-jurisdictional fingerprint searches. The Criminal Procedure (Identification) implementation project is establishing measurement-collection units across police districts, prisons and central agencies. The Digital Police Portal gives authorized users national searches for criminals or suspects and access to utilities such as Cri-MAC and the National Database on Sexual Offenders. These layers enlarge identity resolution and cross-case linking, which can improve investigations but also make an inaccurate or outdated entry propagate farther. citeturn15view3turn14search14

The inventory also shows why product names can mislead. “MARVEL” is a state-controlled institutional vehicle; “MahaCrimeOS” is an application associated with it. “ICJS 2.0” is an integration and modernization program, not an AI model version. “Trinetra 2.0” appears to identify a revised operational product, but its precise technical baseline remains undisclosed. “CMAPS” had phases and proposed increments rather than a published software release history. TSCOP, YAKSH, AI-SPS and Smart Prahari lack publicly accessible model cards, release notes or version-controlled evaluation reports. In high-stakes systems, the inability to identify which version produced a recommendation frustrates reproducibility, legal challenge and post-incident accountability.

A further pattern is the dependence of status claims on weak proxies. A launch ceremony proves announcement. A mobile app listing proves that software is distributed, not that every promised component operates. A supplier case study can establish commercial participation, but productivity figures may be anecdotal or selectively reported. A tender proves an intention to procure, not successful commissioning. A CAG audit or a technically detailed privacy policy is often more revealing than repeated references to “AI-powered policing” because it identifies data sources, controls, missing modules and actual workflows.

Data ecosystem, institutional flow, and legal status of persons

The national architecture begins at the police station or field unit. Information may enter through a written or electronic complaint, an FIR, a general or station diary entry, an emergency call, a missing-person report, a traffic or vehicle inquiry, an arrest or search record, a witness statement, a seizure memo, a photograph, fingerprints or other measurements, an investigation case diary, a final report, or a charge sheet. CCTNS was explicitly designed to computerize FIR, investigation and challan or charge-sheet processes, and its national portal supports searches involving accused persons, charge-sheeted persons, convicts, habitual offenders, proclaimed offenders, vehicles, property, missing persons and unidentified bodies. citeturn15view5

That first layer is not a neutral sample of “crime.” It is a record of incidents that became known to the police and were categorized, registered and entered. A complaint may not become an FIR; an FIR may identify no offender, one alleged offender or several; later investigation may add or remove names; a charge sheet may differ from the FIR; a court may discharge or acquit a person; an appellate court may reverse a conviction. Unless systems preserve each stage and prevent earlier allegations from being displayed as current truth, a search result can collapse accusation and adjudication into a single “criminal record.”

Emergency-response systems create a different stream. Calls to police control rooms may produce incident locations, caller numbers, timestamps, dispatch decisions and response records even where no FIR is registered. CMAPS’s audited dependence on Delhi’s PA-100 control-room data illustrates how a system’s apparent crime map can reflect calls or dispatch events rather than legally registered offenses. Combining emergency calls and FIRs without clear labels can double-count one event or treat unverified reports as equivalent to investigated offenses. citeturn11view1

Surveillance streams add much greater volume. City and district command centers may ingest fixed CCTV, police cameras, automatic number-plate recognition, traffic cameras, body- or vehicle-mounted feeds, live incident video and facial images. A facial-recognition engine generally converts an image into a template and searches a watchlist or gallery for similar templates. It does not determine guilt or predict future conduct. Any investigative meaning comes from the watchlist, the match threshold, image quality, corroboration and the officer’s subsequent decision.

Biometric and identification infrastructure is expanding independently of predictive analytics. The Criminal Procedure (Identification) Act, 2022 authorizes collection of broad “measurements” for specified categories of persons and provides for records to be retained in digital or electronic form for seventy-five years. For a person without a previous qualifying conviction who is released without trial, discharged or acquitted, destruction is contemplated after legal proceedings are exhausted, subject to the statutory exception permitting a court or magistrate to direct otherwise. The scale and duration of this repository make reliable outcome updating essential. citeturn21search2turn21search6

Under ICJS, information can move longitudinally across institutions. A simplified organizational flow is:

The MHA’s “one data once entry” objective can reduce repetitive transcription and inconsistent duplicates, but it also means that an error entered once may travel through police, prison, prosecution, forensic and court interfaces. Interoperability therefore changes the severity of error. A misspelled name in one station is local; a wrongly merged identity propagated nationally can affect arrest decisions, passport or employment verification, bail arguments, prison processing and future police encounters. citeturn15view3

The CAG’s 2025 Odisha CCTNS audit demonstrates that these are not theoretical concerns. Audit found gaps and duplications in general-diary serial numbers, the ability to manipulate local desktop dates and create back-dated entries, 1.22 lakh instances of duplicate diary serial numbers associated with back-dated insertion, and tens of thousands of records with malformed identifiers. It warned that weak application controls created risks to record integrity and investigation, including the possibility of deletion or manipulation without a clear trail. citeturn16view1turn16view2turn16view3

The legal and practical labels attached to people must remain distinct:

An accused person is someone against whom a criminal allegation has crystallized in a criminal process, commonly through naming in an FIR, arrest, a police report or proceedings before a court. The label is not a conviction. The person retains the presumption of innocence, and the evidentiary posture can change as investigation and trial progress.

A convicted person has been found guilty by a competent court, subject to appeal, suspension or reversal. A database should record the offense, court, date, sentence, appeal and current status rather than expose a timeless “convict” flag after a conviction is set aside.

A suspect is an investigative classification, not a general national adjudicated status. Suspicion may arise from description, proximity, association, intelligence or an officer’s hypothesis. Its evidentiary threshold is lower and its duration often undefined. A “suspect search” may be useful to investigators, but reuse for employment verification or future risk scoring would be particularly vulnerable to arbitrariness.

