Introduction
In September 2022 a 102-year-old man, Dhuli Chand, rode through Rohtak on a chariot, with a band playing, in a mock wedding procession. The theatre had a plain purpose: he wanted the officials of Haryana to accept that he was alive. His old-age pension had stopped after the State’s family database, the Parivar Pehchan Patra, recorded him as dead.1 His case was not unusual. The State told its own legislature in August 2023 that pensions had been halted for 277,115 elderly persons and 52,479 widows over three years because they were recorded as deceased, and the Chief Minister conceded that 44,050 of 63,353 old-age pensioners whose benefits had been stopped were subsequently found eligible.2 According to the reporting, the government did not answer requests for information about the logic or source code behind the system.3
No judge decided Dhuli Chand’s case, and the reporting records no officer’s written reasons for it. A database did the sorting, and a human being had to undo it. Episodes of this kind give a face to a structural shift. Welfare targeting, tax scrutiny, policing, credit scoring and recruitment are increasingly mediated by statistical systems that place people in categories no statute defined. Such systems are often presented as neutral, but an algorithm carries human choices about which data to use, what to optimise and how much error to tolerate.4 The India AI Governance Guidelines of the Ministry of Electronics and Information Technology (MeitY), released in November 2025, list bias and discrimination among the principal categories of AI risk, with the example of inaccurate data used to decide on future employment.5 The same document concludes, on its present assessment of risk, that a separate law to regulate AI is not needed because existing laws can address many of the harms.6 If that is right, the Constitution is the first of those existing laws, and the task of saying what it demands falls to constitutional lawyers rather than to engineers.
This paper takes up that task. Its title compresses the argument. Code, when a public authority adopts it to make or shape decisions about persons, is state action. State action by code can be challenged as arbitrary under Article 14. And where such code burdens people through proxies and patterns rather than through an expressly forbidden classification, Articles 14 and 15 must be read substantively, so that the effect of a rule counts for as much as its form. Three questions follow. First, when does a decision produced by software, possibly built and hosted by a private firm, become an act of the “State” for the purposes of Article 12 and Part III? Second, what does arbitrariness mean when the decision-maker is a statistical model whose reasoning no person can fully reconstruct, and how far do E.P. Royappa, Maneka Gandhi and Shayara Bano supply a workable test? Third, can the recognition of indirect discrimination in Lt. Col. Nitisha v. Union of India be carried over to algorithmic systems, and what remedial structure would give that recognition practical force?
Comparative scholarship has begun to answer parts of these questions. Aziz Huq has argued that American constitutional doctrine, built around discriminatory intent and individual litigation, fits machine learning badly because the relevant harms flow from system-level design choices; he proposes a combination of ex ante regulation and aggregate litigation.7 The Cambridge volume edited by Micklitz and others places the same difficulty within a wider worry that algorithmic power is increasingly exercised by private actors and escapes the checks built for officials.8 Sulmicelli, writing in the Indian Journal of Law and Technology in 2025, adds a constructive claim: that AI can also strengthen the detection and enforcement of rights, provided it sits inside four safeguards, namely transparency, accountability, contestability and rights-compatibility by design.9 Indian writing has been quicker to diagnose than to doctrinalise. Vidushi Marda mapped the limits of data-driven decision-making in Indian policy as early as 2018,10 and two recent commentaries press for a constitutional doctrine of algorithmic review grounded in Articles 14 and 21 and for a right to human decision-making.11,12 What remains underdeveloped is the Article 15 side of the problem: how Indian equality law, having now acknowledged indirect discrimination, should deal with models that never mention caste, religion or sex yet distribute burdens along exactly those lines. This paper tries to join the two strands.
Policy documents, chiefly NITI Aayog’s Responsible AI papers of 2021 and the MeitY Guidelines of 2025, are examined as indications of the standards the executive itself accepts, since those acceptances bear on what counts as reasonable State practice.13 Facts about deployments come from the reporting and public documents cited in the notes and should be read with the caution that attaches to journalism and official statements. “Code” is used broadly to mean any formalised decision procedure, from a weighted scoring rule to a trained model, that an authority uses to classify, rank or predict individuals. Generative systems, whose constitutional difficulties lie more in speech and deception, fall outside the scope.
The argument unfolds in five sections. Section II builds the state action foundation, tracing the instrumentality and public-function tests and the horizontal reach of Part III of the Constitution. Section III develops algorithmic arbitrariness under Article 14 and connects it to reasoned decision-making and fair procedure. Section IV turns to indirect discrimination and substantive equality, examines proxy variables, and measures the NITI Aayog and MeitY frameworks against constitutional standards. Section V concludes with recommendations.
