Judicial Disengagement and the Automated Welfare State

1 Introduction

The State¹ in India is now increasingly relying on automated systems for the administration of welfare policies. The administrative tasks that were earlier carried out end-to-end by humans now have an algorithm to share the burden. These algorithmic systems are being used to cross-check databases, to flag duplications across the recorded entries, to flag potential frauds and build risk profiles, and to require biometric matching before welfare benefits can be accessed, among other things. These are not mere software upgrades; rather, these represent a systematic automation in governance. Decisions as basic as whether a family will receive food rations, whether wages will be paid to a labourer, or whether an elderly citizen will receive her pension, which until now required human adjudication, are being left to be decided by an algorithm.

To break it down, the automated systems are the platform or point of access to avail any welfare benefit; algorithmic processing is the procedural work done by the AI to process a person’s application to verify the information it contains against government datasets; and outputs are the decisions given post such processing accepting or rejecting such applications. This structure has altered the entire ecosystem of welfare delivery, and has at times detrimentally affected the fundamental rights of citizens.

Now this alteration has also given rise to a constitutional problem in India. It is settled law under Article 21 that any procedure depriving any person of her right to life and personal liberty must be just, fair and reasonable. It is also well understood that the right to live with dignity rests on the shoulders of policies of a welfare state and their equitable benefit delivery. However, the established jurisprudence presumes that a procedure governing access to such entitlements is open to inspection. When a citizen is wrongfully excluded by algorithmic processing, the exclusion is a mathematical output rather than a reasoned administrative decision. Such a procedure is not aligned with the constitutional standards of transparency in decision making.

The Supreme Court of India was recently asked to address this exact issue in Narendra Kumar Goswami v. Union of India² on August 13, 2026. The Court, however, declined to decide the petition; observations from the Bench, as reported, indicated two grounds. First, the Court considered the issue of algorithmic governance as a technical one requiring special expertise. Second, the Court considered the issue as belonging to the policy domain of the executive and the legislature. This paper argues that neither ground sits comfortably with the Court’s own historical jurisprudence. It also argues that the constitutional standard required to regulate these systems is the one that the Court already possesses.

It is to be noted that the original petition in Goswami covered four distinct areas of state automation, but this paper restricts its scope to analyse the issue associated with welfare administration. The remaining domains raise distinct questions that require a different constitutional treatment.

Methodologically, this paper conducts a doctrinal analysis, supported by a single comparative reference point. The paper proceeds in six parts. It starts by tracing the jurisprudential shift under Article 21 and the gap it reveals regarding automation. It then sets out the factual matrix of Goswami, what was asked and what was refused. It then tests the grounds of refusal against the Court’s own pattern of judicial intervention in cases of legislative vacuum. It then turns to automated welfare systems presently in operation, and the exclusions they have produced in the absence of any legislative or judicial supervision. Next, it makes a comparative analysis contrasting Goswami with a Dutch case arising from a similar factual matrix. Finally, it argues that the Court need not devise new constitutional safeguards because the existing proportionality test is a sufficient mechanism to hold the State accountable within such an ecosystem.

2 From A. K. Gopalan to Maneka Gandhi, and the Contemporary Algorithmic Gap

2.1 The Shift from Gopalan to Maneka Gandhi

In 1950, the Supreme Court took a strictly textual approach to read “the procedure established by law” under Article 21. As long as the procedure was prescribed by a validly enacted law, the Court would not enquire into its fairness.³ Decades later, in 1978, this position was fundamentally changed. The Court ruled that any procedure depriving a person of life or personal liberty must be just, fair, and reasonable, rather than arbitrary or oppressive.⁴ This jurisprudential shift has been thoroughly studied but what matters the most today are the standards that Maneka Gandhi proposed. An entire procedure can actually be examined by the courts to assess its fairness; however, this approach presumes the presence of an identifiable human decision maker, reasons for that decision, and a factual record. Thus, the test of fairness is applied by examining how a particular decision was reached. Yet all these legitimate assumptions give out when procedure is automated.

