The Speaking Algorithm: Constitutionalising Reasons for Automated Administrative Decisions in India

Abstract 

The digitisation of government is moving beyond electronic record-keeping and online service  delivery toward data-driven administration. Algorithms are increasingly capable of sorting  applications, detecting anomalies, identifying risk, prioritising cases, and supporting decisions  concerning access to public benefits and services. This development creates a distinctive  administrative-law problem. A traditional administrative order can be tested through the reasons  recorded by the decision-maker. An algorithmic decision may instead present the citizen with a  result—“rejected,” “flagged,” “ineligible,” or “high risk”—without revealing how the result was  produced, what information materially influenced it, or whether a human authority independently  considered the case. 

Indian law does not yet contain a comprehensive statutory regime governing automated  administrative decision-making. Nevertheless, the Constitution and existing administrative-law  doctrine provide a substantial normative foundation for regulating such decisions. Article 14  constrains arbitrariness; Article 21 requires fair, just and reasonable procedures where protected  interests are affected; natural justice ordinarily requires meaningful participation where adverse  decisions are contemplated; and the Supreme Court has repeatedly stressed the importance of  reasoned or “speaking” orders. The recognition of informational privacy under Article 21 adds  another dimension, particularly where automated systems depend on extensive data processing.  The Digital Personal Data Protection Act, 2023, although enacted, is subject to phased  commencement and therefore cannot presently be treated as a fully operational answer to  algorithmic public decision-making. The India AI Governance Guidelines, released in November  2025, provide principles of fairness, accountability and understandable-by-design systems, but  remain principally a governance framework rather than a comprehensive source of enforceable  procedural rights. 

This article argues that Indian administrative law should develop a doctrine of algorithmic  reason-giving. Such a doctrine would not require government to disclose source code or  proprietary technical architecture in every case. It would instead require affected persons, at an  appropriate level, to know the legal rule applied, the principal factual inputs relied upon, the  material role played by the algorithm, the nature of human review, and the reasons why the individual’s circumstances satisfied or failed the governing criteria. The central claim is simple:  automation may change the mechanics of administration, but it cannot dilute the  constitutional obligation to make public power intelligible, reviewable and contestable. 

Keywords: Administrative Law, Article 14, Article 21, Algorithmic Decision-Making, Artificial  Intelligence, Natural Justice, Speaking Orders, Right to Reasons, Data Protection, Explainability,  Constitutional Governance 

Introduction 

Imagine that a citizen applies for a government benefit for which she appears, on the face of the  statutory eligibility criteria, to qualify. Her application is processed digitally. An automated  system cross-references multiple databases, detects what it considers an inconsistency, assigns a  risk score, and recommends rejection. A government officer then approves the recommendation  with a standard electronic message: “Application rejected as per system-generated eligibility  assessment.” 

The citizen asks a straightforward question: Why? 

The answer matters for more than curiosity. Without knowing what went wrong, she cannot  meaningfully challenge the decision. She may not know whether the system relied on an  incorrect address, outdated income data, duplicate records, an erroneous identity match, a proxy  variable, or an inference that no ordinary applicant could reasonably anticipate. A court  examining the decision may confront a further difficulty: the official file contains the final  outcome but not a comprehensible account of the computational path that led to it. 

This is the emerging problem of automated administration. 

Governments have long used rules, forms, classifications and statistical techniques. Algorithms  are therefore not legally interesting merely because they involve computers. Their significance  lies in what happens when computational systems begin to influence the exercise of public  power. The State may use algorithmic tools to allocate benefits, detect tax irregularities, identify  suspicious transactions, manage public-health interventions, administer education systems,  process immigration and identity records, or determine which cases deserve further investigation. 

The traditional language of administrative law was developed around human decision-makers.  Yet its central concerns were never exclusively about human psychology. They were about  power, reasons, fairness, equality, participation and reviewability. Those principles are precisely what become difficult when a decision is generated through a technological process  that may be complex, probabilistic, continuously updated and partly opaque. 