Code as state action
A. Why attribution comes first
Review of an algorithm cannot begin with the algorithm. It begins with the question whether anyone answerable under Part III has acted. Fundamental rights in India ordinarily run against “the State”, which Article 12 extends to the Government and Parliament of India, the Governments and Legislatures of the States, and “all local or other authorities within the territory of India or under the control of the Government of India.”14 When a welfare officer cancels a pension, attribution is trivial. When the cancellation follows from a score computed on a platform that a vendor designed, hosts and updates, the chain of agency is harder to see. The phrase “state action” is American in origin, but Article 12 sets a more functional inquiry, and that is what makes the Indian argument easier to sustain than it would be in the jurisdiction where much of the literature is written.
B. The authority that adopts the code
The first route to attribution is also the simplest. If the body using a system is itself “State” under Article 12, its output is that body’s act, whatever the origin of the code. Departments, tax administrations and police forces need no instrumentality analysis, and for bodies in the grey zone the familiar factors from Ajay Hasia (financial control, monopoly status, deep and pervasive control, governmental character of the function) continue to apply.15
The more useful question is what counts as the decision. A statutory authority must bring its own mind to bear and may not act at the dictation of another.16 A rejection order issued because a score crossed a threshold that the officer cannot interrogate is therefore not the vendor’s act. It is the authority’s, and an unreflective one. The authority cannot plead the machine as an intervening cause, because reliance is itself a series of choices: to procure a particular system, to set its thresholds and outcome variable, to accept its outputs without independent inquiry. Each is an exercise of public power. Huq’s survey of regulatory, allocative and punitive uses of machine learning shows that the consequential choices are made at the level of design and policy rather than at the moment of application.17 This paper accordingly treats three moments as State conduct open to review: procurement, configuration and reliance. What the Constitution examines is the decision to rely, not merely the single output.
C. Private vendors and the public function
Many Indian systems are built or operated by private firms or by entities of mixed character. Under Article 12 case law, a body is drawn into “State” through the tests in Ajay Hasia and R.D. Shetty, which look to substance over corporate form, though Pradeep Kumar Biswas later insisted that the Government’s control must be pervasive and particular to the body concerned.18,19 On that approach, a software supplier with no governmental character is unlikely to qualify as an Article 12 authority.
Three further routes close the gap. First, the writ jurisdiction under Article 226 extends to “any person or authority”, and Binny Ltd. v. V. Sadasivan confirms that a private body discharging a public function can be answerable in public law.20 A vendor that runs a beneficiary-verification service and effectively settles eligibility questions that statute commits to officials is performing such a function. Second, the Constitution is not wholly vertical. Article 15(2) binds private persons with respect to shops, restaurants and public places,21 and a Constitution Bench majority in Kaushal Kishor v. State of Uttar Pradesh accepted that rights under Articles 19 and 21 may, in some circumstances, be enforced against non-State actors.22 Third, and most importantly, the State cannot contract out of Part III. If the vendor is private and the authority relies on its product, the authority remains answerable for what it adopts. Pasquale’s chapter in the Micklitz volume warns of a “contractual creep” toward automated adjudication and argues for procedural protections that cannot be bargained away; he recounts Australia’s Robodebt scheme and Michigan’s MIDAS, where flawed data and assumptions produced erroneous automated charges, reportedly 93 per cent of them in Michigan.23
Indian executive documents lean the same way. NITI Aayog’s second Responsible AI paper urges that government procurement follow responsible-AI principles and that procurement documents address liability where those principles are violated, and the MeitY Guidelines state that accountability should follow the function performed, the risk of harm and the due diligence undertaken.24 These are not binding, but they show that the executive itself does not regard outsourcing as dissolving responsibility.
D. Secrecy as a defence
Characterising code as State action alters the force of the usual reply that the model is proprietary. Pollicino and De Gregorio locate the democratic difficulty of the algorithmic society in private power and in an opacity and lack of accountability that, they note, has been defined as an “algocracy”.25 Where the State relies on a model to decide rights, commercial confidentiality cannot be treated as an automatic answer. Even the Right to Information Act, 2005 exempts commercial confidence only where the competent authority is not satisfied that a larger public interest warrants disclosure, and it contains a wider public-interest override.26 The Haryana reporting, in which questions about the system’s logic went unanswered, shows what is lost when secrecy is accepted as the default.
E. Three objections
The first objection is that code is merely a tool, no more a State actor than a calculator. The answer lies in what the tool does. A calculator does not choose the variables, the cut-off or the objective; a scoring model does, and those choices are the substance of the decision. The second objection is breadth: if every database is state action, the category loses meaning. But characterisation only opens the door to scrutiny. It does not decide violation, and courts can calibrate review through reasonableness and proportionality, applying greater rigour where a system’s output is decisive or heavily weighted and liberty or livelihood is at stake. The third objection concerns the human in the loop. If an officer signs every order, is the model still the real decision-maker? The attribution then runs to the officer, but the arbitrariness question shifts rather than disappears. Sharma notes that officials who defer to risk scores without independent assessment effectively abdicate their discretion,27 and Olsen, Slosser and Hildebrandt argue that the duty to give reasons in administrative law applies to computationally aided decisions no less than to others.28
Indian courts have not shied from examining technological design as State conduct. In Puttaswamy (Aadhaar) the Court tested the architecture of the identification scheme, including its authentication, retention and private-sector uses, against Part III, and read down or invalidated several provisions.29 The step taken here is to extend that willingness from the architecture of a national database to the logic of a scoring model.
