Governing by Algorithm: Constitutional Limits on Automated State Decisions in India

1 Introduction

The constitutional problem created by automated government is often framed too dramatically. The question is not whether a machine may ever participate in public decision-making. Public administration has always relied on instruments that classify, aggregate and standardise information. Digital systems can reduce delay, expose inconsistent treatment, assist overburdened officials and make public services easier to administer at scale. The harder question arises when computational systems cease to be merely clerical and begin to structure the reasons for State action: a welfare claimant is flagged as ineligible, an applicant is ranked out of a public opportunity, a tax case is selected or assessed through a digital workflow, or access to an essential service turns on an automated risk or identity determination. At that point, efficiency collides with constitutional accountability. A citizen may confront an adverse result without knowing that automation mattered, what information drove the result, or whether any official was capable of reconsidering it.¹

Indian law does not yet contain a general statute governing automated public decisions. The Digital Personal Data Protection Act, 2023 (“DPDP Act”) is relevant because automated administration depends upon data processing and inference, but it does not create a GDPR-style right directed specifically at automated decisions. More importantly, its substantive processing duties and individual rights are being brought into force in stages. Under G.S.R. 843(E) of 13 November 2025, sections 3–17 and much of the core rights-and-obligations architecture fall within the group commencing eighteen months from publication; as of 30 August 2026, that group is not yet operative.² India’s 2025 AI Governance Guidelines move closer to the present problem. They place people at the centre, call for meaningful control and human oversight, encourage grievance redress, transparency and algorithmic auditing, and recognise threats to livelihood and well-being as especially salient. Yet they remain governance guidance rather than a self-executing constitutional remedy.³

The gap is therefore neither a complete absence of law nor a simple demand for new AI legislation. Indian constitutional and administrative law already insist that coercive or materially adverse public decisions remain non-arbitrary, procedurally fair, reasoned and judicially reviewable. The Supreme Court’s post-2021 jurisprudence has strengthened those requirements. T. Takano v. Securities and Exchange Board of India connected fair adjudication to disclosure of relevant material; State Bank of India v. Rajesh Agarwal treated natural justice as a substantive restraint on administrative arbitrariness; and Madhyamam Broadcasting Ltd. v. Union of India located reason-giving, disclosure and reviewability within a constitutional “culture of justification.”⁴ In 2026, however, State Bank of India v. Amit Iron Pvt. Ltd. supplied an equally important limit: natural justice does not invariably require an oral or personal hearing. Notice, disclosure, a meaningful opportunity to answer and a reasoned decision may satisfy fairness, depending upon statutory context and consequences.⁵

Existing Indian scholarship has already identified algorithmic administration as a problem for administrative law, and recent writing has pressed for a constitutional right to reasons in an emerging algorithmic State.⁶ The unresolved question is more specific: what procedural entitlement follows once an algorithm has materially shaped an adverse State decision? A general “right to a human decision” is both over-inclusive and under-theorised. It would protect trivial decisions merely because they are automated while saying too little about nominally human decisions that simply rubber-stamp an algorithmic score. This article develops a narrower doctrine. It argues that Articles 14 and 21 should guarantee meaningful human reconsideration where two conditions coincide: an automated system has materially contributed to State action, and that action has a materially adverse effect on rights, benefits, legal status, public opportunities or access to essential public services. The inquiry is doctrinal and comparative rather than empirical. It reads recent Indian public-law decisions together with emerging policy and data-protection developments, and uses the European Union and United Kingdom as contrasting regulatory models. The objective is not to constitutionalise technical design preferences. It is to preserve a basic proposition of public law: when the State acts seriously against a person, responsibility for the decision must remain intelligible, contestable and attributable.