India has not enacted a general law establishing comprehensive procedural rights for persons  subject to automated public-sector decisions. Nor has the Supreme Court, as of September 2026,  laid down a general constitutional test specifically for government decisions produced by  artificial intelligence or machine-learning systems. The legal task, therefore, is not to pretend that  a settled doctrine already exists. It is to ask whether existing doctrine can—and should—be  adapted to the new institutional reality. 

This article argues that it can. More importantly, it argues that the duty to give reasons must  evolve from a duty merely to record a conclusion into a duty to provide sufficient reasons  for an affected person to understand and contest an algorithmically mediated decision. 

The distinction is crucial. “The system rejected your application” is not a reason. It is a  description of an event.

The Research Problem 

The core legal problem may be stated as follows: 

When a public authority relies materially on an algorithm in reaching an adverse decision  affecting a person’s rights, interests, benefits or legitimate expectations, what explanation is  constitutionally required, and who bears responsibility for providing it? 

Three difficulties follow. 

First, automated decisions may be difficult to explain even to their operators. A machine learning model can identify statistical relationships without producing a human-readable  narrative corresponding to conventional legal reasoning. 

Second, the algorithm is rarely itself the formal legal decision-maker. The final order may be  signed by an official who did not design, train or fully understand the system. This raises an  issue of institutional accountability: can an authority discharge a legal duty to give reasons by  merely endorsing a computational output? 

Third, transparency itself has limits. Government may have legitimate interests in cybersecurity,  privacy, prevention of fraud, protection of third-party data, or confidentiality of security systems.  A constitutional right to reasons therefore cannot simply become a right to obtain source code in  every dispute.

The challenge is to identify a legally workable middle ground between algorithmic opacity and  technological over-disclosure. 

III. Research Questions and Objectives 

The article addresses four questions: 

1. What constitutional and administrative-law principles presently regulate reason-giving,  natural justice, equality and privacy in Indian public administration? 
2. How do these principles apply when public decisions are materially influenced by  algorithms? 
3. Does existing data-protection law adequately protect procedural interests arising from  algorithmic administrative decision-making? 
4. What form should an Indian doctrine of algorithmic reason-giving take? 

Its principal objective is not to advocate a blanket prohibition on automated administration.  Algorithms can improve consistency, speed and the capacity of government to process large  volumes of information. The objective is instead to develop legal conditions under which  efficiency does not become a substitute for accountability. 

Methodology 

This article adopts a doctrinal and comparative methodology. It examines the Constitution of  India, the Supreme Court’s jurisprudence on Article 14, Article 21, natural justice, privacy and reasons in administrative decisions, the Right to Information Act, 2005, the Digital Personal Data  Protection Act, 2023, the Digital Personal Data Protection Rules, 2025, and the India AI  Governance Guidelines released by the Ministry of Electronics and Information Technology in  2025. 

The article also draws comparatively on international developments, particularly European data protection and artificial-intelligence governance instruments, not to transplant foreign law into  India but to identify regulatory techniques relevant to a constitutional administrative system. 

From Human Decision-Makers to Computational  Administration 

An algorithm, in its simplest sense, is a structured procedure for producing an output from  inputs. Not every algorithm uses artificial intelligence, and not every automated government  process is genuinely autonomous. A tax portal calculating a liability according to explicit  statutory rules is legally different from a machine-learning model predicting the likelihood that a  taxpayer’s return requires scrutiny. 

This distinction is important because legal risk increases when the relationship between input  and outcome becomes less transparent. 

Consider three levels of automation: 

First, mechanical automation. The law supplies a determinate formula and software merely  performs the calculation. 

Second, rules-based decision support. The system applies programmed conditions and  generates a recommendation, while an officer formally decides. 

Third, predictive or machine-learning systems. The system may identify correlations or  generate risk scores based on historical data, potentially using variables whose relationship to the  legal criterion is not obvious. 

The third category presents the most difficult administrative-law questions. A model trained on  historical enforcement data may reproduce historical patterns. If past enforcement  disproportionately targeted a particular class or locality, the system may learn that pattern and  treat it as predictive of future non-compliance. The resulting discrimination may be statistically  difficult to detect and individually difficult to prove. 

