Contemporary discourse surrounding surveillance and digital systems focuses extensively on artificial intelligence. Media headlines, civil rights litigation, and legislative debates routinely zero in on predictive policing algorithms—software like Geolitica (formerly PredPol) or ShotSpotter—interrogating their statistical validity and racial biases. This focus, while well-intentioned, treats a systemic symptom as the root cause. Essentially, these discussions tend see AI as an autonomous entity superimposing itself onto modern law enforcement. However, this is not the case. AI is merely the topmost layer of a massive "stack"—which we will refer to as the "digital policing stack"—of pre-existing digital policing infrastructure, and is essentially the analytical component, which gives meaning to the vast amount of data, collected through the "sensory organs," that is the means through which such data is collected. Instead of arguing for a traditional system of AI-regulation,
I rely on a combination of the techno-legal and socio-technical frameworks to argue that the challenge starts even before deployment of AI-driven algorithmic policing (referred to interchangeably as "AI-based policing" or "algorithmic policing") actually occurs, and that primary focus should be on regulating the pre-existing data layer that makes algorithmic policing possible, and not the algorithm alone. To do so, we must first understand two things: (1) what I mean when I use the term "digital policing," and; (2) the normative elements of it.
At its core level, digital policing refers to the integration of modern analytical and data-driven methods into law-enforcement, characterised by the so-called "dataification" of everyday activities such that vast amounts of information can be easily analysed by the state. Of course, most digital infrastructures involved in policing rely on the digitization and centralized storage of records rather than artificial intelligence. A major example is the Crime and Criminal Tracking Network & Systems (CCTNS), a national database that connects over 15,000 police stations to store digitized information on complaints, active investigations, and criminal cases.
Nonetheless, as algorithms become increasingly more powerful and pervasive, it is set to be an inevitable inclusion in the future of law-enforcement. As discussed, algorithms are not independent tools, but a part of the digital policing stack, operating on top of pre-existing digital infrastructure. For instance, since 2021, Tamil Nadu has been using a Facial Recognition Software (FRS) that is integrated with the CCTNS to match video feeds against digitized criminal photographs. Similarly, the Organised Crime Network Database (OCND) operates as a stack on top of the National Intelligence Grid (NATGRID) to analyse patterns, retrieve data and act as a tool to detect cross-state terror financing. The digital policing stack consists of several constituent elements that layer on top of one another to move policing from reactive investigation toward predictive, continuous, and data-driven governance.
We note that at each layer, there is a unique risk that is specific to that layer, and there have been well-documented harms emerging from this set-up. However, much of AI governance—especially in the field of predictive policing—assumes that predominant risk emerges from the algorithm itself despite, somewhat counterintuitively, recognising that a major cause for algorithmic harm is the datasets that exist within it. This is a category error. Legal risk is not something that switches on only once an algorithm is deployed; it accrues progressively, at every point where the state exercises a choice over information. To see why, it helps to ask a simple question at each layer: what has the state done to the data, and does that act carry independent legal significance?
Data collection determines what information exists. Before any classification or analysis happens, the state has already made a threshold decision about which facts about a citizen enter the record at all—biometric markers, forensic samples, CCTV footage, complaint histories. This is not a neutral act of "gathering." Under the Puttaswamy framework, collection itself engages the right to privacy, because the very existence of a record about a person, retrievable and re-usable indefinitely, is a distinct harm from whatever use is eventually made of it. A CCTNS entry sits dormant for years, but its creation already forecloses a certain kind of anonymity.
Classification determines what that information means. For instance, the Tamil Nadu Police recently adopted the Sexual Offender Profiling, Evaluation, Classification, Tracking, Risk Assessment and Unified Monitoring System ("SPECTRUM"), which assigns sexual offenders to colour-coded risk categories based on their perceived likelihood of reoffending. SPECTRUM does not add new facts about a person; it adds a judgment about the facts already collected—Red, Amber, or Green. That judgment then travels with the underlying record and conditions how every downstream system treats that person, independent of whether the classification was accurate, contestable, or even disclosed. This is where the state moves from record-keeper to adjudicator, quietly, without the due process architecture that ordinarily accompanies an adjudicative act. Classification is a legally consequential exercise precisely because it is sticky: once a person is tagged, the tag does the work that used to require a fresh determination each time.
Interoperability determines who can access and use it. A record confined to a single database, however invasive its contents, is bounded by the purpose for which it was collected. The moment it becomes interoperable—as NATGRID's federated-access model or the Inter-operable Criminal Justice System are designed to allow—that boundary dissolves. Purpose limitation, one of the DPDP Act's foundational principles, is not violated by any single act of collection; it is violated by the infrastructural decision to make data legible and queryable across agencies that had no role in collecting it. Interoperability is, in effect, a standing authorisation for future uses that were never contemplated at the point of consent (to the extent consent was sought at all, given the state's exemptions under the Act for law-enforcement purposes).
Decision-support determines how it shapes policing. This is the layer where classification and interoperability convert into operational consequence—a patrol pattern, a surveillance frequency, a threshold for "caution" in an officer's interaction with a specific individual. South Zone Inspector General Vijayendra Bidari Bidari's own description of SPECTRUM makes this explicit: the system does not merely store a categorisation; it dictates a procedure. That is a rule with the practical force of law, generated and applied without legislative authorisation, judicial oversight, or a codified standard of review.
