Watts, Water, and Waste: The Environmental Price of Artificial Intelligence and the Law’s Failure to Collect It

Written by - Arunima Jha, High Court of Bombay & Arkadyuti Sarkar, High Court of Calcutta

Artificial intelligence is sold and adopted as a clean and weightless technology. Yet, its physical substrate consumes electricity on a national scale, drinks fresh water in cities that cannot spare it, and leaves a widening trail of electronic waste. The carbon footprint of AI systems alone is projected at between 32.6 and 79.7 million tonnes of carbon dioxide in 2025, with a water footprint reaching hundreds of billions of litres. No jurisdiction has enacted a binding, purpose-built framework that holds AI infrastructure accountable for this harm. This article argues that India’s constitutional environmental jurisprudence, its polluter pays and precautionary principles, and the activist mandate of the National Green Tribunal together supply the doctrinal tools to govern AI’s environmental cost. The deficit is not conceptual but political. By comparing the Indian jurisprudence with the European Union, the United States, and China, the article maps the existing jurisprudential gaps and proposes the elements of a dedicated AI environmental liability regime. India has the instruments; what it lacks is the will to use them.

PART I: THE PROBLEM

1. Introduction

1.1 The Myth of the Clean Industry

Artificial intelligence occupies a peculiar place in the public imagination. It is spoken of as intangible, frictionless, and post-industrial, a matter of code and clever mathematics rather than of steel, concrete, and combustion. A user typing a query into a chatbot perceives software, not infrastructure. That perception is the foundational error this article seeks to correct. Behind every user-entered prompt sits a physical economy of hyperscale data centres, racks of power-hungry graphics processing units, industrial cooling systems, diesel backup generators, and a global supply chain that begins in cobalt and lithium mines. The model is software; the cost is overwhelmingly physical.

The scale of that physical reality is no longer speculative. A single AI-focused data centre can consume as much electricity as over 100,000 households, approximately, and draw up to several million litres of water a day for cooling. Constructing such a facility, before it processes a single instruction, embeds enormous quantities of carbon in its concrete, steel, and semiconductors. The gap between how AI is perceived and what it actually demands of the natural world is the problem from which every legal question in this article follows.

1.2 The Legal Silence

Against this expanding footprint stands a near-total legal silence. No jurisdiction in the world has enacted a binding, purpose-built framework imposing environmental liability on AI systems or the data centres that run them. The European Union, through its Artificial Intelligence Act, has gone furthest, yet its environmental provisions amount to little more than a mandate to develop voluntary standards. In the United States, the only dedicated federal proposal never actually advanced beyond committee. While India to date has neither any AI-specific environmental regulation, nor mandatory environmental disclosure tied to computational activity, nor a liability mechanism designed for diffuse, infrastructure-driven harm of this kind. The central question of this article is thus twofold: can existing legal frameworks be stretched to cover AI’s environmental harm, and if they cannot, what must replace them?

1.3 Scope and Argument

This article takes Indian law as its primary frame, with comparative reference to the European Union, the United States, and China. The analysis is doctrinal, drawing on environmental statutes, corporate disclosure law, and constitutional jurisprudence. Its principal claim is that the polluter pays principle, the precautionary principle, the public trust doctrine, and the body of fundamental-rights jurisprudence built around Article 21 of the Constitution collectively demand accountability for the environmental harm caused by artificial intelligence. The claim is normative as much as descriptive: the doctrinal foundation already exists, and the failure to build upon it is a choice, not an inevitability.

1.4 Methodology

This article adopts a doctrinal legal methodology combining comparative jurisprudence and policy analysis. The primary references include: constitutional provisions, statutes, judicial decisions, and regulatory instruments for comparative analysis of the European Union, the United States, and China to identify regulatory approaches suited to the Indian context.