Algorithmic arbitrariness under Article 14
A. From classification to arbitrariness
For its first two decades Article 14 was applied mainly through the doctrine of reasonable classification: a law or order survived if it rested on an intelligible differentia that bore a rational nexus to the object sought.30 E.P. Royappa altered the centre of gravity. Equality, Bhagwati J. wrote, is a dynamic concept that cannot be confined within traditional limits, and in its positive sense it is the opposite of arbitrariness; the two are “sworn enemies”.31 Royappa itself concerned executive action, the transfer of a senior civil servant, and Maneka Gandhi carried the principle across Part III by insisting that State action affecting liberty must follow a procedure that is right, just and fair rather than arbitrary, fanciful or oppressive.32 Ajay Hasia then held that arbitrary State action violates Article 14 even where no classification can be identified.33 In Shayara Bano, Nariman J. applied to legislation the standard of “manifest arbitrariness”, meaning something done capriciously, irrationally or without adequate determining principle, or something excessive and disproportionate.34
Two clarifications keep this lineage in proportion. The manifest arbitrariness standard was stated by a plurality and was developed for statutes, where courts have long been cautious.35 That caution has little purchase here, because the systems at issue are instruments of administration, and executive action has been tested for arbitrariness since Royappa. The older, plainer standard is therefore enough. An algorithmic determination is arbitrary when the authority cannot show a determining principle, cannot show that the principle tracks the object, or cannot show that the decision was reached through a procedure a reasonable person would call fair.
B. Four marks of algorithmic arbitrariness
Unstateable criteria. A classification must have a differentia that can be articulated. A trained model sorts people on correlations that its designers may not be able to express, and an official relying on it may know only that the person fell below a cut-off. If the authority cannot say what distinguishes the included from the excluded, it cannot demonstrate the rational nexus. Sharma draws a useful distinction here between complexity, which is not suspect in itself, and opacity that shields the State from review.36 The constitutional objection is to the second.
Wrong objective or poor data. A model optimises whatever outcome variable its designers give it, and designers often have to choose among imperfect proxies for the policy goal.37 In a much-cited study, an algorithm used to allocate extra care to patients in the United States used past health expenditure as a stand-in for health need; because less money had historically been spent on Black patients with the same illness burden, the tool under-referred them.38 A scoring rule that targets the wrong quantity is arbitrary in the classic sense that it takes irrelevant considerations into account. The Haryana record illustrates the data side of the problem: a widow and her nine-year-old granddaughter were each recorded as earning six lakh rupees a year, while, according to her family, the household’s only earner made about seven thousand rupees a month.39 The Chief Minister’s own concession, set out in Section I, that most of the stopped pensions he counted were later found to belong to eligible people speaks to a procedure whose reliability was not established before it was used to deprive.
Absence of reasons. The duty to record reasons is a safeguard against arbitrariness and the condition of meaningful review. The Supreme Court has held that an administrative authority exercising judicial or quasi-judicial functions must record its reasons, in part so that its decision can be reviewed, and the logic of that rule applies with equal force to eligibility determinations that cut off a livelihood.40 A model’s output is not itself a reason. Olsen, Slosser and Hildebrandt make a point worth adopting: no decision-maker, human or machine, lets us see “why” in a literal sense, so the legal question is whether the explanation supplied is accurate and sufficient for the purpose of contestation.41 The standard, then, is not that the code be open to inspection, though inspection may sometimes be needed. It is that the authority can state, in terms the affected person can answer, which facts about them drove the outcome. The MeitY Guidelines make this a design commitment in their sutra of “Understandable by Design”, although “to the extent technically feasible” is a soft qualifier for a constitutional setting.42
Fettered discretion. Where officers treat a score as dispositive, the discretion Parliament conferred is not exercised at all. This is arbitrariness by delegation, and it is aggravated by automation bias, the tendency to defer to an output because it looks objective.
C. Fair procedure and the limits of the individual hearing
Article 14 and Article 21 operate together after Maneka Gandhi, and the content of fair procedure includes notice, an opportunity to be heard and a reasoned decision.43 Applied to automated decisions, these requirements have an unusual consequence: the burden of correction should not fall on the person harmed. In Haryana, according to the reporting, residents had to establish that they were alive. A system that reverses the ordinary presumption of life, and then requires a pensioner to travel to a redress camp, is not offering a hearing in any real sense.
Huq’s analysis explains why the hearing alone is inadequate. The due process inquiry, he observes, asks about the reliability of a procedure as a whole, yet litigation proceeds case by case, and individual litigants have little incentive or capacity to uncover systemic design flaws; error in one case is not evidence of a flawed model, since most algorithms err sometimes.44 He concludes that a mix of ex ante regulation and aggregate litigation suits the machine-learning state. Indian law is better placed than American law to supply the second element. Public interest litigation, with its relaxed standing rules and its tradition of continuing supervision, was built for systemic wrongs affecting people who cannot easily sue.45 A petition challenging a State-wide eligibility algorithm fits that tradition. What the courts lack is a recognised vocabulary of what to ask for, which Section III.B above is intended to provide.