Yet the constitutional problem does not depend on whether a programmer deliberately  discriminated. Article 14 is concerned with the effect and rationality of State action, not merely  the subjective intention of the person who wrote the software. 

The technical structure of the decision therefore cannot determine the applicable legal standard.  If the outcome is an exercise of public power, constitutional scrutiny follows the function, not  the software architecture.

The Constitutional Framework 

Article 14 and the Constitutional Prohibition on Arbitrariness 

Article 14 guarantees equality before the law and equal protection of the laws. Indian  constitutional doctrine has progressively moved beyond a narrow conception of classification  toward a broader concern with arbitrary State action. 

In E.P. Royappa v. State of Tamil Nadu, the Supreme Court famously connected equality with non-arbitrariness. Later doctrine, including Shayara Bano v. Union of India, recognised “manifest arbitrariness” as a constitutional ground of invalidity in the appropriate context. The  Court described arbitrariness as conduct lacking an adequate determining principle and involving  irrationality or disproportionality. 

This jurisprudence is particularly relevant to algorithmic administration. 

An algorithm is not automatically non-arbitrary merely because it produces the same result every  time. Consistency is not the same as constitutional reasonableness. A system may  consistently apply an irrational proxy, perpetuate historically distorted data, ignore relevant  factors, or assign vastly different consequences to similarly situated persons. 

Indeed, automation can make arbitrary administration harder to see. Human discretion leaves  identifiable trails: notes, conversations, instructions and reasons. Computational systems may  transform discretion into apparently neutral scores. The numerical character of an outcome can  create a false impression of objectivity. 

Article 14 should resist that tendency. A score is not self-justifying. Public authorities must be  able to explain why the variables selected, the method used and the outcome produced bear a  rational relationship to the legal objective. 

This leads to an important doctrinal proposition: algorithmic opacity should not receive  greater constitutional deference than human opacity. Government cannot escape Article 14  simply by purchasing or deploying a system it does not understand.

Article 21 and Fair Procedure 

Article 21 provides that no person shall be deprived of life or personal liberty except according  to procedure established by law. The transformation of Article 21 in Maneka Gandhi v. Union of India established that the relevant procedure must be fair, just and reasonable rather than arbitrary or oppressive. Maneka Gandhi v. Union of India,   Although not every government algorithmic decision will implicate life or personal liberty in the  same manner, modern Article 21 jurisprudence extends significantly into autonomy, dignity and  informational privacy. 

The constitutional recognition of privacy in Justice K.S. Puttaswamy (Retd.) v. Union of India is  particularly important. The nine-judge Bench recognised privacy as a constitutionally protected  right connected with liberty, dignity and autonomy and articulated proportionality as a standard for assessing permissible restrictions. 

Algorithmic government frequently depends upon the collection, aggregation and inference of  personal information. A system may not merely process information supplied by an individual; it  may infer characteristics from multiple data points. Consequently, the legal issue is not only  whether data were lawfully collected. It is also what governmental power is being exercised on  the basis of that data. 

Suppose a benefits algorithm uses an applicant’s transaction history, location information and  household records to infer “economic reliability.” Even where every individual data point was  lawfully obtained, the resulting inference may significantly affect the person’s autonomy and  access to State support. Privacy therefore intersects with procedural fairness. 

The constitutional concern may be framed this way: if the State uses intimate or inferential  information to make a consequential decision, should the affected person be entitled to  know enough about that use to challenge its accuracy and legality? 

In a constitutional democracy, the answer should ordinarily be yes. 

VII. Natural Justice in the Age of Automation 

The classic principles of natural justice include, most prominently, the rule against bias and audi  alteram partem—the opportunity to be heard. The Supreme Court has repeatedly rejected the  simplistic idea that administrative decisions are necessarily beyond these procedural  requirements. 