2.2 What the Fairness Standard Assumes

When an algorithmic system gives decisions, it is simply a technologically generated output. Such an output cannot be fairly assessed using the same approach. There is no identified decision-maker who can be held accountable for erroneous outputs; the output is rarely supplemented by a statement of reasons; and there is no accessible record against which the decision may be examined. The term “technological due process” has been used by legal scholars to describe this problem. Danielle Keats Citron observed that automated administrative systems collapse individual adjudication into rigid rulemaking while observing the procedural safeguards of neither.⁵

The Indian law currently lacks any statutory framework to address this specific procedural gap. On the one occasion on which the issue was called for judicial intervention, the Supreme Court refused to engage with it. The following sections trace the consequences of that judicial disengagement.

3 The Goswami Case: What Was Asked and What Was Refused

3.1 The Petition and Its Prayers

On August 13, 2026, a three-judge bench of the Supreme Court of India, comprising Chief Justice Surya Kant, Justice Joymalya Bagchi, and Justice V. Mohana, heard a public interest petition filed by advocate Narendra Kumar Goswami.⁶ The petitioner prayed before the Court to lay down constitutional safeguards to govern the state’s deployment of “high-risk” artificial intelligence systems across four critical domains: welfare, policing, surveillance, and content moderation. Within the specific sphere of welfare, the petition described a range of statutory entitlements which reveal immediate risk of algorithmic exclusion. These entitlements generally relate to food rations, public health benefits, daily wages, old-age pension benefits, scholarships, and targeted subsidies.⁷ For the purpose of conceptual comprehensibility, this paper takes up only the welfare aspect of the petition.

It is to be noted that the petition was framed as a claim of constitutional rights against the automation of sovereign power, rather than a mere wishlist for technology policy and AI regulation. It sought a judicial declaration that deploying artificial intelligence in public governance without a governing statutory framework, without mandatory human oversight, and without a guaranteed right to an explanation is violative of fundamental rights guaranteed under Articles 14, 19(1)(a), 19(1)(g), and 21 of the Constitution.⁸ In the alternative, it sought a declaration that any State AI system producing civil consequences must satisfy minimum constitutional safeguards. The petition articulated these minimum standards as the strict requirement of legality, transparency, non-arbitrariness, proportionality, human oversight and effective remedy.⁹

3.2 Two Grounds of Refusal

The bench questioned what tangible relief it could grant in matters of such nature. According to contemporaneous reporting, Chief Justice Kant acknowledged that the petition was very comprehensive but indicated two reasons for the Court’s refusal to take up the matter: first, that algorithmic governance is a highly technical issue for which the judiciary is not the expert, and second, that the regulation of such technology falls squarely within the policy domain of the executive and legislature. Accordingly, the Court asked the Centre to consider the representation the petitioner had already submitted, permitted him to supplement it with a copy of the writ petition, and disposed of the matter without expressing any opinion on the merits. Notably, the Court, before concluding, observed that the use of artificial intelligence within the judiciary itself was already regulated by the Court’s internal guidelines.¹⁰

This refusal by the Apex Court rests on two pillars of judicial restraint. The first is institutional propriety, based on the premise that such technological regulation belongs to the executive and legislature. The second is institutional incompetence, based on the premise that algorithmic systems are too technically complex for judicial evaluation. Both are established grounds for declining constitutional intervention. Yet, in the field of fundamental rights, both grounds remain to be tested. The first ground must be tested against the Court’s own extensive record of intervening to protect fundamental rights where the Parliament has left a vacuum. The second ground must be tested by asking whether the Court in fact requires technical expertise to apply existing constitutional safeguards to an automated state procedure. It is precisely these questions that this paper now explores.