2 Automated State Action and the Problem of the Nominal Human

Automated decision-making is not a binary category. A system may make a final determination without intervention; it may recommend an outcome that officials usually accept; it may assign a risk score or eligibility classification; it may prioritise cases for scrutiny; or it may filter evidence before an official sees the file. The legal importance of automation therefore cannot depend exclusively on who clicks the final button. A formally human decision may be functionally algorithmic when the officer lacks access to the underlying reasons, is institutionally expected to follow the score, or cannot practically revisit the model’s assumptions. Conversely, a genuinely assistive tool may improve administration without displacing human judgment. Scholarship on public-sector automation accordingly emphasises that automation changes the “decision space” between officials and citizens, not merely the identity of the final decision-maker.⁷

India’s faceless income-tax architecture illustrates the procedural stakes even though a faceless assessment should not automatically be described as an AI decision. Digital administration can separate the affected person from the decision-maker, channel communication through portals and standardise workflow. In D.B. Engineering Pvt. Ltd. v. National Faceless Assessment Centre, the Delhi High Court set aside an assessment where a hearing requested under the then-applicable statutory scheme was denied. The Bombay High Court reached a similar result in Sulzer Pumps India Pvt. Ltd. v. National Faceless Assessment Centre, and the Patna High Court in Vaibhav Singh v. National Faceless Assessment Centre treated failure to act on an e-portal request for video hearing as a clear violation of natural justice.⁸ These cases do not establish a constitutional right to an oral hearing in digital administration. Their deeper significance is institutional: digitisation does not dilute the duty to provide the procedure that law requires, and a portal cannot become the point at which a legally relevant request disappears.

The same danger becomes sharper when an algorithm contributes substantive judgment. Automation bias may induce officials to treat computational output as objective, while proprietary design may prevent them from explaining why the system produced a particular result. A “human-in-the-loop” label then obscures rather than cures the problem. The constitutional question should therefore turn on material contribution. If the automated output substantially influenced the adverse result—by supplying a decisive score, classification, inference, recommendation or evidentiary filter—the State should not escape procedural scrutiny merely because an official formally adopted it. This functional approach is consistent with the broader public-law concern with real decision-making power rather than administrative form.

3 Constitutional Foundations of Contestability

Article 14 supplies the first foundation. Modern Indian equality doctrine does not tolerate State arbitrariness simply because the challenged action is technologically mediated. In Madhyamam Broadcasting, the Supreme Court described proportionality as part of a constitutional culture in which public power must be justified. It treated reasons as fundamental to fair administration because they expose the link between material, inference and conclusion, discipline the decision-maker, and permit effective judicial review. The Court also rejected procedures in which the affected party was denied material necessary to challenge the case against it, observing that non-disclosure can reduce review to an exercise conducted in opacity.⁹ Algorithmic State action presents the same structural risk in a new form. A decision that cannot be connected to intelligible reasons is difficult to distinguish from an unexplained assertion of power.

T. Takano sharpens the disclosure principle. The Supreme Court held that material relevant to adjudication cannot be insulated from disclosure merely through the authority’s assertion that it did not rely on it, where the material has a nexus with the proceedings and could have influenced the outcome.¹⁰ Transposed carefully to automated administration, the principle does not require routine publication of source code. It does require access to decision-relevant information: the data or factual inputs attributed to the person, the principal factors or rules that mattered, and enough information about the system’s role to permit a meaningful challenge. A State agency should not be able to answer a request for reasons with the circular statement that “the system generated the result.”

Article 21 provides the second foundation. Since Maneka Gandhi v. Union of India, procedure affecting life or personal liberty must be fair, just and reasonable; Articles 14 and 21 together constitutionalise procedural fairness rather than confining it to statutory grace.¹¹ Rajesh Agarwal illustrates how this operates outside the conventional courtroom. Classification of a borrower’s account as fraud carried serious civil consequences. The Court read audi alteram partem into the regulatory process, requiring notice, disclosure sufficient to answer the case, an opportunity to represent, and a reasoned order.¹² The doctrinal lesson is not that every serious administrative decision must reproduce a trial. It is that a process producing material civil consequences must offer a real opportunity to affect the outcome.

Amit Iron is therefore central rather than inconvenient to the present argument. The Court rejected the proposition that natural justice always includes a personal hearing and relied on Constitution Bench authority recognising the contextual character of audi alteram partem. A written opportunity accompanied by relevant material and followed by a reasoned order can be adequate. At the same time, the Court reaffirmed disclosure as the rule for a forensic audit report relevant to fraud classification and insisted that the decision-maker engage with the response.¹³ Meaningful human reconsideration is compatible with that reasoning. It is not a demand for an interview with every claimant. It is a demand that a competent human officer be capable of understanding the contested basis of an automated outcome, considering the person’s answer and changing the result when the answer warrants it.

