"(a) What Are the Ethical Issues Involved in the Use of AI in Data-Driven Policing? (b) Place Yourself in Ravi's Role and Discuss the Alternatives Available" — UPSC Mains 2026 GS4 Q8
A complete, examiner-standard 20-mark case study answer for the UPSC Mains 2026 GS Paper 4 predictive policing case — with a stakeholder table, the bias feedback-loop diagram, an options comparison, a phased course of action, and the Puttaswamy proportionality framework.
UPSC Mains 2026 GS Paper 4 set a 20-mark case study on AI-enabled predictive policing, algorithmic bias and community trust. Below is a full model answer with a static-portion refresher.
Ravi, SP of a riot-prone district, introduced an AI predictive-policing system capturing crowd biometrics and matching them to a data library. It flagged an immigrant, low-income neighbourhood as a hub of gang violence and drug trafficking; focused patrolling, preventive detentions and checkposts followed, and public order visibly improved. Community leaders, civil rights lawyers and activists have now submitted a memorandum alleging the system rests on historical data shaped by social bias and discriminatory policing, that surveillance has created a climate of tension, and that residents do not know what data is held against them.
(a) What are the ethical issues involved in the use of AI in data-driven policing?
(b) Place yourself in Ravi's role and discuss the alternatives available. Justify the action that optimises compliance with ethics.
Model Answer
Stakeholders and Their Interests
| Stakeholder | Interest at Stake |
|---|---|
| Residents of the flagged neighbourhood | Privacy, dignity, freedom from stigma and arbitrary detention — and protection from the gang violence and drug trafficking they themselves suffer |
| Wider district population | Public order, safety, freedom from riots |
| Ravi as SP | Effective policing, constitutional propriety, institutional legitimacy, personal accountability for a system he introduced |
| Police personnel | Clear operating instructions, protection from liability for acting on flawed outputs |
| Civil society and rights advocates | Non-discrimination, due process, transparency |
| State and technology vendor | Reputational and legal exposure; the precedent this sets for other districts |
(a) Ethical Issues in AI-Driven Policing
1. Skewed Input
Training data is arrest and FIR data, which records where police looked, not where crime occurred
2. Algorithmic Flag
System identifies the historically over-policed area as high risk
3. Intensified Policing
Patrolling, checkposts and detentions concentrate there
4. Apparent Confirmation
More arrests are recorded, which feeds back as proof the prediction was correct
- The loop is self-validating — because the system's success is measured against data the system itself generates, a biased model can appear highly accurate. Improved arrest figures are therefore not evidence that the flagging was correct.
- Proxy discrimination — even if religion, caste or nationality are never used as variables, locality, income level and name patterns can encode them. Article 15's prohibition can be violated by effect without being violated by design.
- Collective attribution — treating a neighbourhood as suspect imposes suspicion on residents who have committed no offence, which is a form of collective stigmatisation rather than individualised policing.
- Mass biometric capture — recording the faces of everyone in a crowd surveils the innocent majority to identify a guilty few. Under K.S. Puttaswamy (2017), such intrusion must satisfy legality, legitimate aim, necessity and proportionality, and procedural safeguards. Public order is a legitimate aim, but that satisfies only the first limb.
- Absence of statutory basis — legality requires a law authorising the intrusion. Deployment resting on administrative decision alone is the weakest point of the entire arrangement.
- Chilling effect — the "climate of tension" reported in the memorandum is itself a harm, deterring lawful assembly, movement and expression protected under Article 19.
- Adverse record without notice — that residents do not know what data is held against them denies the elementary requirements of natural justice: notice, hearing and the opportunity to correct. An error in the database is uncontestable because it is invisible.
- The black-box problem — where the reasoning is not explainable, neither the officer acting on it nor the person affected can meaningfully evaluate it. Preventive detention on an unexplainable output is deprivation of liberty without articulable grounds.
- Presumption of innocence inverted — predictive systems act on people before an offence, shifting policing from response to pre-emption and placing the burden of rebuttal on the individual.
- Diffused responsibility — when a wrongful detention follows an algorithmic flag, accountability disperses between vendor, system and officer. Automation bias compounds this: personnel tend to defer to machine outputs even against their own judgment.