In Swadeshi Cotton Mills Co. Ltd. v. Union of India, the Court reaffirmed the relevance of  natural justice to administrative action and treated the opportunity to respond as an important safeguard even where governmental power is being exercised administratively. Swadeshi Cotton  Mills Co. Ltd. v. Union of India, Automation does not make a hearing less necessary. In some situations, it makes the content of  the hearing more important. 

Imagine an individual is told that her application was rejected because the system detected a  “mismatch.” What can she possibly say in response unless she knows what was mismatched? 

A hearing is meaningful only if the person can engage with the material that materially  influenced the proposed decision. A procedural ritual in which the State says, “The computer  says no,” and invites the citizen to respond without revealing the basis of the conclusion would  be a hollow version of audi alteram partem. 

This suggests that natural justice in an algorithmic setting has two linked dimensions: procedural participation—the person must have an opportunity to respond; and  epistemic access—the person must receive enough information to make that response  meaningful. 

The second principle is increasingly unavoidable. A hearing without intelligibility is  participation in form rather than substance. 

VIII. The Doctrine of the Speaking Order 

The strongest doctrinal bridge to algorithmic decision-making may be found in India’s  established requirement that administrative and quasi-judicial decisions, particularly those  prejudicially affecting rights or interests, should give reasons. 

In Kranti Associates (P) Ltd. v. Masood Ahmed Khan, the Supreme Court reaffirmed the  importance of reasons and emphasised that an order affecting parties should “speak” rather than  present an unexplained conclusion. The Court connected reason-giving with fairness,  accountability and meaningful judicial review. Kranti Associates (P) Ltd. v. Masood Ahmed  Khan,  

The related principle in Mohinder Singh Gill v. Chief Election Commissioner is that a statutory  order must ordinarily stand or fall on the reasons stated in the order; later explanations cannot  simply repair an inadequately reasoned decision. Mohinder Singh Gill v. Chief Election  Commissioner, The Supreme Court has continued to reaffirm this proposition. 

These doctrines acquire a new significance when the underlying decision is computational. 

A conventional official may write: 

“The application is rejected because the applicant does not satisfy Condition X. The relevant  record shows Y and Z.” 

An algorithm cannot simply replace this with: 

“The model-generated score is below threshold.” 

That statement identifies the output but not the legal reasoning. 

A proper administrative explanation should instead connect law → relevant facts →  computational process → human decision → outcome. 

That does not mean the officer must reproduce mathematical equations in every individual order.  The law generally requires reasons relevant to the decision, not a technical manual. But where  the algorithm materially contributes to the adverse outcome, the explanation must disclose its legally significant role. 

The emerging distinction should therefore be between technical explainability and legal  explainability. 

Technical explainability asks: “Can an engineer reconstruct why the model generated this  prediction?” 

Legal explainability asks: “Has the State disclosed enough for the affected person and a  reviewing court to determine whether the decision was lawful, rational, procedurally fair and  non-discriminatory?” 

The second question is the constitutionally relevant one.

Does the Digital Personal Data Protection Act Solve the  Problem? 

The Digital Personal Data Protection Act, 2023 (“DPDP Act”) is an important part of India’s  contemporary data-governance framework. It establishes provisions concerning lawful  processing, notice, consent, legitimate uses, obligations of Data Fiduciaries, rights of Data  Principals, grievance redressal and exemptions. 

But two cautions are necessary. 

First, the DPDP Act is not a comprehensive automated-decision-making statute. It does not  simply reproduce a broad European-style general right not to be subject to automated decisions.  Its structure is fundamentally concerned with the governance and protection of digital personal  data rather than with creating an all-purpose administrative-procedure code for algorithmic  public decisions. 

Second, and particularly important as of the date of this article, the Act has been brought into  force in stages. The Central Government’s 13 November 2025 commencement notification  brought certain institutional and miscellaneous provisions into force immediately, specified  further provisions for one year later, and provided that a substantial group—including Sections 3  to 5, Sections 7 to 10 and Sections 11 to 17—would come into force eighteen months after  publication. 