The privacy dimension is complementary. Puttaswamy recognised privacy as a constitutional guarantee tied to dignity, autonomy and informational control; the Aadhaar judgment subjected large-scale State data architecture to legality and proportionality review.¹⁴ Automated administration adds a further concern because decisions may be driven not only by data supplied by the citizen but by inferred attributes, correlations and profiles. Data accuracy and lawful processing are therefore prerequisites to fair decision-making, but privacy law cannot substitute for administrative justice. Even when data are lawfully processed, the inference drawn from them may be contestable, the model may use a poor proxy, or the individual case may contain context that the system cannot recognise. The DPDP framework, once its substantive provisions become operative, may strengthen data rights and fiduciary duties; it still does not answer who must reconsider an adverse State outcome or what reasons that reviewer must give.¹⁵

4 From a Right to a Human Decision to a Right to Meaningful Reconsideration

A categorical right to a human decision has intuitive appeal but a weak constitutional fit. Human decision-makers are neither inherently transparent nor invariably fair. They may be inconsistent, biased, hurried or incapable of processing complex data as accurately as well-designed tools. Aziz Huq’s influential critique is therefore persuasive to an important extent: neither accuracy, participation nor reason-giving supplies a general justification for banning machine decisions as a class.¹⁶ Stefan Schäferling reaches a more rights-protective position in the context of governmental automation, but the disagreement itself reveals why the Indian doctrine should not rest on a romantic premise that human judgment is always superior.¹⁷

The constitutional interest is better described as contestability backed by accountable human authority. Reconsideration matters after an adverse output because some forms of error are visible only from the affected person’s perspective: an income entry is duplicated; a database incorrectly links identities; an employment gap has a lawful explanation; a welfare rule has been applied to an atypical household; a model treats a proxy as determinative although the statutory standard demands individual judgment. The right should therefore arise when automation materially contributes to materially adverse State action, not whenever software is used somewhere in the workflow.

Pragya Prasun v. Union of India offers a narrow but revealing bridge. In the context of inaccessible digital know-your-customer processes, the Supreme Court recognised digital access as part of Article 21 and directed institutions to create a human-review mechanism for rejected applications, with a designated officer capable of overriding automated rejection in appropriate cases.¹⁸ The case arose from disability access and cannot simply be transformed into a general AI holding. Its institutional logic is nevertheless important: a safeguard is meaningful only if the human reviewer possesses authority to depart from the automated result. A reviewer who merely verifies that the system was followed is not reconsidering the decision.

This distinction also explains why “right to explanation” and “right to reconsideration” should not be collapsed. Explanation makes a decision intelligible; reconsideration makes it alterable. Reasons without remedial authority may reveal error but leave the individual trapped by it. Conversely, human review without intelligible reasons risks becoming an unstructured second opinion. Constitutionally adequate contestability requires both: information sufficient to identify the basis of the adverse result, and an institutional route through which a responsible officer can correct it.

5 A Constitutional Framework for High-Impact State Automation

The proposed framework begins with a trigger rather than a sectoral list. Two conditions should be satisfied. First, the automated system must have made a material contribution to the decision. “Material” should capture outputs that substantially shape the result—such as a score, ranking, flag, eligibility determination, predictive classification or recommendation that the official treats as presumptively authoritative. Second, the State action must be materially adverse. The clearest cases are denial, suspension or substantial reduction of welfare benefits; exclusion from public employment or education; tax or regulatory determinations carrying significant financial consequences; and denial of essential public services or legal status. Trivial personalisation, clerical automation and tools with no meaningful effect on the outcome should fall outside the doctrine.

Once the trigger is met, six safeguards follow from existing public-law principles. First, the person should receive notice that an automated system materially contributed to the decision. Hidden automation prevents an affected person from knowing what kind of error to challenge. Second, the State should disclose decision-relevant material. This ordinarily includes the material inputs attributed to the person and the principal factors, threshold or rule that drove the adverse outcome. Full source-code disclosure should not be the default; it can be both technically unhelpful and legitimately restricted. What matters is decision-useful disclosure.