A habitual offender is not simply a person with multiple police contacts. The applicable meaning depends on legislation, rules and judicial interpretation. The Supreme Court has warned against hereditary, caste-linked or community-based understandings of habitual criminality, especially concerning denotified and nomadic communities. In Sukanya Shantha, the Court held caste-based prison classifications unconstitutional and directed that “habitual offender” treatment conform to valid legislation rather than stigmatizing community identity. citeturn23search7turn9search2turn9search4

A history-sheeter is ordinarily a person placed under continuing police surveillance through state police rules or standing orders, not a person convicted of a new offense by virtue of the history-sheet entry. In Amanatullah Khan v. Commissioner of Police, Delhi, the Supreme Court stressed that a history sheet is an internal police document, must not indiscriminately include innocent relatives, requires periodic senior review, and demands special care to avoid prejudicial treatment of people from socially and economically disadvantaged communities. citeturn23search3turn9search6

A gang member may be a person alleged or proceeded against under a state gang or organized-crime statute, included in a police gang chart or register, or merely identified by an analytic system as associated with known individuals. Those are not interchangeable. A co-residence, family relationship, telephone contact, social-media connection or presence in the same location may generate an analytic edge without proving participation in a criminal organization.

A person merely associated with another record is the broadest category. Witnesses, complainants, victims, relatives, visitors, phone subscribers, vehicle owners, landlords and people appearing in CCTV can all become linked to a case. Association graphs visually encourage inference: a dense cluster looks meaningful even when its edges represent fundamentally different events. Systems must label edge type, date, source, reliability and direction, and must prevent an association from automatically becoming a risk feature.

The danger of status collapse is heightened by national antecedent searches. The official CCTNS description states that the portal holds information concerning accused and convicted persons and can be used for police and employment-related antecedent verification. Without outcome reconciliation and context-sensitive display, an old acquittal, mistaken identity or unproven accusation may affect a person well beyond the original investigation. citeturn15view5

External datasets require separate legal authority. CCTV collection, telecom interception, call-detail acquisition, open-source social-media monitoring, driver and vehicle databases, and private vendor data are not made lawful merely because CCTNS or a command center can technically ingest them. Each source has its own collection authority, retention logic and reliability. ICJS authorization to share records among criminal-justice pillars does not by itself authorize unlimited enrichment from every available government or commercial database.

CMAPS, CCTNS and ICJS as the foundation of Indian police analytics

CMAPS is the clearest case study of how predictive-policing language can outrun implementation. Delhi Police and the Advanced Data Processing Research Institute of ISRO entered a memorandum in December 2015. The resulting web application was placed at police headquarters and made accessible through browsers to police stations and district units. Its intended role was described as decision support: plotting crime, identifying concentration by region, frequency and type, and eventually incorporating security and open-source intelligence functions. citeturn11view1

The project plan contained four phases. The first covered crime analytics. The second contemplated a security module with a situation database and threat rating. The third proposed a news module capable of geotagging and clustering reports. The fourth proposed social-media or website extraction, social-network analysis and text annotation. In 2017, Delhi Police and ADRIN also discussed criminal profiling and a mobile CMAPS application, with expected completion in early 2018. citeturn11view1turn11view2

The CAG found that this planned architecture had not materialized. As of September 2019, CMAPS mainly fetched data from PA-100, while comprehensive integration with CCTNS had not been achieved. The security, news and social-media modules were not operational as designed. The additional criminal-profiling and mobile work had made no progress, and the audit record showed no relevant communication between Delhi Police and ADRIN after March 2018. The CAG characterized project monitoring as inadequate and found diminished institutional interest. citeturn11view1turn11view2

This matters technically because crime mapping is not synonymous with prediction. A map showing the previous month’s robberies is descriptive. A density surface computed from past crimes can be exploratory. A forecast must specify a future interval, geographic unit, predicted outcome and evaluation criterion. It must be tested out of sample or prospectively against an appropriate baseline, such as historical frequency or officer judgment. No such CMAPS performance report was identified in the CAG audit.

CMAPS also illustrates dependence on hidden infrastructure. Academic research on Delhi’s predictive-policing data practices reported that geocoding and mapping workflows were affected when an ArcGIS license expired and that police data production involved substantial manual and organizational choices. The researchers emphasized representation, measurement and historical bias: a model can only learn from offenses and locations rendered visible through police recording practices. citeturn0search12turn0search15

A predictive crime map can be mathematically accurate with respect to recorded events and still misdescribe victimization. Suppose Police Station A registers complaints promptly while Police Station B discourages registration. A model will infer greater risk in A. If patrols then increase there, officers will observe and record more street offenses, creating a feedback loop. Conversely, hidden domestic violence, labor exploitation, caste violence or cyber fraud may remain geographically underrepresented because reporting and detection do not arise through visible patrol encounters.

The 2026 Delhi hotspot research is a meaningful development because it describes a method, not just a brand. The researchers used a nonparametric spatiotemporal kernel-density approach, released algorithmic material and reported use in allocating patrol vehicles. A kernel method gives greater weight to crimes close in space and time while allowing influence to decay. This is future-oriented place prediction, although the publication’s masked data limits external reconstruction and the deployment’s institutional relationship to legacy CMAPS remains unclear. citeturn0search21turn0academia29

The appropriate evaluation of such a system is not merely “what percentage of future crimes fell inside predicted hotspots?” A model can obtain a high hit rate by declaring a large part of the city high-risk. Evaluation should report the predicted area, precision or predictive-accuracy index, temporal stability, comparison with simple historical baselines, displacement to adjacent areas, changes in reporting, patrol time imposed on neighborhoods, stops and searches generated, and whether crime outcomes improved without disproportionate burdens.

CCTNS creates the substrate on which more advanced analysis becomes possible. The 2018 government status brief described its objectives as computerizing FIRs, investigations and charge sheets; providing pan-India crime and criminal searches; producing state and national reports; offering citizen services; and sharing data with courts, prisons, prosecution, forensics and fingerprint systems. It also stated that records may be in English or regional languages and that police users can search categories such as accused, charge-sheeted, convicted, habitual and proclaimed offenders. citeturn15view5

The official architecture is federated in organizational terms even where data are replicated or made nationally searchable. States and Union Territories retain their police administration and may customize workflows, engage different system integrators and vary in digitization quality. The central portal aggregates search and reporting. This helps explain why asking for “the CCTNS version” has no simple answer: CCTNS is a program, network, common-application framework and set of state implementations rather than one uniform application binary.