Accordingly, on 9 September 2026, the substantive rights and obligations contained in that eighteen-month tranche should not be described as fully operative. The same caution applies to  the Digital Personal Data Protection Rules, 2025: Rules 1, 2 and 17–21 commenced  immediately, Rule 4 after one year, while Rules 3, 5–16, 22 and 23 were prescribed to  commence eighteen months after publication. 

This phased architecture illustrates why constitutional administrative law remains indispensable.  A citizen confronted with an automated public decision cannot be left procedurally unprotected  simply because the data-governance statute is incomplete, inapplicable to a particular processing  activity, or not yet fully commenced. 

Data protection and administrative fairness overlap, but they are not substitutes.

The India AI Governance Guidelines: Progress Without a  Complete Remedy 

India’s regulatory approach to artificial intelligence has also developed significantly. On 5  November 2025, MeitY released the India AI Governance Guidelines, framing a national,  principle-based approach to safe, responsible and inclusive AI deployment. The framework  identifies seven guiding principles, including trust, people-first governance, fairness and equity,  accountability, “understandable by design,” and safety, resilience and sustainability. 

The guidelines are significant precisely because “understandable by design” places  comprehensibility within the architecture of responsible AI. India’s Responsible AI materials  separately describe explainability in terms of enabling users to seek explanations for significant  decisions and providing those explanations in human-understandable language. 

Yet the legal status of this framework must not be overstated. Government itself has described  the Guidelines as relying substantially on existing laws and sectoral regulation rather than  creating a new general AI statute with independent appeals or oversight machinery. 

Thus, the Guidelines can inform statutory interpretation, administrative standards and good  governance, but they cannot simply be treated as if they themselves created enforceable  constitutional rights. 

This distinction is important academically and legally. Soft law can shape the content of  reasonableness without replacing hard law. 

A Proposed Doctrine of Algorithmic Reason-Giving 

The existing principles of Article 14, Article 21, natural justice and reasoned decision-making  can be developed into a modest doctrine of algorithmic reason-giving. 

The doctrine should rest upon five minimum requirements. 

1. Notice of Material Algorithmic Involvement 

Where an algorithm materially influences an adverse public decision, the affected person should  ordinarily be informed that such a system was used. 

A citizen should not discover only during litigation that a supposedly human decision was  substantially determined by a risk score generated by software. 

This requirement is conceptually simple: the person must know the nature of the decision making process that affected her. 

2. Disclosure of the Legal Criterion 

The government must identify the statutory, regulatory or policy criterion against which the  person was assessed. 

This prevents the algorithm itself from becoming the effective source of law. A model may  operationalise an existing legal criterion; it should not silently invent new eligibility  requirements.

3. Meaningful Explanation of Principal Inputs 

The affected person should receive information about the material categories of facts or  indicators relied upon. 

This does not require publication of every variable in every system. It does require enough  information to permit error correction and legal challenge. 

If a benefit was rejected because of an alleged income discrepancy, the person must be told that  income data formed part of the assessment and, where appropriate, what record produced the  discrepancy. 

4. Human Accountability 

The existence of human review must mean more than placing a signature after an automated  recommendation. 

An officer who formally decides must be capable of considering an individual’s response,  correcting obvious errors and departing from the automated recommendation where law or facts  require it. 

Otherwise the “human in the loop” risks becoming a constitutional fiction.

5. A Route to Reconsideration 

An automated decision affecting significant rights or interests should be contestable before a  competent human authority. 

This is particularly important because statistical systems operate through generalisations while  legal adjudication often turns on individual circumstances. 

A person may be an outlier in statistical terms and nevertheless fully entitled in legal terms. 

XII. Why Source-Code Disclosure Should Not Be the Default 

A serious legal framework must acknowledge the strongest counterargument: complete  algorithmic transparency may be impractical or undesirable.

Source code can expose cybersecurity vulnerabilities. Proprietary systems may contain  commercially confidential elements. Publishing every technical detail can also overwhelm  ordinary citizens rather than empower them. 

For that reason, “explainability” should not mean unlimited disclosure. 

The legally relevant standard should be sufficient explanation for the exercise of rights, not  maximal technical disclosure. 