4 When the Court Has Filled Vacuums Before

The first ground of institutional propriety can be evaluated against the Court’s own established jurisprudence of intervening to protect fundamental rights where legislative vacuum occurs. This approach is illustrated in Vishaka v. State of Rajasthan, where the Court dealt with the pressing issue of widespread workplace harassment, but there was no legislative regulation. Rather than directing the petitioners to approach the executive, the Court itself framed guidelines under Article 32. It declared that these guidelines would be binding until a law is made by the Parliament. The Court drew the substance of these safeguards from an international document: the Convention on the Elimination of All Forms of Discrimination Against Women (CEDAW), even though it was not enacted as a domestic law.¹¹

4.1 Understanding the Structural Parallel

Here, a structural parallel can be drawn with Goswami. Both cases exhibit an apparent legislative vacuum, ongoing harm, an affected class of citizens, and a prayer seeking the Court’s intervention or procedural rules rather than asking for a permanent statutory code. This parallel is not merely structural. The petition in Goswami expressly prayed that any guidelines framed by the Court operate as law under Articles 141 and 142 until Parliament enacts legislation,¹² and sought the constitution of a High-Powered Expert Committee, including Supreme Court nominees, to frame them.¹³ The petitioner, in other words, requested the Court to do precisely what it had done in Vishaka. However, there is a stark difference in their factual context. Where Vishaka arose from a specific criminal incident with an identifiable victim, Goswami sought anticipatory relief in the welfare domain with no individual complainant. Yet, both face the same institutional problem of an unregulated space affecting fundamental rights. In one case, the Court stepped in with interim safeguards, but in the other, the matter was sent to the executive.

Vishaka is not the only instance where the Court stepped up when the legislature could not. The Supreme Court has repeatedly crafted safeguards when the existing legal framework could no longer adequately counter emerging harms. In M.C. Mehta v. Union of India, the Court upon realising the inadequacy of the common-law rule in Rylands v. Fletcher formulated the doctrine of absolute liability.¹⁴ Similarly, the proportionality test which the Court now applies to rights claims was not made by the Parliament. It was judicially constructed in Modern Dental College and Research Centre v. State of Madhya Pradesh and applied by the Court in K.S. Puttaswamy v. Union of India to address the problems of modern data practices.¹⁵

4.2 Was the Restraint the Right Move?

This does not mean that the Court should intervene every time there is a legislative void. Judicial overreach is frowned upon and courts cannot take over legislative functions. Judicial restraint has its place. However, the reasons given by the Court in Goswami present an inconsistency with its own past actions. Creating a workable framework, albeit on an interim basis, to protect fundamental rights was the exact step the Court took in the case of Vishaka. Technical complexity did not stop the Court there. Nor has the lack of familiarity with technical subject matter prevented it from articulating an extensive privacy framework in Puttaswamy, or from regulating hazardous industry in M.C. Mehta. The real issue here is not whether the Court should take over, but whether its claim of institutional incompetence holds up against its own established record.

5 What Runs While the Courts Wait

The Court in Goswami treated the core issue of algorithmic governance as a policy issue to be considered by the executive in due course. But this treatment assumes that there is time. There is not. Automated systems are already making exclusionary decisions across Indian welfare administration, and the harms are documented rather than anticipated. However, this problem is not India-specific. A global transition has been observed and what the United Nations has termed “digital welfare states” has made the tension between automated government decision-making and rule-of-law principles a subject of international academic attention.¹⁶

5.1 Automation in Indian Welfare Delivery

The operation of these automated systems runs across multiple welfare sectors through standardised patterns. In Telangana, the Samagra Vedika platform has, since 2016, cross-matched citizens across up to thirty government databases to build a comprehensive 360-degree profile and flag eligibility, resulting in wrongful exclusions, including a widow denied benefits because the system attributed a vehicle to her deceased husband.¹⁷ In Haryana, the Parivar Pehchan Patra works as a primary platform for state welfare schemes, where documented failures include residents wrongly recorded as deceased and losing pensions, alongside auto-populated income figures determining eligibility without input from citizens.¹⁸ A third instance arises under the Targeted Public Distribution System (Control) Amendment Order, 2025, which mandates periodic e-KYC for all households covered by the National Food Security Act and provides for deletion of those found ineligible.¹⁹ Where authentication is not completed, names face removal, and authentication fails at the point of sale where fingerprints do not match. This causes serious problems for the elderly and manual labourers. Another important example of automated deletions was seen under the since repealed MGNREGA framework (before its replacement by VB-G RAM G in December 2025) where roughly 27 lakh workers were reported removed from the database during the shift to Aadhaar-based payment and app-based attendance verification.²⁰