ICJS extends the potential analytic unit from “police record” to “criminal-justice trajectory.” A police allegation can be compared with prosecution action, forensic results, custody information and a court outcome. Properly designed, that could correct police data—for example, by attaching an acquittal or discharge. Poorly designed, it can produce a richer dossier in which every stage remains searchable indefinitely without contextual hierarchy. citeturn15view3

ICJS 2.0’s official goals include reducing duplicate entry, improving data quality, enabling analytics and AI/ML, and reducing paper dependence. Those are aspirations, not a quality certification. The Odisha audit shows that system-generated identifiers, timestamps and audit trails can fail even within a mature state deployment. Integration should therefore be preceded by source-level validation rather than treating central linkage as a cure for local data problems. citeturn15view3turn16view1turn16view2

CCTNS also changes the stakes of ordinary police classifications. A handwritten local history sheet might once have been known to a limited circle. A digitally indexed designation can become discoverable across jurisdictions and incorporated into a vendor’s risk or network model. A field marked “habitual offender” may then function as a high-weight feature even where the underlying classification has not been judicially reviewed. The system’s capacity to search that category is officially documented; the governance of how it may be used in scoring is not. citeturn15view5

CCTNS and ICJS should therefore be evaluated along two axes. The first is administrative performance: uptime, completeness, timely entry, deduplication, user training, security and successful exchange. The second is rights-sensitive data governance: lawful collection, role-based access, source and status labeling, correction, retention, audit logs, downstream-use controls and remedies. A system can score well on the first while failing the second.

State and district case studies, vendors, and procurement

Telangana is India’s most developed example of converged command-and-control policing, but also one of the least transparent at the level required for independent algorithmic assessment. TSCOP has been described as a police mobile platform allowing officers to access operational and criminal information in the field, perform checks and use face-recognition-related capabilities. Hyderabad’s broader command-center model connects extensive CCTV and operational monitoring. Official district descriptions confirm real-time monitoring, response and coordination functions, while investigative and international reporting has described a much larger city surveillance environment. citeturn1search4turn1search10turn1search16

Public discussion sometimes calls this infrastructure predictive because it aggregates large volumes of data and supports rapid alerts. Yet neither aggregation nor alerting proves future-crime prediction. A face match against a wanted-person list is retrospective identification. An alarm when a listed vehicle enters an area is rules-based detection. A dashboard showing repeat-offense locations is descriptive. A model estimating that a person will offend next week would be person-based prediction and requires much stronger evidence.

The 2024 reporting about TSCOP data being offered or exposed online is relevant not only as a cybersecurity incident but as evidence of the platform’s data sensitivity. Reports described data involving police officers and persons in criminal or suspect-related records, including facial images. Publicly available information did not allow this investigation to determine the complete scope, root cause, number of affected people or remedial audit. citeturn1search9turn1search13turn1search17

A mature governance response would publish a redacted incident report, affected data classes, exposure period, access-log findings, notification policy, credential and architecture changes, vendor responsibilities and an independent security assessment. Merely restoring a service does not resolve risks from copied data, particularly biometric images that cannot be reissued like passwords.

Maharashtra’s MARVEL model is institutionally distinctive. The Home Department’s official record verifies that the state approved a special-purpose vehicle involving the Government of Maharashtra, IIM Nagpur and Pinaka Technologies, with government share capital. This provides a formal route for co-developing policing technology but also complicates procurement transparency, intellectual-property ownership, subcontracting and public-law accountability. citeturn2search4

MahaCrimeOS is publicly described as an AI investigation platform developed through MARVEL with Microsoft and CyberEye participation. The Microsoft case study says it extracts and organizes multilingual case files, drafts notices, helps analyze call records and proposes investigative steps using Azure OpenAI and Microsoft Foundry technologies. These are generative-AI and knowledge-management functions. They may save time, but a partner case study cannot establish accuracy, legal reliability or whether officers over-rely on generated suggestions. citeturn2search6

The distinction between an investigative copilot and predictive policing is important. A system that summarizes a completed FIR or suggests procedural steps is not predicting criminal conduct. It may nevertheless create serious risks: hallucinated facts, omitted exculpatory material, mistranslation, loss of source provenance, disclosure of case files to cloud infrastructure, and automation bias. Every generated assertion used in a notice, remand application or charge sheet should link to the exact source passage and be affirmatively verified by the investigating officer.

The March 2026 Maharashtra resolution supporting a ₹2 crore AI-powered investigation-platform pilot is stronger evidence than press statements because it confirms administrative funding and a limited operational scope. Later reporting of use in Nagpur city or rural police stations and proposed expansion to roughly 1,100 stations should be treated as reporting on rollout plans until completion, acceptance testing and statewide operational orders are published. citeturn2search1turn2news31

Uttar Pradesh’s Trinetra and YAKSH illustrate vendor-mediated continuity. Trinetra was launched in 2018 with publicity describing a large database of offender photographs and AI-assisted recognition. By 2024, Trinetra 2.0 was reported to use Staqu’s CrimeGPT technology and a larger digitized record base. Publicly reported capabilities center on face matching, audio or voice-related cues and natural-language search. citeturn3search17turn3search11

YAKSH appears to extend this model into multimodal criminal intelligence. Reports attribute to it searches across FIRs, case files, CCTV, faces and voices, combined with gang or association analysis and beat-level verification. Some reporting refers to “top-10” offender scores or lists. The central unanswered questions are what that score represents, which features determine it, who can alter it, how frequently it is reviewed, and what consequences follow. citeturn3search2turn3search20turn3search13

A ranking can be predictive, descriptive or administrative. Counting pending cases is descriptive. Ranking people by officer-entered “importance” is administrative. Estimating reoffending probability is predictive. Public material does not establish that YAKSH’s ranking is a validated probability estimate. It should not be described as a future-offending model without disclosure of its target and methodology.

The vendor relationship matters. Staqu’s JARVIS One and CrimeGPT descriptions offer plausible evidence of underlying software functions, but supplier claims are not independent validation. Procurement documents should identify whether the state licensed an off-the-shelf platform, commissioned customization, transferred intellectual property, or relies on a continuing hosted service. They should also specify model and data ownership, subcontractors, cloud location, breach liability, audit access, post-contract data return and the ability to export logs in usable form.

AI-SPS provides a rare example in which an official privacy page narrows a publicity claim. The Ghazipur district page states that Igile Technologies customized its ADVAS solution and identifies attendance, facial images, demographic data, device identifiers, location, GPS special-duty reporting, patrol tracking, “Milaan validations,” real-time monitoring and management analytics. It also states that personal data may be used for user categorization or classification and retained for at least three years after account deletion for legal compliance. citeturn22search0

Those functions can affect police employees’ privacy and labor conditions, but they do not establish crime prediction. Launch journalism that labels the whole platform predictive may be referring to planned crime mapping or to a separate dashboard not described in the privacy policy. The responsible conclusion is that operational workforce and patrol-management capabilities are verified, while crime-prediction functionality remains unverified.