A court may therefore distinguish between: 

• source code; 
• system architecture; 
• training-data documentation; 
• variables and categories materially used; 
• decision thresholds; 
• individual-level factual inputs; 
• the legal rule implemented; 
• audit records; and 
• the human official’s final reasons. 

Different information may appropriately receive different degrees of protection. 

What cannot be accepted, however, is a categorical claim that procurement contracts, proprietary  technology or computational complexity immunise the State from explaining its decisions. 

The State is not a private consumer trapped by a vendor’s terms of service. If government  chooses a technological system to exercise public power, the constitutional responsibility  remains governmental. 

XIII. Judicial Review and the Problem of the “Black Box” 

Judicial review provides a natural institutional setting for developing algorithmic administrative  law. 

A court traditionally asks whether the authority acted within jurisdiction, complied with statutory  requirements, observed procedural fairness, considered relevant matters, ignored irrelevant ones,  and avoided arbitrariness or mala fides. 

Each of those questions can be translated into the algorithmic setting.

Relevant considerations: What data and variables influenced the result? 

Irrelevant considerations: Did the system use proxies that have no legitimate relationship to the  statutory objective? 

Natural justice: Was the affected person given meaningful notice and opportunity to challenge  decisive factual assumptions? 

Reasonableness: Is there a rational relationship between the model’s inputs and the statutory  purpose? 

Equality: Does the system produce unjustified disparate effects on similarly situated classes? 

Proportionality: Is reliance on extensive personal data justified by the legitimate governmental  objective, and is the intrusion proportionate? 

The Supreme Court’s privacy jurisprudence gives proportionality particular relevance where  data-intensive State action implicates Article 21. The Court has described legality, legitimate  State aim and proportionality as central requirements for constitutionally justified privacy  

intrusions. 

Algorithms should therefore be subjected to ordinary judicial review, not a special presumption  of technological correctness. 

The court may have to examine technical evidence more frequently than before. But complexity  should affect the method of review, not the availability of review. 

XIV. International Comparative Perspective 

Comparative law reinforces this direction. 

European data-protection law has long confronted the problem of significant decisions produced  through automated processing, while the European Union’s AI regulatory framework has moved  toward a risk-based system governing high-risk uses. 

More broadly, the Council of Europe Framework Convention on Artificial Intelligence and  Human Rights, Democracy and the Rule of Law represents a significant international  development. Adopted in 2024 and opened for signature on 5 September 2024, it is the first  international legally binding treaty specifically addressing AI through the combined lenses of  human rights, democracy and the rule of law. It expressly emphasises transparency, oversight, accountability, equality, privacy, procedural safeguards and the ability of affected persons to challenge decisions substantially based on AI systems. 

The comparative lesson is not that India must copy another jurisdiction’s legislation. Rather, it is  that a rights-based AI framework increasingly treats contestability as a central legal  requirement. 

That concept fits comfortably within Indian constitutionalism. 

The right to challenge State action is meaningful only when the citizen can identify the basis on  which the State acted. 

Challenges in the Indian Context 

A number of institutional difficulties remain. 

Fragmented Government Technology Procurement 

Government agencies may procure algorithmic systems from external vendors. Contractual  dependence can produce uncertainty about data ownership, audit rights and access to technical  documentation. 

Procurement rules should therefore treat explainability, auditability and public-law compliance  as substantive contractual requirements rather than optional technical features. 

Data Quality 

Many algorithmic harms begin before the algorithm is run. A perfectly designed model cannot  rescue inaccurate or outdated government databases. 

A constitutional right to challenge an automated decision must therefore include a practical  capacity to challenge the factual inputs underlying it. 

Administrative Capacity 

Judges and civil servants cannot be expected to become specialist programmers. India therefore  needs multidisciplinary capacity—lawyers, administrators, statisticians, computer scientists and  auditors capable of working together. 

Automation Bias 

Officials may develop excessive trust in system outputs because the output appears  mathematically objective. This phenomenon creates a danger that human review will become  ceremonial. 

Training and institutional design must therefore encourage officials to ask whether the automated  recommendation is legally justified, not merely whether the software operated correctly. 