D. Proportionality and the choice of technology
Puttaswamy (Privacy) placed proportionality at the centre of rights review, requiring a legitimate aim, a rational connection, necessity, and a balance between the burden on the individual and the benefit to the State, and Anuradha Bhasin applied the test to orders that restricted internet access while requiring that such orders be published and reasoned.46 The choice of technology belongs inside that analysis. Where a transparent decision rule or a simple eligibility check would meet the policy objective with lower risk of error, adopting an opaque statistical model is hard to justify as necessary. The question for the authority is not only whether the model is accurate, but whether a less intrusive and more legible alternative was available.
E. Burden and comparative anchors
The facts about a model’s construction and validation lie with the State and its vendor. A petitioner who can show only that an adverse decision followed an automated output is at a structural disadvantage. A calibrated presumption would respond: once a prima facie case is made that a rights-affecting decision rested materially on automated output and was unexplained, the State should be required to produce the validation record, the features and the reasons. The proposal borrows from the general rule that facts especially within one party’s knowledge must be proved by that party, and from Sharma’s call for a rebuttable presumption of arbitrariness.47
Comparative practice shows both the risk and the remedy. In State v. Loomis, the Wisconsin Supreme Court accepted a proprietary risk score in sentencing, but only because it was not determinative and on condition of written warnings about its limits, while acknowledging that the defendant could not examine the method.48 By contrast, the District Court of The Hague in 2020 set aside the Netherlands’ SyRI welfare-fraud risk legislation as insufficiently transparent and verifiable to satisfy the proportionality demands of the European Convention.49 Sulmicelli’s four pillars of transparency, accountability, contestability and rights-compatibility by design summarise what those decisions, in different registers, require.50 Article 14, properly read, already contains them.
Indirect discrimination and substantive equality under Articles 14 and 15
A. The doctrinal gap
Article 15(1) forbids the State from discriminating against any citizen “on grounds only of religion, race, caste, sex, place of birth or any of them.”51 Read literally, the clause invites a search for the forbidden ground as the operative reason for a decision. Algorithmic systems rarely supply one. A scoring model does not record that it disfavoured an applicant because of caste; it records that the applicant lived in a particular locality, held a certain kind of bank account or spoke a particular language. If equality law waits for the prohibited ground to appear on the face of the decision, it will find nothing to examine.
Indian courts have, however, been moving away from a form-bound reading. Anuj Garg looked past a protective purpose to the stereotype underneath a law restricting women’s employment, and Navtej Singh Johar read the grounds in Article 15 broadly enough to reach the stereotypes attached to sex and sexual orientation.52 American doctrine offers a cautionary contrast. Under Washington v. Davis, a facially neutral measure with disparate racial impact offends the Equal Protection Clause only if traced to discriminatory purpose.53 Huq’s critique is that purpose is a concept of individual psychology and sits awkwardly with systems whose outputs arise from data, objectives and design choices that no single person intended; he suggests that the true equality concern with such systems is their contribution to what he calls pernicious social stratification.54 India need not inherit that difficulty, and Nitisha shows why.
B. How a neutral model discriminates
Bias enters a model at several points. Designers choose a target variable that may be a poor stand-in for the policy goal; they train on historical records that already embody unequal treatment; they select features that correlate with protected traits, such as residential area; and removing an explicit protected attribute does nothing to sever those correlations.55 Barocas and Selbst describe how each of these steps can produce disparate impact without anyone intending it, and Wachter, Mittelstadt and Russell caution that legal tests of disparity are contextual and cannot be reduced to a single statistical measure.56,57
The Indian setting supplies its own proxies. Residential pin codes, surnames, school attended, spoken language, handset type and patterns of mobile or bank use can each correlate with caste, religion, gender or region. This claim is a matter of social structure rather than of any particular system, and whether a given deployment relies on such variables is an empirical question that only disclosure can answer. In the Haryana case, the public record does not show whether anyone tested the family-income predictions for differential error across groups; the reporting records no answer from the State to questions about the system’s logic. That silence is itself the problem. Without group-level error data, neither the court nor the affected person can tell whether a pattern of wrongful exclusion is random or structured.