5.2 The Inversion of Burden of Proof

What these systems have in common is not that automation in administration has failed, but that it has placed the citizen on the wrong side of the failure. The system-generated output (automated decision) is presumed to be correct. The burden lies entirely on the wronged citizen to first discover the exclusion, then to identify its cause and finally disprove it. All of this without access to the reasoning behind the decision, and against data the citizen has never seen. This is precisely the issue identified by the technological due process literature almost two decades ago. As demonstrated by Danielle Keats Citron, automated administrative systems erode the foundational protections of a hearing. They dilute the requirement of notice and, most importantly, promote the idea among adjudicating officers that machine outputs are correct.²¹ Both tendencies are visible in the Indian examples above. In Telangana, officials were required to consult the algorithm’s output before recording their own decision, but in practice they generally deferred to it.²²

It was these automated systems that the petition in Goswami asked the Union to disclose. But the Court’s response leaves these systems running, without the articulation of any safeguards against which they might be tested. Judicial restraint may have merit in theory, but the practical cost is paid by the most vulnerable citizens.

6 A Court That Did Engage: SyRI

6.1 The SyRI System and the Judgment of the District Court of The Hague

The Netherlands operated Systeem Risico Indicatie (SyRI), a statutory mechanism that linked data held across various government departments to generate automated risk scores identifying individuals considered likely to commit social security fraud.²³ SyRI closely resembles the Indian welfare systems discussed above. Both involve cross-departmental data linkage, risk profiling of welfare recipients, lack of prior notice, and outputs generated by algorithmic logic.

A coalition of civil society organisations challenged the legislation that authorised SyRI in NJCM c.s. v. The State of the Netherlands before the District Court of The Hague.²⁴ The Court decided against the legislation, holding it incompatible with Article 8²⁵ of the European Convention on Human Rights. It was further held that the system failed the proportionality assessment under Article 8 for two reasons: the absence of transparency about the system’s logic, and the inadequacy of safeguards for those profiled.²⁶ The Court also drew on principles from the General Data Protection Regulation as an interpretative aid, and received an amicus curiae submission from the UN Special Rapporteur on extreme poverty and human rights.²⁷

6.2 Ordinary Procedural Tools with Non-Technical Expertise

It is noteworthy that this Hague court possessed no greater technical expertise in automated systems than the Supreme Court of India. It did not go into the details of the system’s underlying code, nor did it require the help of technical experts. What this shows is that reviewing an automated system does not require specialised technical expertise.²⁸ The reach of this judgment should not, however, be overstated. As scholars note, the decision was limited to the circumstances of the case and says relatively little about automated fraud detection in general.²⁹ Its immediate practical effects on administrative practice were limited. Furthermore, the specific European rights framework is different from the existing Indian framework. Yet the point drawn here is narrower. The question before a constitutional court is not how the algorithm mathematically works, but whether its use by the state can be justified. What this Dutch decision demonstrates is the willingness of a court to take up the matter for review rather than treating it as beyond judicial competence. This willingness is in sharp contrast to the judicial disengagement shown by the Supreme Court of India in Goswami.

6.3 Limits of SyRI and the Legislative Comparison

Europe also has a legislative safeguard in the form of Article 22 of the GDPR, which India lacks. That provision governs decisions based solely on automated processing.³⁰ However, its scope has been questioned by scholars. It remains contested whether Article 22 establishes a genuine right to an explanation or merely provides a limited right to be informed about the logic involved. This ambiguity, surrounding even the finest available statutory model, strengthens the case for judicially articulated standards, particularly in the Indian case where there is no equivalent provision available in the Digital Personal Data Protection Act, 2023. In the absence of both legislative and constitutional safeguards, the automation of state systems runs a risk of systemic harms. This is well documented in the case of Australia’s automated Robodebt scheme, where automated income averaging continued unchecked until it was challenged in the Federal Court, with a Royal Commission subsequently finding the scheme to have been unlawful.³¹