Third, the initial decision or review notice should contain intelligible reasons. “Algorithmic score: high risk” is not a reason in the constitutional sense. The explanation must connect the legal criterion to the facts and identify the role played by automated output. Fourth, the affected person must be permitted to contest factual error, data quality, the applicability of the rule, exceptional circumstances and any legally impermissible proxy or inference. Depending on context, written representation may be sufficient. An oral or video hearing should be required where statute provides it, credibility is material, disability or communication needs make writing inadequate, or the complexity and consequences of the case make oral engagement necessary. This contextual position respects Amit Iron rather than evading it.

Fifth, reconsideration must be undertaken by a designated human officer with both competence and authority. The officer must have access to the information needed to evaluate the automated contribution, understand the legal standard independently of the model, and possess genuine power to depart from the recommendation. Where the agency itself cannot interrogate a vendor system sufficiently to meet these obligations, procurement is the constitutional problem: the State cannot outsource a public function on terms that make lawful review impossible. The 2026 Parliamentary Standing Committee on Personnel, Public Grievances, Law and Justice moved in precisely this institutional direction, recommending government-wide safeguards for AI-assisted administration, proper record-keeping, enterprise-level controls over providers and retention of final decision-making authority by designated human officers.¹⁹

Sixth, the agency must preserve a reviewable record. Courts need not audit model architecture in every dispute, but they must be able to determine what system was used, what version operated, what information entered the decision, what output emerged, how the human officer treated it and why the final conclusion followed. Logging and provenance are therefore not merely technical best practices; in high-impact administration they are conditions for effective judicial review. India’s AI Governance Guidelines already endorse accountability, understandability, human oversight, grievance mechanisms, transparency and algorithmic auditing.²⁰ Constitutional doctrine can give these aspirations a more precise procedural consequence when State action becomes materially adverse.

The framework should contain proportionate exceptions. Emergency action may occasionally precede full process where delay would defeat a legitimate and urgent public purpose, but prompt post-decisional reconsideration should follow. Confidentiality, cybersecurity and trade secrets may justify tailored redaction or controlled disclosure, not a blanket refusal to provide decision-useful reasons. National-security settings may require a different calibration, as Madhyamam itself demonstrates, yet even there secrecy must be justified rather than treated as self-validating. The purpose of the framework is not maximum disclosure. It is minimum constitutional contestability.

6 Comparative Perspective: Lessons from the European Union and United Kingdom

European law offers two useful lessons and one caution. Article 22 of the General Data Protection Regulation protects individuals, subject to exceptions, against certain decisions based solely on automated processing that produce legal or similarly significant effects; where an exception applies, safeguards include human intervention, the opportunity to express a view and the ability to contest the decision.²¹ The “solely automated” formulation, however, risks excessive formalism. In SCHUFA Holding (Scoring), the Court of Justice treated automated credit scoring as an automated individual decision where a third party accorded the score a determining role.²² The functional insight is more valuable for India than the precise statutory test: law should examine what actually drove the outcome.

The EU Artificial Intelligence Act adds a different safeguard. Article 86 grants, for specified high-risk systems and adverse decisions, a right to obtain a clear and meaningful explanation of the role of the AI system and the main elements of the decision. The Regulation’s general application date began on 2 August 2026, subject to its detailed transitional provisions.²³ In CK v. Magistrat der Stadt Wien (Dun & Bradstreet), decided under the GDPR, the Court of Justice further held that meaningful information about automated logic must enable the data subject to understand and challenge the decision; protection of trade secrets cannot convert the right into an empty formality.²⁴ These developments support decision-useful explanation rather than indiscriminate demands for source code.

The United Kingdom now provides a revealing counter-model. Section 80 of the Data (Use and Access) Act 2025 replaced the previous UK GDPR structure with a more permissive regime for significant solely automated decisions in many settings. Yet the reform retained safeguards requiring information, an opportunity to make representations, human intervention and contestation.²⁵ This matters because contestability is not simply a by-product of European hostility to automation. Even a legislature deliberately widening the space for automated decisions preserved an avenue for consequential decisions to return to human judgment.