C. Nitisha as the bridge
In Lt. Col. Nitisha v. Union of India, decided on 25 March 2021, the Supreme Court examined the criteria used to grant permanent commission to women Short Service Commission officers in the Army. The criteria were neutral on their face. Their effect, the Court found, was to disadvantage women whose career records had been shaped by years of exclusion from the opportunities that men enjoyed.58 Chandrachud J. observed that indirect discrimination was still at a “nascent stage” in India, defined it as discrimination caused by facially neutral criteria, and insisted that the inquiry “looks, not at the form of the impugned conduct, but at its consequences.”59 Drawing on Griggs v. Duke Power Co., he accepted that neutral practices can operate as “built-in headwinds” for disadvantaged groups regardless of intent.60
The Court then adopted the two-step test from the Supreme Court of Canada’s decision in Fraser v. Canada: first, whether the impugned rule disproportionately affects a particular group; second, whether it reinforces, perpetuates or exacerbates the disadvantage that group already suffers.61 Two further holdings matter for algorithmic cases. Indirect discrimination can be established without statistical evidence, which is important where the State controls the data;62 and structures that freeze an unjust status quo are themselves the target of substantive equality.63
The test translates readily. At the first step, the question is whether a model’s errors or burdens fall disproportionately on a protected or disadvantaged group, which requires disaggregated performance data that only the deployer is positioned to produce. At the second, the question is whether the model reinforces existing disadvantage. Here, systems trained on historical allocations are structurally exposed: past patterns of State attention are themselves products of the disadvantage in question, and a model that learns from them will tend to repeat them. Huq makes the point that race may structure past deployment of State resources in ways that are hard to disentangle from training data.64 A third stage remains, as in Griggs, where the State may justify a disparity by showing that the criterion is necessary to a legitimate end, and the question then becomes whether a less discriminatory alternative of comparable accuracy existed.65 For code, that is a practical inquiry: could another feature set, threshold or model have achieved the same objective with a smaller disparity?
D. Substantive equality and the limits of “fairness through unawareness”
The recognition of indirect discrimination belongs to a larger commitment. Sandra Fredman has described substantive equality as having redistributive, recognition, participative and transformative dimensions, asking not only whether like cases are treated alike but whether law redresses disadvantage, counters stigma and stereotype, enables voice and accommodates difference.66 Khaitan likewise places the systemic disadvantage of groups at the centre of discrimination law.67 Measured against these standards, a model that is blind to protected traits is not therefore fair. Huq notes that mere awareness of race by officials is not the same as impermissible classification, and the same logic applies to model design: attribute-aware testing and correction may be necessary to detect and reduce disparity.68
The Indian text is hospitable to that approach. Articles 15(3) to 15(6) and 16(4) authorise group-conscious measures for women, children, socially and educationally backward classes, Scheduled Castes and Tribes, and economically weaker sections.69 A State that deploys a ranking system in education, employment or welfare must ensure it does not silently undo these measures. Sulmicelli observes that constitutional systems with stronger positive rights foundations create legal space for AI-augmented enforcement that the American order cannot easily accommodate, and the same can be said of India.70 He also identifies a limit condition of particular relevance here: the systematically weaker performance of AI tools in non-English linguistic environments.71 In a country of many languages, a system that performs well only in a few is not a neutral instrument. Its failures may fall on speakers of other languages, who may also be among the poorer and more remote citizens.
E. What the policy documents offer
NITI Aayog’s Responsible AI papers of 2021 are an early official attempt to translate constitutional values into AI principles. Part 1 acknowledges that AI may be biased against sections of society and warns that mistakes in beneficiary identification can exclude people from guaranteed services. Its principles of equality, which requires that individuals in the same circumstances relevant to the decision be treated equally, and of inclusivity and non-discrimination, which forbids denying opportunity to a qualified person because of identity, restate Articles 14 and 15 in recognisable form.72 Part 2 adds recommendations on procurement, risk assessment, audits “to the extent possible” and an apex advisory body.73 Two limits stand out. The equality principle is formal: it asks for like treatment, not for scrutiny of effects, and it predates Nitisha by weeks. And the recommendations are advisory and do not create duties enforceable by those harmed. NITI Aayog’s 2025 growth report, AI for Viksit Bharat, treats fairness mainly as a measure to be tracked in financial services and says little about discrimination in public deployments.74
The MeitY Guidelines of November 2025 go further in one respect. Their “Fairness and Equity” sutra speaks of outcomes that are fair, unbiased and non-discriminatory, which is nearer to an effects-based standard than the 2021 formulation, and the Guidelines list bias and discrimination as a risk category, endorse graded liability, call for grievance redressal and refer to a voluntary fairness-assessment standard issued by the Telecommunication Engineering Centre.75 But the Guidelines also advise that a separate AI law is unnecessary at present, place “Innovation over Restraint” among their sutras, and mention constitutional provisions only in listing the existing legal framework.76 They impose no specific duty on public deployers to assess equality impact before rolling out a system.
Neighbouring statutes do not fill the gap. The Digital Personal Data Protection Act, 2023 allows the State to process data for benefits and services without consent in specified circumstances and empowers the Government to exempt notified State instrumentalities, and it contains no provision on automated decision-making comparable to those found in European data protection law.77 Its Rules were notified on 13 November 2025, with most provisions to take effect after eighteen months.78 The comparison with the United Kingdom is instructive. In R (Bridges) v. Chief Constable of South Wales Police, the Court of Appeal found a breach of the public sector equality duty because the force had not taken reasonable steps to discover whether its facial recognition software carried racial or gender bias.79 India has no statutory equality duty, but the Constitution contains an analogue in Article 38(2), which directs the State to strive to minimise inequalities in income and to endeavour to eliminate inequalities in status, facilities and opportunities, among individuals and groups.80 Read with Articles 14 and 15, it supports a constitutional expectation that equality impact be assessed before a system is deployed, not discovered afterwards through complaints.