2. Watts, Water, and Waste: The Scale of the Problem

2.1 Energy: The Watt Problem

Energy is the most quantified dimension of AI’s environmental cost. The carbon footprint of AI systems alone is projected to fall between 32.6 and 79.7 million tonnes of carbon dioxide in 2025. In the United States, a detailed study of more than two thousand facilities found that data centres accounted for over four percent of national electricity consumption, that fifty-six percent of that power was derived from fossil fuels, and that the sector generated more than 105 million tonnes of carbon dioxide equivalent, with a carbon intensity roughly forty-eight percent above the national average. Energy is consumed across the full lifecycle of a model, in the compute-intensive training phase and again, cumulatively and indefinitely, in the inference phase each time the model is queried. In India, the harm has an additional and more direct legal dimension: data centres routinely operate large diesel backup generators during grid outages, and generator emissions constitute “air pollution” within the clear meaning of the Air (Prevention and Control of Pollution) Act, 1981 - a pathway to liability under existing law that requires no doctrinal extension whatsoever.

The corporate record exposes the contradiction at the heart of industry climate pledges. Google has reported that its greenhouse gas emissions rose by roughly half since 2019, driven by the energy intensity of AI compute, and it has abandoned its claim to operational carbon neutrality. Microsoft has reported emissions well above its 2020 baseline despite a pledge to become carbon negative, attributing the increase to the construction of AI-optimised data centres. Both firms remain in self-declared breach of their own targets. For India, the danger is magnified: a national grid still heavily relying on thermo-electric sources means that every additional watt drawn by a domestic data centre adds to a higher carbon penalty than the same watt drawn from a grid rich in renewables or nuclear power. Computational growth layered onto a coal-fired grid is a multiplier of harm, not a neutral input.

2.2 Water: The Invisible Consumption

Water is the least visible and least disclosed of AI’s appetites. A mid-sized data centre can consume well over a million litres a day just for cooling, while the indirect water cost of generating its electricity is often larger still. A leading analysis grounded in Phoenix, Arizona, found that roughly a third of data centres nationally sit in areas of high or extremely high water stress, and projected that water use tied to data centre electricity demand in the Phoenix region could rise by as much as four hundred percent, with regional water stress increasing by up to thirty-two percent if all planned facilities come online. The opacity is structural: leading developers publish little to no facility-level water data, leaving regulators and citizens unable to measure the harm they are asked to absorb without clarity.

India’s exposure is acute. The country supports eighteen percent of the world’s population relying on four percent of its fresh water, and its data centre boom is concentrated in Mumbai, Navi Mumbai, Hyderabad, Chennai, and Pune, with all these cities already competing for scarce supply. Industry projections see data centre water consumption in India rising from roughly 150 billion litres in 2025 toward 358 billion litres by 2030. A water-cooling facility sited in a water-stressed basin converts a global technological ambition into a local subsistence crisis, drawing from the same aquifers and reservoirs on which households and agriculture depend.

2.3 Waste: Hardware and Supply Chain Harm

The third dimension is waste, and it is the one least amenable to a snapshot statistic because it accrues across a lifecycle involving GPU turnover rates, impacts of semiconductor manufacturing, extraction of rare earth metals, and other indirect contributors to the AI-environmental relation. According to a 2024 research article titled “E-Waste Challenges of Generative Artificial Intelligence,” suggested that AI-generated e-waste can cumulatively hit 1.2 to 5 million tonnes between 2020 and 2030, based on adoption patterns. The research further estimated AI becoming a significant contributor to global e-waste surge unless the present hardware obsolescence is reduced.

The relentless cycle of GPU obsolescence, as each hardware generation is superseded by a faster successor, accelerates the generation of electronic waste. Upstream, the extraction of cobalt, lithium, and coltan inflicts ecological damage at the source, often in jurisdictions with weak environmental enforcement. The appropriate unit of legal analysis is therefore not the operating data centre in isolation but the entire chain, from mine to manufacture to deployment to disposal.