Smart Prahari is closer to classic place-based prediction. Reports say Washim police used several years of past crime data to identify patterns by location, time and offense type and to recommend patrol routes through a mobile application. The described method could produce forecasts even if technically simple; sophisticated machine learning is not required for predictive policing. citeturn4search9turn4search16

Its locally developed, reportedly open-source character could reduce vendor lock-in and permit inspection. It does not remove the need for governance. A zero-cost model can still intensify unequal patrols, and “open source” should mean that the exact deployed code, dependencies and model parameters are inspectable—not merely that developers used open-source libraries.

Public procurement is one of the most effective intervention points because safeguards can be made contractual before deployment. For central government entities, General Financial Rules Rule 144 and current procurement manuals emphasize transparency, fairness, competition, economy, efficiency and accountability; common goods and services available through the Government e-Marketplace are generally required to be procured through GeM. States have their own financial rules, tender portals and procurement arrangements, and police technology may be acquired through open tender, limited tender, proprietary justification, nomination, grant, memorandum or a special-purpose vehicle. citeturn21search3turn21search7turn21search27turn21search31

Ordinary information-technology tenders are inadequate for algorithmic policing. A specification focused on server capacity, number of cameras, response time and “AI features” does not define lawful use. A proper tender must include the decision being supported, excluded uses, training-data provenance, minimum performance by operational context, subgroup testing, false-match reporting, audit-log format, data-retention schedule, human-review procedure, independent test access, breach notification, model-update controls, source-code or escrow rights, and termination obligations.

Acceptance testing should use representative Indian operational data, including regional scripts, transliteration variants, low-quality CCTV, multiple people with similar names, old photographs, masks, head coverings and changing addresses. Testing by the vendor on its own benchmark is insufficient. For generative tools, evaluation should include fabricated facts, incorrect statutory citations, mistranslation, omission of exculpatory material and prompt-injection or malicious-document attacks.

Contracts should prohibit undisclosed reuse of police data for general model training. They should identify every processor and cloud region, require encryption and strong authentication, and give the police—and an independent auditor—the ability to examine access logs. The public should receive a redacted contract and impact assessment; narrow redactions can protect exploitable security details and legitimate trade secrets without concealing cost, purpose, performance obligations, retention or accountability.

The recurring procurement risk is responsibility diffusion. Police may say the vendor built the algorithm; the vendor may say police selected the data and threshold; a cloud provider may say it only supplied infrastructure; an institutional partner may characterize the work as research. The deploying public authority must remain legally responsible for the decision system as a whole. Contractual allocation cannot displace constitutional accountability.

Accuracy, data quality, caste and minority impact

Predictive policing fails in at least four distinguishable ways. Representation error occurs when recorded data omit much of the underlying phenomenon. Measurement error occurs when variables inaccurately capture what they purport to represent. label error occurs when the target—for example “offender,” “gang member” or “high-risk location”—is inconsistently assigned. deployment error occurs when an otherwise reasonable model is used for a different population, purpose or threshold than the one for which it was tested.

Indian recorded-crime data are especially sensitive to representation error because policing is complaint-driven but registration is institutionally mediated. Victims may not report because of fear, social pressure, distance, language, economic dependence or distrust. Police may classify a report as a non-cognizable entry, petition or station diary item rather than an FIR. Some offenses are detected primarily through proactive police activity, so recorded prevalence closely follows enforcement intensity. The Delhi data research emphasizes that predictive outputs inherit choices made during data production rather than representing crime independently of police practice. citeturn0search12turn0search15

A model trained on arrests is not learning who committed crimes; it is learning who was arrested. A model trained on charge sheets is learning whom police and prosecutors proceeded against. A model trained on convictions has a more adjudicated target but introduces years of delay, plea and trial differences, legal-representation inequalities and offense-specific attrition. None is automatically an unbiased proxy for future offending.

Inconsistent FIR registration can distort geography and group comparisons. A district that improves registration may appear to experience a crime surge. A police station that records every emergency call may look more dangerous than one that records only FIRs. A neighborhood with visible street activity may generate more stop, narcotics and public-order records than a gated neighborhood where comparable conduct is less observable.

CCTNS audit findings show how mundane database flaws can become model features. Duplicate or malformed general-diary identifiers, manipulable timestamps and absent backend logs affect sequence, frequency and recency—precisely the variables commonly used in risk scoring and hotspot analysis. If back-dated entries are treated as contemporaneous, temporal models can infer nonexistent clusters or repeat activity. citeturn16view1turn16view2turn16view3

Identity resolution creates another class of errors. Indian names may appear in multiple scripts and transliterations; initials, patronymics, caste or community names and inconsistent surname usage complicate matching. Addresses may be informal, incomplete or frequently changing. Migrant workers may share rooms or phone numbers. Police systems may contain aliases that are spelling variants rather than deliberate identities. The official CCTNS description confirms that records may be in English or regional languages and that national searches operate across person categories, making cross-language matching a core operational problem. citeturn15view5

Entity-resolution systems normally trade false positives against false negatives. Broad fuzzy matching finds more true matches but also merges unrelated people. Strict matching misses records entered under another spelling. Biometrics can help but do not eliminate the problem: old or low-quality photographs, partial faces and large watchlists can produce candidate lists requiring expert review.

NIST’s large-scale evaluations demonstrate that face-recognition performance depends on the algorithm, application, demographic group and image quality. Its demographic studies found substantial differentials in many algorithms, particularly for false positives, while noting that effects vary by developer and dataset. NIST also emphasizes that one-to-many identification over large galleries differs from simple one-to-one verification. These results do not prove that a particular Indian deployment is biased; they show why each deployed algorithm, threshold, camera environment and watchlist must be tested locally. citeturn23search0turn23search4turn23search8turn23search11

NITI Aayog’s responsible-AI materials similarly identify safety, reliability, equality, inclusion, non-discrimination, privacy, security, transparency and accountability as governing principles, and specifically use facial recognition as a high-risk case study. These documents are policy guidance rather than binding police regulation. citeturn23search1turn23search5turn23search9

The distinction between discriminatory intent, design and impact is essential. No finding in this report establishes that a named system was created with an intention to discriminate against a caste, religion, tribe or political group. Design discrimination may arise, however, where a system treats police contact, residence, family association or history-sheet status as a risk feature despite their unequal social distribution. Impact discrimination may arise when apparently neutral hotspot or watchlist rules concentrate surveillance and intervention on protected or disadvantaged groups without adequate justification.