Conclusion
This paper began with a pensioner forced to perform his own survival. His predicament contains the whole argument in small. A public authority adopted a system, relied on its output, and left the person it harmed to carry the burden of proving the system wrong. The preceding sections have contended that the Constitution does not permit this arrangement and that its existing doctrines, properly extended, supply the means of saying so.
Three conclusions follow. First, code adopted by a public authority to decide matters affecting persons is state action. Attribution turns on the decision to procure, configure and rely on a system, not on who wrote the software, and the State cannot contract out of Part III. Second, an automated determination is arbitrary under Article 14 when the authority cannot state its operative criteria, cannot show that the model tracks the policy object, cannot give reasons that the affected person can contest, or has surrendered its discretion to a score. Third, Articles 14 and 15 must be read substantively. After Nitisha, a facially neutral criterion that disproportionately burdens a disadvantaged group and entrenches that disadvantage is open to challenge without proof of intent and, in appropriate cases, without statistics. Models that discriminate through proxies are the paradigm case for that reasoning.
The policy documents surveyed here are valuable as statements of aspiration and as admissions of what the executive accepts. NITI Aayog’s 2021 principles and the 2025 MeitY Guidelines both name fairness, accountability and understandability as goals. But they are voluntary, they rest on the premise that existing law suffices, and they do not tell a judge or a pensioner what to do when a system fails. The constitutional argument is what turns their commitments into obligations. Four proposals follow.
A presumption in litigation. Where a petitioner shows prima facie that a rights-affecting decision rested materially on an automated output and was not explained, courts should require the State to produce the system’s purpose, features, validation record, disaggregated error rates and the reasons for the individual decision. Failing that, the decision should be presumed arbitrary. The presumption is rebuttable and should be calibrated to the stakes.
Equality impact assessment before deployment. Grounded in Articles 14, 15 and 38(2), a duty to assess disparate impact before a public system goes live should be adopted by statute or binding executive instruction. It should cover data provenance, candidate proxies, performance by group and by language, less discriminatory alternatives and a plan for monitoring, with publication of a summary and independent audit for high-stakes uses. The institutions envisaged in the MeitY Guidelines, including the proposed AI Governance Group and AI Safety Institute, could host the technical function.81 Procurement contracts should reserve audit rights and documentation access, and should not allow proprietary secrecy to defeat judicial or regulatory review.
A right to human review. Persons should be entitled to notice that an automated system contributed to an adverse decision, to the factors relied upon, to a human review that is capable of departing from the output, and to contest the outcome before it takes effect. These elements track what Chaudhary proposes as a right to human decision-making.82 They require that the burden of correcting error fall on the State, and they would need a legislative home, perhaps through amendment of the data protection framework, which presently has no automated-decision provision.
Aggregate remedies. Individual hearings cannot reveal systemic defects. Huq’s preferred combination of ex ante regulation and aggregate litigation is more achievable in India than in the United States because public interest litigation already exists.83 Courts should be ready to appoint expert panels, take confidential disclosure of source material, and issue continuing mandamus requiring audits and corrections across a programme.
These proposals should not be read as hostility to technology. Sulmicelli observes that the same capabilities can strengthen the detection and enforcement of rights, and that legitimacy depends on transparency, accountability, contestability and rights-compatibility by design.84 Properly governed, models could help identify wrongful exclusion in the State’s own welfare rolls. The Constitution’s demand is that such tools serve the persons they classify rather than the convenience of those who classify them.
Limits remain. The paper has relied on public documents and reporting rather than audits of Indian systems, and the doctrinal bridges it proposes have not been tested by a court in an algorithmic case. These are reasons for caution about the pace of judicial development, not for doubting the direction. When the next Dhuli Chand comes before a court, the question should not be whether he can prove he is alive. It should be whether the State can prove that its system deserved to be believed.
*****
Footnotes
1. Kumar Sambhav, Tapasya & Divij Joshi, In India, an Algorithm Declares Them Dead; They Have to Prove They’re Alive, Al Jazeera (Jan. 25, 2024), https://www.aljazeera.com/economy/2024/1/25/in-india-an-algorithm-declares-them-dead-they-have-to-prove-theyre
2. Id. (reporting figures given to the Haryana Legislative Assembly and statements by the Chief Minister in August 2023).
3. Id.
4. Oreste Pollicino & Giovanni De Gregorio, Constitutional Law in the Algorithmic Society, in Constitutional Challenges in the Algorithmic Society 3, 4–5 (Hans-W. Micklitz et al. eds., 2021), https://doi.org/10.1017/9781108914857.002
5. Ministry of Electronics & Info. Tech., Gov’t of India, India AI Governance Guidelines: Enabling Safe and Trusted AI Innovation 25 (Nov. 2025), https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc2025115685601.pdf [hereinafter MeitY Guidelines].