- Legitimacy as a policing asset — the Peelian principle holds that police effectiveness depends on public cooperation. Order secured at the cost of a community's trust is unstable, since intelligence, witnesses and cooperation dry up.
- Distributive burden — the benefits of improved order are district-wide while the costs of surveillance fall on one vulnerable population least able to contest them. On a Rawlsian test, an arrangement that improves aggregate outcomes by burdening the worst-off is difficult to justify.
- Efficiency versus justice — the utilitarian case is real: order has improved and residents of that very neighbourhood are among the victims of the gang violence. But a rights-based view holds that some means are impermissible regardless of aggregate benefit.
(b) Alternatives Before Ravi
| Option | Merits | Demerits |
|---|---|---|
| 1. Continue unchanged, citing results | Sustains visible order; avoids operational disruption | Ignores a credible rights complaint; compounds harm if bias is real; invites litigation and eventual judicial intervention on worse terms |
| 2. Suspend the system entirely | Immediately halts alleged harm; signals responsiveness | Forfeits a genuine capability; abandons crime victims in the same neighbourhood; treats an unverified allegation as proven |
| 3. Refer upward and await instructions | Procedurally safe for Ravi personally | Abdication — harm continues during the delay, and the officer who introduced the system owns responsibility for it |
| 4. Suspend high-harm uses, audit, then resume with safeguards | Precautionary where the stakes are liberty; preserves capability; verifies rather than assumes; restores legitimacy | Demands time, technical capacity and political courage; may satisfy neither critics nor operational staff immediately |
The Course That Optimises Ethical Compliance: Option 4
- Receive the delegation formally and record the memorandum — acknowledgement is not concession, and dismissing a documented complaint would itself be an ethical failure.
- Issue a written order barring coercive action on algorithmic output alone — no preventive detention, stop or search solely on a system flag; independent corroboration and recorded reasons mandatory. This is the single highest-impact safeguard and requires no external approval.
- Pause indiscriminate crowd biometric capture pending review, while retaining case-specific identification of persons already accused in registered offences.
- Review recent detentions from that locality for evidentiary sufficiency independent of the algorithm; release where none exists.
- Commission an independent bias audit — a third-party technical review with State Crime Records Bureau and academic participation, examining training-data provenance, disparate-impact rates and false-positive rates by locality. The critical question: is the model trained on crime data or on arrest data?
- Create a notice-and-correction mechanism — a defined route for a resident to learn what is recorded, contest it and have it corrected, with a designated grievance officer. This addresses the memorandum's most concrete and least contestable complaint.
- Publish a use policy — data categories collected, purpose limitation, retention and deletion periods, access controls and audit logging.
- Verify before accepting — the memorandum contains allegations, not findings. Precaution justifies suspending high-harm uses; it does not justify accepting the claim uncritically.
- Community policing alongside technology — beat officers, mohalla committees and youth engagement in the flagged neighbourhood, so that residents experience the police as protectors and not only as watchers.
- Address causes, not only symptoms — de-addiction services, schooling and livelihood linkages coordinated with the district administration. The system identified a symptom of concentrated deprivation; enforcement alone cannot resolve it.
- Train personnel against automation bias — algorithmic outputs are probabilistic inputs to judgment, never substitutes for it.
- Seek a statutory and SOP framework — write to state headquarters recommending clear authorisation, oversight and audit norms, since legality cannot be supplied at district level.
- Report transparently — publish periodic aggregate data on flags, detentions and outcomes disaggregated by locality, allowing external scrutiny of disparate impact.
- It refuses a false choice — the case appears to demand either technological effectiveness or rights protection. Option 4 shows the trade-off is largely artificial: most of the harm comes from how the system is used, not from its existence.
- Precaution is proportionate to the stake — where the potential harm is wrongful deprivation of liberty falling on a vulnerable group, acting before certainty is justified; the cost of a temporary pause is far lower than the cost of continued wrongful detention.
- It satisfies the proportionality test — narrowing use to corroborated, case-specific application moves the practice toward the least restrictive means and supplies the procedural safeguards Puttaswamy requires.