International instruments reinforce the same orientation without supplying a ready-made Indian rule. The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law requires procedural safeguards where AI significantly affects human rights; its explanatory materials identify human oversight, including ex ante or ex post human review, as a possible safeguard where AI substantially informs or takes a rights-impacting decision.²⁶ UNESCO’s Recommendation on the Ethics of Artificial Intelligence similarly links transparency, explainability and human oversight to human-rights-respecting AI governance.²⁷ India need not transplant these instruments. Their value is confirmatory: contemporary AI governance increasingly treats explanation and review as complementary, not competing, safeguards.

7 Objections, Limits and Institutional Design

Three objections deserve weight. The first is administrative cost. A right to reconsideration could recreate the very delays automation is meant to reduce. The answer lies in the trigger. The proposed doctrine does not require human re-adjudication of every automated output. It attaches to materially adverse action after automation has materially contributed to the result, and ordinarily becomes operative when the person contests that result. Agencies can triage review, standardise disclosure and use automation to assemble the record. What they cannot do is make consequential error practically irreversible.

The second objection is that human review may be performative. Automation bias can lead reviewers to defer to a model even when authorised to depart from it; poorly trained officers may be less competent than the system they review. That objection strengthens rather than defeats the proposed doctrine. “Meaningful” reconsideration must be evaluated institutionally: the reviewer needs training, adequate time, access to the relevant material and authority to reject the output. Aggregate reversal rates, recurring grounds of correction and system-level error should inform internal audit. Human review is not valuable because human intuition is sacred; it is valuable because public law requires an accountable point at which legally relevant reasons and individual context can change the State’s position.

The third objection concerns secrecy and technical complexity. Some models are proprietary, some are difficult even for developers to interpret, and some public systems implicate cybersecurity or fraud-prevention concerns. A constitutional right framed as full technical transparency would be fragile for precisely these reasons. The better requirement is functional disclosure. The affected person needs information sufficient to contest the adverse application to her; the reviewing officer and, where necessary, the court may require deeper documentation under confidentiality protections. The CJEU’s recent insistence that meaningful explanation need not amount to disclosure of source code but must still enable challenge provides a useful comparative illustration.²⁸

The larger institutional lesson is that constitutional compliance must begin before deployment. Agencies procuring high-impact systems should contract for access to logs, documentation, version history, audit support and explanations adequate to discharge public-law duties. A vendor’s confidentiality clause cannot determine the content of Article 14. Nor should an agency deploy a system whose outputs it is incapable of defending through reasons. Ex ante governance and ex post reconsideration serve different functions: audits identify systemic defects; reconsideration protects the person whose case has already gone wrong. Neither is a substitute for the other.

8 Conclusion

Automated administration does not require Indian constitutional law to invent a technological exception to familiar principles. Nor does it justify a categorical constitutional preference for humans over machines. The durable concern is the exercise of public power without a meaningful route from adverse output back to accountable judgment.

Articles 14 and 21 already provide the necessary doctrinal materials: non-arbitrariness, fair procedure, disclosure of relevant material, reasoned decision-making and effective judicial review. The post-2021 cases make those commitments unusually clear, while Amit Iron prevents them from hardening into an indiscriminate oral-hearing rule. Read together, they support a narrower proposition suited to automated government. Where an automated system materially contributes to materially adverse State action, the affected person should be entitled to know that automation mattered, understand the principal basis of the result, contest error and context, and obtain reconsideration from a human officer with genuine authority to depart from the system.

The proposed material contribution–material adversity test also avoids a weakness visible in foreign data-protection law. Constitutional accountability should not disappear because a human has formally touched the file. What matters is whether the automated system exercised real decisional influence and whether the resulting State action seriously affected the person. India’s emerging AI-governance policy already recognises human control, accountability, audit and grievance redress; Parliament’s 2026 committee work likewise insists that designated human officers retain final authority in public administration. The constitutional task is to convert those governance instincts into a disciplined rule of contestability. Efficiency may justify automation. It does not justify an administrative dead end.