6. Id. at 10.
7. Aziz Z. Huq, Constitutional Rights in the Machine-Learning State, 105 Cornell L. Rev. 1875, 1883–84 (2020), https://publications.lawschool.cornell.edu/lawreview/wp-content/uploads/sites/2/2020/12/Huq-final.pdf
8. Constitutional Challenges in the Algorithmic Society (Hans-W. Micklitz et al. eds., 2021), https://doi.org/10.1017/9781108914857 [hereinafter Micklitz et al.].
9. Sergio Sulmicelli, AI-Augmented Constitutionalism: Framing the Alliance Between AI and Fundamental Rights in Comparative Perspective, 21(2) Indian J.L. & Tech. art. 4 (2025), https://doi.org/10.55496/SPUE2081
10. Vidushi Marda, Artificial Intelligence Policy in India: A Framework for Engaging the Limits of Data-Driven Decision-Making, 376 Phil. Transactions Royal Soc’y A, art. 20180087 (2018), https://doi.org/10.1098/rsta.2018.0087
11. Paras Sharma, The Ghost in the Administrative Machine: Toward a Constitutional Doctrine of Algorithmic Review in India, Law Sch. Pol’y Rev. (Mar. 18, 2026), https://lawschoolpolicyreview.com/2026/03/18/the-ghost-in-the-administrative-machine-toward-a-constitutional-doctrine-of-algorithmic-review-in-india/
12. Raghvendra Kumar Chaudhary, Constitution Never Approved Algorithmic Governance: Why India Needs Right to Human Decision-Making, LiveLaw (Aug. 6, 2026), https://www.livelaw.in/amp/articles/constitution-never-approved-algorithmic-governance-india-right-human-decision-making-544523
13. NITI Aayog, Responsible AI #AIFORALL: Approach Document for India, Part 1 – Principles for Responsible AI (Feb. 2021), https://www.niti.gov.in/sites/default/files/2021-02/Responsible-AI-22022021.pdf [hereinafter NITI Part 1]; NITI Aayog, Responsible AI #AIFORALL: Approach Document for India, Part 2 – Operationalizing Principles for Responsible AI (Aug. 2021), https://www.niti.gov.in/sites/default/files/2021-08/Part2-Responsible-AI-12082021.pdf [hereinafter NITI Part 2].
14. India Const. art. 12.
15. Ajay Hasia v. Khalid Mujib Sehravardi, (1981) 1 S.C.C. 722, ¶ 9 (India).
16. Comm’r of Police v. Gordhandas Bhanji, A.I.R. 1952 S.C. 16 (India).
17. Huq, supra note 7, at 1890–1905.
18. Ajay Hasia, supra note 15; R.D. Shetty v. Int’l Airport Auth. of India, (1979) 3 S.C.C. 489 (India).
19. Pradeep Kumar Biswas v. Indian Inst. of Chem. Biology, (2002) 5 S.C.C. 111, ¶ 40 (India).
20. India Const. art. 226; Binny Ltd. v. V. Sadasivan, (2005) 6 S.C.C. 657 (India).
21. India Const. art. 15(2).
22. Kaushal Kishor v. State of Uttar Pradesh, (2023) 4 S.C.C. 1 (India).
23. Frank Pasquale, Inalienable Due Process in an Age of AI: Limiting the Contractual Creep Toward Automated Adjudication, in Micklitz et al., supra note 8, at 42–43, https://doi.org/10.1017/9781108914857.004
24. NITI Part 2, supra note 13, at 11, 20, 27; MeitY Guidelines, supra note 5, at 13, 32.
25. Pollicino & De Gregorio, supra note 4, at 3–4.
26. Right to Information Act, 2005, No. 22 of 2005, § 8(1)(d), (2) (India).
27. Sharma, supra note 11, pt. II.
28. Henrik Palmer Olsen, Jacob Livingston Slosser & Thomas Troels Hildebrandt, What’s in the Box? The Legal Requirement of Explainability in Computationally Aided Decision-Making in Public Administration, in Micklitz et al., supra note 8, at 219.
29. Justice K.S. Puttaswamy (Retd.) v. Union of India (Aadhaar), (2019) 1 S.C.C. 1 (India).
30. State of W.B. v. Anwar Ali Sarkar, 1952 S.C.R. 284 (India); Budhan Choudhry v. State of Bihar, A.I.R. 1955 S.C. 191 (India).