Vulnerable Citizens 

Procedural complexity disproportionately harms those with limited digital literacy. An  explanation available through a technical portal is not necessarily an accessible explanation. 

The constitutional standard of procedural fairness must therefore be evaluated from the  perspective of the affected person, not only the sophistication of the administration. 

XVI. Recommendations 

India does not necessarily need, at this stage, an enormous omnibus AI statute governing every  conceivable use of artificial intelligence. A more immediate and constitutionally grounded  approach would be to develop sector-specific safeguards around consequential public decisions. 

First, executive authorities should adopt a mandatory algorithmic impact assessment for  systems capable of materially affecting rights, benefits, entitlements or significant governmental  burdens. 

Second, government agencies should maintain an algorithmic decision record containing the  legal basis of the system, purpose, principal data categories, model version, responsible  authority, audit history and significant changes. 

Third, affected persons should receive an individualised statement of reasons where an  automated or algorithmically assisted decision is materially adverse. 

Fourth, significant decisions should ordinarily retain a genuine human reconsideration  mechanism. 

Fifth, courts and tribunals should be empowered to require production of relevant technical and  audit material under suitable confidentiality safeguards where necessary for effective judicial  review.

Sixth, public procurement frameworks should impose clear obligations on vendors to provide  sufficient documentation for legal accountability. 

Seventh, Parliament or the competent rule-making authorities should consider an explicit  procedural framework for high-impact automated public-sector decision-making, drawing  upon Article 14, Article 21, natural justice and the emerging norms of responsible AI. 

Finally, India should distinguish between two questions that are often conflated: “Is the  algorithm transparent?” and “Is the governmental decision contestable?” The second is the  constitutional question. Perfect transparency may be unattainable. Effective contestability is not. 

XVII. The Central Principle: The State Cannot Delegate  Constitutional Accountability 

The most important issue is ultimately institutional. 

Suppose a government department says that an external vendor supplied the model. The vendor  says that the model is proprietary. The official says that she merely relied on the output. Each  actor points elsewhere when the citizen asks who is responsible. 

Constitutional law should reject this fragmentation. 

Public power does not cease to be public merely because its exercise has been technologically  outsourced. 

The State may delegate technical work. It cannot delegate away constitutional accountability. 

This principle follows from the structure of judicial review itself. Government bodies exercise  statutory powers because the law authorises them to do so. A computational system possesses no  independent constitutional authority. Its output becomes legally relevant only because a public  institution chooses to rely upon it. 

Therefore, the responsible public authority must remain answerable for the legality of the  decision, including the integrity of the process by which it was reached.

XVIII. Conclusion 

The central mistake in debates over algorithmic administration is to assume that the legal  challenge begins with artificial intelligence. 

It does not. 

The challenge begins with an older question: What must the State tell a person when it  exercises public power against her? 

Indian administrative law already contains a powerful answer. Public decisions must not be  arbitrary. Fair procedure matters. Natural justice can apply to administrative action. Decisions  affecting individuals should ordinarily be reasoned. Courts require intelligible reasons in order to  review governmental power. Privacy and dignity place constitutional limits on data-driven State  action. 

Algorithms do not eliminate these principles. They test whether the legal system is prepared to  apply them when decision-making is distributed across databases, models, software and officials. 

The constitutional objective should not be to make every algorithm perfectly transparent. Nor  should it be to prohibit government from using technological tools. The appropriate objective is  more modest and more important: no person should lose a legal entitlement or suffer a  significant adverse governmental consequence merely because the State has placed a  computer between itself and its duty to give reasons. 

A speaking order has traditionally meant an order that explains itself. In an algorithmic  administration, that concept must become slightly more demanding. 

The decision must explain not merely what happened, but enough about why it happened to  allow the person affected—and ultimately the reviewing court—to determine whether the  exercise of public power was lawful. 

The future of administrative law may therefore not belong to the “automated decision.” 

It should belong to the speaking algorithm: a technological system whose use remains  embedded in a human institution that can explain, justify, reconsider and defend the  exercise of public power. 