Caste can enter a model directly, through an explicit field, or indirectly through surname, locality, occupation, police classification, social network or the historical distribution of enforcement. Religion may be inferred through names, institutions, neighborhoods or event participation. Scheduled Tribe or Adivasi status can correlate with geography and policing under forest, excise or public-order laws. Migrant status can correlate with unstable addresses and failed verification. Poverty can correlate with public-space visibility and residence in heavily surveilled localities.

Common Cause and Lokniti’s 2023 policing report found that government CCTV was reported as substantially more prevalent in slums and poor localities than in higher-income areas and that trust in surveillance and police tended to be lower among poorer respondents, Adivasis, Dalits and Muslims. These are survey and observational findings, not proof that every CCTV was installed with discriminatory intent, but they demonstrate unequal exposure that an analytics system may amplify. citeturn23search2turn23search6turn23search13

Earlier police-personnel survey research found significant proportions expressing stereotypes about poor people, slum residents and street vendors being naturally prone to crime. The 2025 report likewise documents police perceptions concerning access to justice for poor people, slum dwellers, sex workers, migrants and denotified or nomadic communities. Survey answers cannot be imputed to every officer or system, but they matter because discretionary labels and manual overrides are produced within institutions, not outside them. citeturn23search10turn23search28

Denotified, nomadic and semi-nomadic communities face a historically specific risk. Colonial “criminal tribe” classifications were formally repealed, yet habitual-offender laws, police registers and social stereotypes have continued to raise constitutional concern. The Supreme Court’s 2024 decisions reject caste-embedded administrative classifications and caution police against arbitrary inclusion and surveillance of disadvantaged people. A model that takes “habitual offender,” “history sheet” or inherited association as ground truth may reproduce a legally discredited classification without ever including caste as an explicit variable. citeturn23search3turn23search7turn9search11

Religious minorities can be affected through public-order and national-security datasets, communal-incident records, protest surveillance and neighborhood-level enforcement. A location model may never receive a religion field but still target a religiously concentrated locality because past deployments produced more recorded incidents there. To determine impact, auditors need patrol, stop, search, questioning, arrest and false-alert rates—not only model accuracy.

Political protesters and civil-society groups raise a distinct Article 19 issue. News extraction, social-network analysis and geotagging were among CMAPS’s planned modules, although the CAG found them unimplemented at the audit date. Such functions can map lawful association and expression. Even without arrests, knowledge that participation is being scored or retained may chill speech, assembly and organization. citeturn11view1turn11view2

Neighborhood hotspot systems create feedback through deployment. More patrol time produces more observation, field inquiries and police-initiated detections. Those new events enter the database and can validate the original hotspot. A responsible evaluation should therefore distinguish citizen-reported crimes from police-initiated records and should test models using victimization or emergency-call information where appropriate, while recognizing those sources’ own biases.

Person-risk models create a stronger feedback loop. If a person is scored high-risk, officers may visit, question or monitor the person more often. Those contacts become new records, which increase the score. Association with other monitored people adds more edges. Without a mechanism for decay, correction and independent review, risk becomes self-confirming.

Accuracy must be measured at the point of use. A face engine may return the correct person within its top ten candidates 95 percent of the time under test conditions, yet frontline officers may treat the first candidate as a confirmed match. A hotspot may contain many crimes but cover so much land that random patrol would be nearly as effective. A language model may summarize most files correctly but introduce one fabricated fact into a remand request. Human review is not a complete safeguard unless reviewers receive source material, uncertainty, alternatives, time and authority to reject the output.

No public documentation located for CMAPS, TSCOP, Trinetra, YAKSH, MARVEL products, AI-SPS or Smart Prahari reported complete operational confusion matrices, calibration or subgroup testing. Absence of public evidence is not proof of poor accuracy. It means claims of accuracy, bias mitigation or effectiveness cannot be independently verified.

Constitutional, statutory, surveillance, transparency, and remedial framework

Article 14 prohibits arbitrary state action and unequal treatment without constitutionally adequate justification. In an algorithmic-policing setting, arbitrariness can arise from undefined criteria, irrational proxies, inconsistent data, unreviewable watchlist placement, secret thresholds or different treatment of similarly situated persons. Equality review must examine effects and administrative structure, not only whether a model explicitly includes caste or religion.

Article 19 protects speech, peaceful assembly, association and movement, subject to constitutionally enumerated restrictions. Police monitoring of protests, political networks, religious gatherings, journalists or online expression can chill participation before any prosecution occurs. A system’s mere existence, if combined with opaque watchlisting, can alter behavior. Restrictions must have a legal basis and satisfy necessity and proportionality rather than rest on a general invocation of public order. The Supreme Court’s communications-restriction jurisprudence has emphasized necessity, proportionality, publication and procedural safeguards. citeturn21search12turn6search3

Article 21 protects life and personal liberty and incorporates privacy, dignity, autonomy and fair procedure. Puttaswamy rejected the idea that privacy is lost whenever a person enters public space or shares information for a limited purpose. The constitutional inquiry asks whether intrusion is backed by law, pursues a legitimate aim, is necessary and proportionate, and is accompanied by safeguards against abuse. citeturn21search0turn21search8

Applied to predictive policing, “legality” requires more than a police department’s general duty to prevent crime. The law should identify what data may be collected and linked, whose data may be retained, what kinds of inference are permitted, who may access outputs, and what consequences may follow. A broad police act or executive standing order may authorize recordkeeping but not necessarily continuous biometric search, social-network inference or algorithmic risk scoring.

“Legitimate aim” will usually be crime prevention, investigation, public safety or efficient deployment. The difficult questions are necessity and proportionality. The state must show why less intrusive alternatives—better emergency response, improved FIR registration, conventional investigation, lighting, victim services or targeted non-personal analysis—would not adequately achieve the aim. The more consequential the output, the stronger the justification and safeguards required.

Predictive output alone should never supply reasonable grounds for arrest, search, detention or coercive questioning. At most it may direct attention for lawful preliminary verification. Otherwise, a statistical association is converted into individualized suspicion without evidence of the person’s conduct.