7 The Test Already in Hand

The proportionality test formulated in Modern Dental College and Research Centre v. State of Madhya Pradesh and adopted in Puttaswamy was not designed merely to protect data privacy. It is a general constitutional safeguard that the Court now routinely applies to evaluate any state action limiting fundamental rights. The test relies on four sequential questions: whether the measure has a legitimate aim, whether there is a rational nexus between this state measure and that aim, whether a less restrictive alternative exists, and whether the infringement of the right is balanced against the public benefit gained.³² But this proportionality-based analysis is not unique to Indian jurisprudence. It has become a standard technique across modern constitutional systems for weighing the public interest served against the State’s interference with basic rights.³³

7.1 Applying the Test to an Automated Exclusion

When this test is applied to automated welfare exclusions, its utility becomes clear. Where an automated system deletes a citizen from a ration database or cancels a pension, the court can evaluate the state action using the four prongs of proportionality:

Legality: is the automation in such systems authorised by a statute? And if in fact a valid statute exists, does it explicitly authorise the deletion of names or cancellation of welfare entitlements via algorithmic processing?

Legitimate aim and rational nexus: is the algorithmic processing serving a legitimate purpose, such as efficient, cost-effective delivery of welfare benefits without leakage? Is there a rational nexus between the state’s use of such automation and the aim it pursues?

Necessity: was the use of such a measure necessary or was there a less restrictive measure available to achieve the state’s aim?³⁴

Balancing: does the administrative gain in efficiency or cost saving justify the extent of harm caused to citizens?

None of these four prongs requires the court to inquire into the technical workings of code or machine learning. Asking questions through each prong can simplify the approach to such complex problems. The claim of institutional incompetence therefore cannot survive. The Court by declining to apply an established standard showed judicial disengagement over an issue which it is fully equipped to handle.

7.2 The Limits of Human Oversight

At the same time, the Court should avoid treating the ‘human in the loop’ in such systems as an adequate safeguard. In practice, this works as an empty formality. Recent scholarship warns that a human supervisor or an official usually rubber-stamps the same machine output rather than reviewing it on a case-by-case basis.³⁵ While applying the proportionality test, the court must take this into account before accepting the State’s argument that a human remains in the loop. Vishaka supplies the form: binding interim safeguards holding the field until Parliament acts. Puttaswamy supplies the content: a standard to put inside them. The Court held both and reached for neither.

8 Conclusion

The growing reliance on algorithmic systems in state governance has revealed an accountability gap. This gap has dual dimensions, one legislative and the other judicial. The legislative gap lies in the absence of any statutory provision in the Indian law equivalent to Article 22 of the GDPR. This paper has examined the second dimension of this gap: the judicial reluctance as displayed in Goswami. The Court possessed both a compelling precedent for its interim intervention and a doctrine to supply its content, yet it chose to disengage citing grounds of policy consideration. In the absence of a proper legislative framework, the Court’s intervention matters more, not less.

The analysis in this paper proposes something narrower. It does not seek to create new fundamental rights to deal with rising state automation. Nor does it call for judicial innovation of new safeguards to deal with it. What it proposes is the application of an existing constitutional test to the given set of problems arising from state action that is escaping scrutiny altogether. However, there is a crucial precondition to applying this test: basic disclosure of the existence of automation in state systems. A constitutional court cannot assess the proportionality of an automated system when its use and operation remain undisclosed. The disclosure affidavit requested in Goswami was not a remedy in itself, rather a precondition to get the actual remedy.

The approach proposed here follows the logic the Court itself has applied every time it issued interim guidelines pending legislation. The judicial guardrails hold the ground until the legislative vacuum is filled. But currently the cost of waiting for Parliament to act is being borne by people already unjustly affected by the automated system. These people are also least equipped to deal with their exclusions by litigation. For them, the interval between judicial disengagement and legislative action is not an abstraction. It is a matter of daily survival.