- It protects long-term effectiveness — legitimacy is not a constraint on policing but a precondition of it. Cooperation, intelligence and witnesses depend on trust.
- It reflects ownership — Ravi sought this system. Integrity requires that he examine its consequences rather than defend the decision that produced them.
Conclusion
AI in policing does not create new ethical categories; it industrialises old ones — bias, opacity and unaccountable discretion now operate at scale and at speed. The mature response is neither technophobia nor technological triumphalism, but disciplined use: the algorithm may direct attention, never authorise coercion. Ravi's obligation is to ensure that a tool he introduced to protect the public does not become an instrument through which one section of that public is governed differently from the rest.
Constitutional and legal framework: Article 14 (equality and non-arbitrariness), Article 15 (non-discrimination, including discrimination by effect), Article 19 (assembly, movement, expression and the chilling effect), Article 21 (life, personal liberty, privacy and due process), Article 22 (safeguards on arrest and preventive detention). K.S. Puttaswamy v. Union of India (2017) — privacy as a fundamental right and the four-fold proportionality test: legality, legitimate aim, necessity or least restrictive means, and procedural safeguards. Maneka Gandhi (1978) on fair procedure; E.P. Royappa (1974) on arbitrariness. Digital Personal Data Protection Act, 2023 and the exemptions available to State instrumentalities; Information Technology Act, 2000; Bharatiya Nagarik Suraksha Sanhita, 2023 provisions on arrest and preventive action; Identification of Prisoners Act, 1920 as replaced by the Criminal Procedure (Identification) Act, 2022 on measurement and biometric collection.
Technology-ethics concepts: algorithmic bias and disparate impact; the distinction between crime data and arrest data; feedback loops and self-fulfilling prediction; proxy variables; the black-box and explainability problem; automation bias; function creep; purpose limitation and data minimisation; human-in-the-loop, human-on-the-loop and human-out-of-the-loop; privacy by design; algorithmic impact assessment. Policing and governance: Sir Robert Peel's principles, particularly that police effectiveness depends on public approval and cooperation; community policing; the Second ARC 5th Report on Public Order; National Police Commission and Prakash Singh (2006) directions on police reform; NITI Aayog's National Strategy for Artificial Intelligence (2018) and Responsible AI principles (2021); India AI Impact Summit, New Delhi, February 2026; the EU AI Act's classification of predictive policing and remote biometric identification as high-risk or prohibited uses; Rawls's difference principle and veil of ignorance; the precautionary principle.
Answer Writing Tips for This Case Study
- The feedback-loop insight is the answer's centre of gravity — the model is trained on arrest data, which records where police looked rather than where crime occurred, so intensified policing produces the arrests that appear to validate the prediction. Explaining this shows you understand why the memorandum's claim is technically credible rather than merely asserting bias.
- Do not treat the community as monolithic. Residents of that neighbourhood are also the victims of gang violence and drug trafficking. Recognising that they need protection and fair treatment is what separates a nuanced answer from an anti-police one.
- Apply the Puttaswamy four-fold test explicitly and note that public order satisfies only the legitimate-aim limb. The weakest link is legality — deployment resting on administrative decision without statutory authorisation.
- The single highest-impact safeguard is a written order that no coercive action may rest on algorithmic output alone. It is within Ravi's own authority, needs no approval, and directly addresses the liberty harm. Lead your action plan with it.
- Do not accept the memorandum uncritically either. It contains allegations, not findings. Precaution justifies suspending high-harm uses pending an independent audit; it does not justify treating the claim as proved. Showing this balance is exactly what "optimises compliance with ethics" is testing.
- Structure part (b) as immediate, short-term and medium-term. A phased plan demonstrates administrative realism; an undifferentiated list of good ideas does not.
- Note that Ravi introduced the system himself. Integrity requires examining its consequences rather than defending the original decision — a point most candidates miss, and one that speaks directly to moral courage.
- Close on the operative principle: the algorithm may direct attention, never authorise coercion. A one-line rule an examiner can carry away is worth more than a paragraph of caution.