The Digital Personal Data Protection Act creates a general framework for digital personal data, but its temporal and substantive limits are significant. The official consolidated text records that most core processing provisions, obligations, rights, enforcement powers and exemptions commence eighteen months after November 13, 2025. Accordingly, on August 2, 2026, rights of access, correction and erasure under sections 11 and 12 and the principal processing obligations had not yet taken effect. citeturn18view0turn18view1

When operative, section 11’s access right excludes some information about sharing with another legally authorized data fiduciary where the sharing responds to a written request for prevention, detection, investigation, prosecution or punishment of offenses. Section 12 provides correction, completion, updating and erasure rights but allows legally required retention. Those provisions may therefore offer less visibility into police-to-police or police-to-agency sharing than individuals need to challenge an erroneous risk record. citeturn18view0

Section 17 permits processing-related exemptions and authorizes the Central Government to notify state instrumentalities as exempt in interests including sovereignty, state security, friendly relations, public order and prevention of incitement to related cognizable offenses. The Act also relaxes certain erasure and correction rules for state processing. These provisions do not erase constitutional review, but they reduce the reliability of the DPDP Act as the sole safeguard for policing systems. citeturn19view0

The final DPDP Rules were notified in November 2025 with staged commencement. They include security, breach and governance requirements and, for significant data fiduciaries, contemplate data-protection impact assessment, audit and attention to algorithmic risk once the relevant provisions become operative. Police-specific duties, watchlist correction and judicial authorization are not supplied by general rules alone. citeturn18view2turn11view3turn11view4

The Criminal Procedure (Identification) Act, 2022 is directly relevant because it creates a legal route for large-scale biometric and biological measurement retention and central NCRB administration. Its seventy-five-year retention period is unusually long relative to many investigative needs. The destruction provision for eligible acquitted, discharged or untried persons will be meaningful only if court outcomes reliably reach the biometric repository, duplicate identities are reconciled, and destruction is technically verified across replicas and downstream systems. citeturn21search2turn15view3

The Bharatiya Nagarik Suraksha Sanhita, 2023 deepens digital criminal procedure by accommodating electronic communications, records and proceedings and providing that trials and proceedings may be held in electronic mode. These reforms can improve traceability but also normalize machine-generated records in investigation and adjudication. Digital provenance, hash verification, chain of custody and disclosure of transformations become essential when an AI tool extracts, translates, enhances or summarizes evidence. citeturn18view5turn21search10

The Bharatiya Sakshya Adhiniyam’s treatment of electronic and digital records does not make an algorithmic conclusion automatically reliable. The underlying record, method of production, integrity, relevance and opportunity to challenge remain legally important. A face-match score or generated summary should be treated as an investigative lead unless properly validated and proved; it should not be presented as self-authenticating fact.

Communications interception is governed separately. The Telecommunications Act, 2023 and the Telecommunications (Procedures and Safeguards for Lawful Interception of Messages) Rules, 2024 provide an authorization and review framework for interception through telecom systems. Information-technology interception and monitoring also operate under the Information Technology Act and associated rules. These frameworks do not automatically govern every police database query, CCTV analysis or historical call-detail request. citeturn17search0turn17search20turn17search28

This fragmentation produces gaps. Live interception may require senior authorization and review, while an analytic platform may reconstruct months of movement and association from stored records without an equivalent independent authorization. Constitutional impact depends on the capability and use, not on whether the state calls it “interception,” “analytics” or “dashboard access.”

Police acts, state police rules, standing orders and manuals govern history sheets, surveillance registers, station diaries, preventive action and intelligence. Because police and public order are primarily state subjects, legal authority and terminology vary across jurisdictions. Digitization does not standardize these underlying rules. A national search can therefore combine categories created under different thresholds.

The Supreme Court’s history-sheet jurisprudence provides minimum procedural guidance. Amanatullah Khan requires internal confidentiality, exclusion of uninvolved relatives, periodic senior review and sensitivity to the disproportionate effect of police surveillance on disadvantaged groups. A digital history-sheet or risk system should at minimum implement review dates, automatic expiry pending renewal, reasons, source records, notification where it would not compromise an active investigation, and a correction mechanism. citeturn23search3turn9search6

The Right to Information Act, 2005 is presently the most practical route for obtaining project approvals, tenders, contracts, minutes, file notings, standard operating procedures, data dictionaries, impact assessments, audit reports and aggregate performance information. The Act defines information broadly to include electronic data, contracts, reports, models and records; section 4 requires proactive publication of functions, decision processes, norms, manuals, document categories, budgets and material facts behind important policies. citeturn20view1

Police authorities may invoke section 8 exemptions for national security, confidential law-enforcement sources, ongoing investigation, personal information, commercial confidence or intellectual property. Those exemptions are not a blanket bar. Section 8(2) permits disclosure where public interest outweighs protected harm, and section 10 requires severance of exempt portions. Aggregate error rates, model purposes, contract values, retention rules and completed audit findings ordinarily require a more specific harm analysis than a generic statement that the system concerns policing. citeturn19view4

Section 24 excludes listed intelligence and security organizations, subject to corruption and human-rights exceptions, and permits states to notify their own exempt organizations. Ordinary police departments are not automatically exempt merely because their functions concern law enforcement, although particular states may have notified units and particular records may fall under section 8. RTI appeals place the burden of justifying denial on the public information officer, and commissions can inspect records that cannot be withheld from them during inquiry. citeturn20view0

A useful RTI request should avoid seeking personal case records and instead ask for the administrative architecture: sanction orders; current version and release date; feature and data-source list; field definitions; standard operating procedures; user roles; retention schedule; number and type of searches; procurement method; bidders; evaluation criteria; contract and amendments; security and privacy assessments; accuracy testing; false alerts; complaints; audit findings; and decommissioning decisions. Requests should separately identify the police department, state crime records bureau, state IT corporation, NCRB, home department and any special-purpose entity because each may hold different records.

Available remedies remain dispersed. A person affected by arrest, surveillance, history-sheeting or consequential reliance on erroneous data may seek constitutional relief under Articles 32 or 226, challenge evidence or procedure in the criminal case, seek correction through the police hierarchy, use human-rights and minority or caste commissions where jurisdiction exists, pursue RTI appeals, or complain to the data-protection framework once applicable. Procurement decisions and expenditure can be examined through departmental vigilance, tender challenges, legislative committees and CAG audit. None presently constitutes a dedicated, rapid appeal against algorithmic watchlisting or risk scoring.

A meaningful remedy must reach the source and every replica. Correcting a police-station entry is ineffective if a national index, vendor cache, biometric watchlist and downloaded intelligence report retain the old record. The responsible authority should be required to issue a digitally traceable correction or deletion instruction to all recipients and certify completion.