31. E.P. Royappa v. State of Tamil Nadu, (1974) 4 S.C.C. 3, ¶ 85 (India).
32. Maneka Gandhi v. Union of India, (1978) 1 S.C.C. 248, ¶¶ 7, 48 (India).
33. Ajay Hasia, supra note 15, ¶ 16.
34. Shayara Bano v. Union of India, (2017) 9 S.C.C. 1, ¶ 101 (India) (Nariman, J.).
35. See State of A.P. v. McDowell & Co., (1996) 3 S.C.C. 709 (India).
36. Sharma, supra note 11, pt. III.1.
37. Huq, supra note 7, at 1885–90.
38. Ziad Obermeyer et al., Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations, 366 Science 447 (2019), https://doi.org/10.1126/science.aax2342
39. Sambhav, Tapasya & Joshi, supra note 1.
40. S.N. Mukherjee v. Union of India, (1990) 4 S.C.C. 594 (India).
41. Olsen, Slosser & Hildebrandt, supra note 28.
42. MeitY Guidelines, supra note 5, at 13, 33.
43. A.K. Kraipak v. Union of India, (1969) 2 S.C.C. 262 (India); Maneka Gandhi, supra note 32.
44. Huq, supra note 7, at 1906–10.
45. Bandhua Mukti Morcha v. Union of India, (1984) 3 S.C.C. 161 (India).
46. Justice K.S. Puttaswamy (Retd.) v. Union of India (Privacy), (2017) 10 S.C.C. 1 (India); Anuradha Bhasin v. Union of India, (2020) 3 S.C.C. 637 (India).
47. Bharatiya Sakshya Adhiniyam, 2023, No. 47 of 2023, § 109 (India); Sharma, supra note 11.
48. State v. Loomis, 881 N.W.2d 749, 763–66 (Wis. 2016).
49. NJCM c.s. v. De Staat der Nederlanden, Rechtbank Den Haag [Hague District Court] 5 Feb. 2020, ECLI:NL:RBDHA:2020:1878 (Neth.) (SyRI).
50. Sulmicelli, supra note 9, pt. III.D.
51. India Const. art. 15(1).
52. Anuj Garg v. Hotel Ass’n of India, (2008) 3 S.C.C. 1 (India); Navtej Singh Johar v. Union of India, (2018) 10 S.C.C. 1 (India).
53. Washington v. Davis, 426 U.S. 229, 239–42 (1976).
54. Huq, supra note 7, at 1917–19.
55. Id. at 1887–88.
56. Solon Barocas & Andrew D. Selbst, Big Data’s Disparate Impact, 104 Calif. L. Rev. 671, 677–93 (2016).
57. Sandra Wachter, Brent Mittelstadt & Chris Russell, Why Fairness Cannot Be Automated: Bridging the Gap Between EU Non-Discrimination Law and AI, 41 Computer L. & Sec. Rev. 105567 (2021), https://doi.org/10.1016/j.clsr.2021.105567
58. Lt. Col. Nitisha v. Union of India, Writ Petition (Civil) No. 1109 of 2020, ¶¶ 42–48, 98, 119 (SC Mar. 25, 2021), 2021 SCC OnLine SC 261 (India).
59. Id. ¶¶ 46, 48, 53.
60. Id. ¶¶ 54–55 (discussing Griggs v. Duke Power Co., 401 U.S. 424 (1971)).
61. Id. ¶ 65 (adopting Fraser v. Canada (Att’y Gen.), 2020 SCC 28).
62. Id. ¶ 68.
63. Id. ¶¶ 66, 71.
64. Huq, supra note 7, at 1911 n.178.
65. Griggs, 401 U.S. at 429–32; see Barocas & Selbst, supra note 56, at 701–12.
66. Sandra Fredman, Substantive Equality Revisited, 14 Int’l J. Const. L. 712 (2016).
67. Tarunabh Khaitan, A Theory of Discrimination Law ch. 5 (2015).
68. Huq, supra note 7, at 1920.
69. India Const. arts. 15(3)–(6), 16(4).
70. Sulmicelli, supra note 9, abstract; see Students for Fair Admissions, Inc. v. President & Fellows of Harv. Coll., 600 U.S. 181 (2023).
71. Sulmicelli, supra note 9, abstract.
72. NITI Part 1, supra note 13, at 15–17, 41.
73. NITI Part 2, supra note 13, at 4, 13, 20–23, 27.
74. NITI Aayog, AI for Viksit Bharat: The Opportunity for Accelerated Economic Growth 22, 24 (2025), https://niti.gov.in/sites/default/files/2025-09/AI-for-Viksit-Bharat-the-opportunity-for-accelerated-economic-growth.pdf
75. MeitY Guidelines, supra note 5, at 13, 25, 32, 52.
76. Id. at 10, 13, 18.
77. Digital Personal Data Protection Act, 2023, No. 22 of 2023, §§ 7(b), 17(2)(a) (India).
78. Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), r. 1(2)–(4) (India) (notified Nov. 13, 2025).
79. R (Bridges) v. Chief Constable of S. Wales Police [2020] EWCA Civ 1058, [2020] 1 W.L.R. 5037 (Eng.).
80. India Const. art. 38(2).
81. MeitY Guidelines, supra note 5, at 1 (Preface).
82. Chaudhary, supra note 12.
83. Huq, supra note 7, at 1883–84.
84. Sulmicelli, supra note 9, pt. III.D.