Predictive AI determines what future inferences the State may draw from it, and it does so through choices that are themselves legally significant, not merely inherited from upstream layers. For instance, the Delhi Police uses the Crime Mapping, Analytics and Predictive System (CMAPS) to identify potential hotspots of crimes by plotting live distress calls and historical crime data onto a geospatial map to identify hotspots, allowing authorities to strategically deploy patrolling resources. A model like CMAPS is not a passive conduit for CCTNS; it makes independent design decisions that carry their own legal weight. Which features are engineered into the geospatial prediction, prior arrests, complaint density, a neighbourhood's historical classification scores, is a choice that determines whose past becomes a template for future deployment. Feature selection is where a classification made under SPECTRUM years earlier gets converted into a live variable weighted against a citizen who may never have been told a "Red" tag existed, let alone that it now feeds a hotspot algorithm patrolling their street. Similarly, Tamil Nadu's FRS makes a threshold decision every time it runs—the confidence score above which a match is treated as actionable. That threshold is the line between an investigative lead and a false accusation, set by an engineer or vendor with no public, contestable standard governing where it should sit. These are not "AI risks" in the generic sense the discourse usually means (bias, hallucination, opacity); they are specific legal choices—about evidentiary weight, about whose data trains the model, about what counts as a sufficient match—made inside the model architecture itself, and they deserve to be treated as such rather than folded into a vague complaint about "algorithmic bias."
So, what would a solution look like? The answer lies in recognising that every layer of the digital policing stack discussed above is an independent exercise of state power over information, carrying its own legal significance irrespective of what happens at subsequent stages. By the time an AI system recommends a patrol route, flags an individual, or predicts a crime hotspot, it is operating on a series of choices the State has already made—what information was collected, how it was classified, who it was shared with, and how it was operationalised. AI is therefore not an independent or objective decision-maker; it is the culmination of an existing socio-technical system.
This is where the sociotechnical framework becomes indispensable, because it explains why isolating the algorithm for scrutiny will always produce an incomplete account of harm. A sociotechnical reading insists that the "social" and the "technical" cannot be prised apart—the outcome of a predictive policing tool is never a property of the model alone, but of the model's interaction with the institutions, workflows, and human judgment surrounding it. Applied to the digital policing stack, this means that CMAPS does not "cause" over-policing of a locality in the way a faulty component causes a machine to fail. It causes over-policing in combination with how classification labels were assigned upstream, how interoperability rules routed that data to a particular unit, and how an officer on the ground chooses to exercise the discretion that decision-support layers structure but do not eliminate. Auditing the algorithm's statistical validity, while necessary, answers only the narrowest of the questions a court or regulator should be asking, and cannot tell you which layer caused the final harm.
This is also precisely why the techno-legal approach acts as a natural governance response. Rather than treating regulation as something imposed externally once an AI system is deployed, it argues that legal, technical and institutional safeguards should be embedded throughout the lifecycle of a system. Applied to the digital policing stack, this means that governance cannot begin at the predictive AI layer because legally significant choices have already been made long before an algorithm is introduced. Instead, safeguards must accompany each transformation of information within the stack: robust standards governing data collection, transparent and contestable criteria for systems such as SPECTRUM, clear limits on interoperability between databases, documented rules governing decision-support systems, and continuous oversight of predictive models after deployment.
This model's lifecycle framework is useful here because it refuses to let the model stage collapse into a single monolithic "deployment" moment. Applied to CMAPS or FRS, this means that a model which performs well during testing may nevertheless generate harmful outcomes once deployed in changing operational environments. A techno-legal response, consistent with the white paper's own proposed controls, would mandate model-specific instruments at each of these sub-stages: red-teaming and bias scoring before a model like FRS is approved for deployment, and runtime monitoring with defined confidence thresholds once it's live, rather than a single upfront certification that assumes a model's behaviour is fixed at the point of training.
Naturally, this does not mean every layer requires the same form of regulation. Data collection raises questions of necessity, proportionality and minimisation. Classification raises questions of transparency, contestability and procedural fairness. Interoperability raises concerns around purpose limitation and access controls. Predictive AI introduces distinct concerns of explainability, validation and accountability. The point is not that every stage should be governed identically, but that every stage should be governed because every stage involves an independent exercise of state power over information.
Interestingly, this understanding is not entirely foreign to Indian law. The Digital Personal Data Protection Act, 2023, already proceeds on the intuition that governance obligations do not arise only at the point where data is ultimately used, but attach throughout its lifecycle. Obligations relating to notice, purpose limitation, data minimisation, accuracy, security safeguards, retention and erasure recognise that legal significance accompanies information as it is collected, processed, stored and ultimately deleted, rather than materialising only when a consequential decision is made. Although the Act creates significant exemptions for law-enforcement processing, its underlying governance philosophy remains instructive. It reflects an appreciation that the legal risks associated with information are cumulative, emerging not from a single act of processing but from a series of decisions about how information is handled over time. In many ways, this lifecycle-oriented approach mirrors the logic underlying the digital policing stack itself.
Most of our discussion has focused on how human error plays a huge role in how the digital policing stack is influenced by human biases. This means that a techno-legal approach, as discussed, cannot operate in isolation, and requires a socio-technical perspective of governance as well. This means that governance should extend to the state officials who interact with decision-support tools: training that emphasises the limitations of algorithmic outputs, documented reasoning when recommendations are followed or overridden, and periodic institutional review to identify patterns of over-reliance or misuse. Eliminating discretion altogether is not the way forward—it has been recognized as an imperative part of law-enforcement procedure.
The challenge, therefore, is to shift the object of governance away from the algorithm alone and towards the information lifecycle that makes algorithmic policing possible. Each stage of the digital policing stack represents an independent exercise of state power over information and therefore warrants its own set of legal, technical and institutional safeguards. Governance should accompany information throughout its lifecycle, from collection and classification to interoperability, operational use and predictive analysis, if India's digital policing infrastructure is to remain both technologically capable and constitutionally legitimate.