International comparisons and the limits of analogy

The FBI’s Threat Screening Center, historically known as the Terrorist Screening Center, is useful as a comparison in watchlist governance, not as an equivalent predictive-policing system. The TSC consolidates federal terrorism-screening information and supports identity checks using biographic and biometric identifiers. Inclusion is based on terrorism-related nomination standards described as requiring articulable intelligence or reasonable suspicion; the system facilitates screening rather than calculating general future-crime probability. citeturn12search0turn12search8

U.S. oversight experience nevertheless illustrates familiar problems: nomination quality, duplicate or incomplete identity records, dissemination rules, redress and the consequences of a false match. Justice Department Inspector General work has examined whether watchlist records were properly and accurately handled. The lesson for India is that even a purpose-limited watchlist requires defined nomination criteria, quality assurance, review, correction and independent audit. citeturn12search4

A comparison with Trinetra, NAFIS or CCTNS search should therefore focus on function. All may assist identity resolution across records. The TSC is terrorism-specific and federally structured; CCTNS spans ordinary crime and state policing; Trinetra includes facial search over a state-oriented gallery. None should be called predictive solely because a match affects future screening.

The United Kingdom’s National Data Analytics Solution, or NDAS, is the closest institutional comparison to police analytics. Early government descriptions discussed analyzing police-held data to assess risk of offending or victimization and emphasized that analytics would support rather than replace officers. Later privacy and governance documents describe a centralized, scalable capability with staged use cases, including modern-slavery analysis. citeturn12search11turn12search5

NDAS has also generated civil-liberties debate, particularly around person-based risk and opaque data use. Its value as a comparison lies partly in documentation: use-case staging, privacy notices, data-protection impact assessment material and local governance records provide a clearer public trail than is available for most Indian systems. Publication does not prove that a system is lawful or unbiased, but it makes scrutiny possible. citeturn12search17turn12search2turn12search9

India’s ICJS has broader institutional interoperability than a single analytic use case, while systems such as YAKSH resemble an analyst interface layered over police records. A responsible Indian framework could adopt the British practice of authorizing and documenting use cases separately instead of granting a platform a general mandate to apply “AI” to all available data.

China’s Police Cloud and Xinjiang’s Integrated Joint Operations Platform occupy a fundamentally different political and legal setting. Human Rights Watch has documented large-scale aggregation of personal and administrative data, algorithmic flagging of activists, dissidents and ethnic minorities, and the use of IJOP flags in coercive investigation and detention. citeturn13search0turn13search2turn13search5

No equivalence should be implied. India has a written Constitution with enforceable fundamental rights, federal and state institutions, electoral competition, courts, an RTI regime, an independent CAG and a public sphere capable of challenging police action. Telangana’s command centers, CCTNS or CMAPS should not be described as “the same as” China’s Xinjiang system.

The comparison is nevertheless legitimate at the level of technical affordances. Large-scale data integration, identity resolution, CCTV, location history, association graphs, watchlists and automated alerts can be used under very different political systems. Technical architecture does not determine constitutional legitimacy. Law, purpose limits, independent authorization, transparency, remedies and actual consequences do.

Across the three comparisons, a functional taxonomy is more useful than national labels:

The central comparative lesson is not that India should copy one jurisdiction. It is that integration and inference should be governed by use, consequence and risk. A terrorism watchlist, a patrol map, a face search and a person-risk score require different legal thresholds and review structures.

Minimum safeguards, proposed statutory framework, and final assessment

India should enact a dedicated Police Data, Surveillance and Algorithmic Accountability Act or an equivalent chapter within comprehensive surveillance reform. General data-protection law is not sufficient because policing involves involuntary collection, secrecy, coercive consequences and records about third parties. The statute should apply to police, state crime records bureaus, NCRB, command centers, special-purpose vehicles, contractors and any government body processing data for law-enforcement analytics.

The statute should begin with precise definitions. “Predictive policing” should mean computational analysis intended to estimate the future location, time or type of offending, victimization or public disorder, or the future conduct or risk attributed to a person or group. “Biometric identification,” “watchlist,” “entity resolution,” “association analysis,” “automated decision system,” “high-impact decision” and “generative investigative system” should be separately defined. This would prevent a facial-recognition purchase from evading controls because it is called “search,” or a risk score from evading controls because it is called “decision support.”

Systems should be placed in risk tiers. Aggregate, anonymized resource planning would occupy a lower tier. Place-based hotspot forecasting would require documented validation and impact monitoring. Facial recognition, persistent tracking and association graphs would require a specific legal basis, necessity review and strong access controls. Person-based future-risk scoring should be presumptively prohibited for coercive decisions and allowed, if at all, only under narrowly defined statutory authorization and independent supervision.

Before deployment, the police should publish an Algorithmic and Surveillance Impact Assessment. It should state the problem, alternatives considered, legal authority, data sources, affected groups, model target, features, protected-attribute and proxy analysis, accuracy, subgroup results, retention, access, human review, security, procurement and expected consequences. Operationally sensitive details could be placed in a confidential annex reviewed by an independent authority; secrecy should not swallow the public document.

A necessity and proportionality assessment should be mandatory. Procurement convenience or vendor availability should not establish necessity. The police should demonstrate that the system addresses a defined problem, that less intrusive means are inadequate, that the expected benefit is measurable, and that the scale and duration are proportionate.

Data quality rules should require source and status labeling. Every person-related record should identify whether it concerns a complainant, victim, witness, suspect, accused, charge-sheeted person, discharged person, acquitted person, convicted person, history-sheeter, habitual offender, proclaimed offender or mere associate. Systems must display the legal source, date, jurisdiction and current outcome and must prevent lower-status labels from being represented as convictions.

Acquittal and discharge updates should be automatic through ICJS, subject to human verification where identities conflict. An acquitted person’s data should not remain in ordinary predictive features merely because a historic FIR remains legally archived. Retention for investigative audit and use for future risk are distinct purposes.

History sheets, gang designations and habitual-offender entries should expire unless affirmatively renewed by a senior officer on recorded, reviewable grounds. Inclusion based solely on caste, tribe, family relation, neighborhood, poverty, political participation or uncorroborated association should be prohibited. The Amanatullah Khan and Sukanya Shantha principles should be codified for digital systems. citeturn23search3turn23search7

The law should prohibit use of protected attributes and close proxies unless strictly necessary for auditing discrimination or addressing a specific legally recognized victimization pattern. Removing caste or religion from the input is not enough; auditors must test surnames, locality, occupation, language, network structure and police labels for proxy effects.