That is not an obstacle to digital government. It is the constitutional price of making digital  government legitimate.

References / Selected Bibliography 

Constitutional and Statutory Sources 

1. Constitution of India, arts. 14, 19, 21, 32, 226.  
2. Right to Information Act, 2005, §§ 3–4, 8, 11. The Act expressly places obligations on  public authorities under Section 4 and establishes the statutory architecture for access to  information.  
3. Digital Personal Data Protection Act, 2023, §§ 3–17, 18–44.  
4. Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), Ministry of Electronics and  Information Technology, Nov. 13, 2025.  
5. Digital Personal Data Protection Act, 2023, Commencement Notification, G.S.R. 843(E),  Ministry of Electronics and Information Technology, Nov. 13, 2025.  
6. Ministry of Electronics and Information Technology, India AI Governance Guidelines:  Enabling Safe and Trusted AI Innovation (Nov. 5, 2025). The Government describes the  framework as principle-based and centred on trust, people-first AI, fairness,  accountability and understandable-by-design systems.  

Judicial Authorities 

7. Maneka Gandhi v. Union of India, (1978) 1 SCC 248.  
8. E.P. Royappa v. State of Tamil Nadu, (1974) 4 SCC 3.  
9. Mohinder Singh Gill v. Chief Election Commissioner, (1978) 1 SCC 405. The Court held  that an administrative order must ordinarily be judged by the reasons stated in it and  cannot be retrospectively supplemented by new reasons.  
10. Swadeshi Cotton Mills Co. Ltd. v. Union of India, (1981) 1 SCC 664. 
11. Kranti Associates (P) Ltd. v. Masood Ahmed Khan, (2010) 9 SCC 496. The Supreme  Court reiterated the importance of reasons in administrative and quasi-judicial decision making.  
12. Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1.  
13. Justice K.S. Puttaswamy (Retd.) v. Union of India, (2019) 1 SCC 1.  
14. Shayara Bano v. Union of India, (2017) 9 SCC 1. The judgment is an important authority  in the Supreme Court’s development of the manifest-arbitrariness doctrine. 
15.  C.P.I.O., Supreme Court of India v. Subhash Chandra Agarwal, (2019) 16 SCC 82. The  decision examines transparency, privacy and access to information under the RTI  framework.

1 Right to Information Act 2005, ss 3?4, 8, 11; Digital Personal Data Protection Act 2023; Digital Personal Data Protection Rules 2025; Ministry of Electronics and Information Technology, India AI Governance Guidelines (5 November 2025).

2 Constitution of India, art 14; E P Royappa v State of Tamil Nadu (1974) 4 SCC 3; Shayara Bano v Union of India (2017) 9 SCC 1.

3 Maneka Gandhi v Union of India (1978) 1 SCC 248; Constitution of India, art 21. 

4 Justice K S Puttaswamy (Retd) v Union of India (2017) 10 SCC 1. 

5 Swadeshi Cotton Mills Co Ltd v Union of India (1981) 1 SCC 664.

6 Kranti Associates (P) Ltd v Masood Ahmed Khan (2010) 9 SCC 496. 

7 Mohinder Singh Gill v Chief Election Commissioner (1978) 1 SCC 405.

8 Kranti Associates (P) Ltd v Masood Ahmed Khan (2010) 9 SCC 496.

9 Digital Personal Data Protection Act 2023, s 1; Digital Personal Data Protection Act 2023, Commencement Notification, GSR 843(E), MeitY (13 November 2025); Digital Personal Data Protection Rules 2025, GSR 846(E), MeitY (13 November 2025). 

10 Ministry of Electronics and Information Technology, India AI Governance Guidelines: Enabling Safe and Trusted AI Innovation (5 November 2025).10 Ministry of Electronics and Information Technology, India AI Governance Guidelines: Enabling Safe and Trusted AI Innovation (5 November 2025). 

11 Justice K S Puttaswamy (Retd) v Union of India (2017) 10 SCC 1.

12 Council of Europe, Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (opened for signature 5 September 2024).