Every model must be evaluated prospectively against a meaningful baseline. Place-based systems should report forecast area, precision, recall, displacement and patrol burden. Face systems should report false matches and non-matches by relevant demographics, image quality, watchlist size and camera conditions. Person scores should report calibration, false-positive rates and differential consequences, although the preferred rule is not to use such scores for coercive action.

No adverse action should be based solely or predominantly on an algorithmic output. “Human in the loop” must mean an identified officer reviews the underlying evidence, records independent reasons and can reject the output without penalty. The output and its uncertainty must be disclosed in any judicial application materially relying on it.

Generative systems such as MahaCrimeOS should preserve source citations for every generated factual proposition. Automatically drafted notices, summaries or investigative plans must be labeled, logged and verified. Generated material should never silently overwrite primary records. Model prompts, retrieved documents and output versions should be retained long enough for legal challenge, with protection for unrelated personal data.

Police personnel should receive role-based access rather than broad platform access. Search justifications, queries, viewed records, exports and changes should be immutably logged. Supervisors and independent auditors should detect celebrity searches, personal disputes, bulk downloads, unusual geographic queries and repeated access unrelated to assigned cases.

Cybersecurity must include independent penetration testing, strong multifactor authentication, device management, encryption, segmentation, least privilege, vendor-access controls and rapid breach notification. Biometric or criminal-record exposures require notice to affected people unless a narrowly reasoned delay is necessary for an active investigation. The TSCOP incident reporting illustrates why police systems cannot rely solely on internal remediation. citeturn1search9turn1search13

Procurement rules should make auditability a non-waivable requirement. Contracts must disclose the deploying authority, vendor, subcontractors, model provider, cloud provider, cost, term, data rights and performance obligations. Proprietary intellectual property should not bar the government, courts or designated auditors from examining code, features, training provenance, thresholds and logs.

Material model updates should trigger reauthorization. A system validated as version one cannot silently shift to a new face engine, language model or training dataset. The deployed version, update date, release notes and evaluation should be recorded. This requirement directly addresses the present inability to determine the precise versions of TSCOP, Trinetra, YAKSH, CMAPS derivatives and district tools.

An independent Police Technology and Data Commission should authorize the highest-risk systems, conduct inspections, receive confidential technical material, issue binding correction and suspension orders, and publish annual reports. Its membership should include policing, criminal procedure, statistics, machine learning, cybersecurity, constitutional law, caste and minority discrimination, disability, gender, child rights and community representation.

Courts and affected persons should have access to meaningful explanations. An explanation need not reveal exploit-sensitive source code, but it must identify the data used, significant factors, record status, uncertainty, officer action and review route. A person should be able to challenge mistaken identity, an outdated case outcome, an incorrect association or unlawful watchlist placement without first proving exactly which hidden system caused the harm.

RTI proactive disclosure should be strengthened by a statutory police-technology register. For every system, the register should list status—announced, tendered, pilot, deployed, suspended or discontinued—current version, geographic coverage, purpose, data classes, vendor, cost, assessment dates, aggregate use, error rates and complaints. This would prevent discontinued prototypes from being repeatedly described as current operational systems and announcements from being mistaken for implementation.

Moratoria should apply where prerequisites are absent. Real-time facial recognition in public spaces, person-based risk scoring, political or protest network analysis, and bulk social-media monitoring should not operate without a specific statute, independent authorization, published assessment and demonstrable necessity. A procurement order or executive memorandum is not an adequate substitute.

The final technical classification is as follows:

The systems that most clearly qualify as predictive are therefore modest and place-based: the published Delhi hotspot work and, subject to confirmation of implementation, Washim’s Smart Prahari. They seek to forecast where or when recorded crime may concentrate. Neither has the public evidence needed to conclude that it reduces crime fairly and causally.

CMAPS is historically important because it was explicitly branded predictive, but the CAG evidence indicates that its verified implementation was dominated by mapping and basic analytics and that several advanced phases stalled. It should not be cited as proof that Delhi operated a mature predictive-policing platform without acknowledging the audit. citeturn11view1turn11view2

CCTNS and ICJS are more consequential than many branded AI projects because they create persistent, national, interoperable data infrastructure. Their principal function is not prediction, but they determine what future models can see. Data quality, legal-status separation and outcome correction within these platforms may therefore matter more than the choice of a particular machine-learning algorithm.

TSCOP, Trinetra and YAKSH appear powerful because they reduce the friction of searching, linking and identifying people. That power can be confused with prediction. On present evidence, their core is retrieval, facial or multimodal matching, operational monitoring and network analysis. YAKSH’s reported ranking features deserve urgent disclosure because they could cross into person-based risk assessment, but the public record does not permit that conclusion.

MARVEL’s current products are better described as AI-assisted investigation and administration. Their risks concern generative reliability, source fidelity, multilingual error, call-record interpretation, cloud governance and automation bias—not proven forecasts of future crime.

AI-SPS is a caution against taking product names literally. Its official privacy documentation verifies an “AI Smart Policing” application but describes employee attendance, GPS duty and patrol management. The “AI” label identifies methods or branding, not predictive policing. citeturn22search0

India’s immediate regulatory priority should therefore not be limited to banning a hypothetical national “pre-crime” score. The more urgent task is to govern the databases, labels, watchlists, biometric searches, association graphs and procurement relationships already being built. A future risk model trained on unreliable or discriminatory infrastructure will inherit those defects. Conversely, transparent, corrected and purpose-limited infrastructure would constrain harm even before specialized AI legislation is enacted.

The final conclusion is one of asymmetrical certainty. It is certain that India has built extensive digital criminal-justice infrastructure and that police forces are purchasing or developing facial recognition, geospatial analytics, network search and generative investigation tools. It is certain that official language increasingly promotes AI/ML use. It is certain, from CAG findings and court decisions, that record quality and discretionary classifications present material risks. What remains uncertain—and should remain expressly marked as uncertain—is which current systems perform operational future-risk calculation, with what variables, error rates and consequences. Until those questions are answered through law, disclosure and independent testing, “predictive policing” in India is less a verified nationwide capability than a heterogeneous field of record integration, surveillance, mapping, identity search, experimental forecasting and aspirational branding.