India occupies a doubly exposed position in this chain. It is simultaneously a fast-growing destination for data centre construction, with hubs emerging around New Town Kolkata and the broader Bengal Silicon Valley corridor, and one of the world’s significant processors of electronic waste, much of it handled in an informal sector with serious consequences for workers and groundwater alike. The country thus bears both the front-end burden of hosting energy- and water-intensive infrastructure and the back-end burden of absorbing the hardware it discards. Any framework that addresses only operational emissions, ignoring the embodied and end-of-life harm, will capture a fraction of the true cost.

While India has E-Waste (Management) Rules, 2022 in place, it was specifically designed to deal with consumer electronics, electrical equipment, and conventional IT hardware. Therefore, it proves insufficient to address AI-related issues like requiring disclosure of AI-related hardware turnover, tracking AI accelerator disposal, imposing environmental reporting obligations on AI developers, and accounting for lifecycle impacts of large-scale AI computation. Thus, AI-generated e-waste often lands in a regulatory blind spot which actually necessitates targeted intervention.

PART II: THE LAW AND ITS FAILURES

3. Indian Environmental Law: Structural Inadequacy

3.1 The Constitutional Foundation

India’s constitutional environmental jurisprudence is among the most expansive in the world, and it furnishes the strongest available normative basis for AI environmental accountability. In Subhash Kumar v. State of Bihar, the Supreme Court held that the right to life under Article 21 includes the right to enjoy pollution-free water and air. That guarantee is reinforced by the Directive Principle in Article 48A, which casts a positive legislative duty on the State to protect and improve the environment, and by the fundamental duty in Article 51A(g) binding every citizen to protect the natural environment. Together, these provisions create both a right to be protected from environmental harm and a corresponding obligation upon the State to regulate the sources of that harm.

The parallel with privacy is instructive. In Justice K.S. Puttaswamy v. Union of India, a nine-judge constitutional bench recognised that fundamental rights can be violated in ways that are invisible, diffuse, and systemic rather than discrete and immediate. AI’s environmental harm shares precisely that character. No single act of pollution is visible to the citizen whose air and water are degraded by the cumulative draw of distant data centres. If the Constitution can reach the invisible architecture of data collection, it can reach the equally invisible architecture of data computation.

3.2 The Environment (Protection) Act, 1986

The principal statutory instrument concerning India’s environmental legislation is the Environment (Protection) Act, 1986. Section 3 of the statute confers on the Central Government broad power to adopt all measures necessary to protect and improve the environment, while Section 7 prohibits the discharge of environmental pollutants in excess of prescribed standards. The difficulty is definitional. The Act considers a “pollutant” as a solid, liquid, or gaseous substance present in a concentration injurious to the environment, and of “environmental pollution” as the presence of such a substance. It is far from clear if the energy overconsumption of a data centre, or the carbon attributable to the grid powering it, fits this frame. The harm of AI is often not a discharge from an identifiable pipe but the aggregate consequence of consumption. The Act was built for physical pollutants emanating from identifiable point sources, and AI infrastructure fits that template only awkwardly.

3.3 The Polluter Pays Principle and Its Limits

3.3A. Sustainable Development as a Governing Principle

The polluter pays and absolute liability principles are the doctrinal heart of Indian environmental law. In M.C. Mehta v. Union of India, the Oleum Gas Leak case, the Supreme Court fashioned a rule of absolute liability under which an enterprise engaged in a hazardous or inherently dangerous activity is liable to compensate for harm without the exceptions that qualify the English rule in Rylands v. Fletcher. In Indian Council for Enviro-Legal Action v. Union of India, the Court affirmed that the financial burden of remedying environmental degradation falls on the polluter who caused it. In Vellore Citizens’ Welfare Forum v. Union of India, it held that the precautionary principle and the polluter pays principle are part of the law of the land. Furthermore, the Vellore Citizens' case has its significance for another reason too. Alongside the polluter pays and precautionary principles, the Supreme Court also considered sustainable development to be an inherent part of India’s environmental jurisprudence. The doctrine discards the myth that economic growth and environmental protection cannot go alongside, and instead suggests that developmental activities need to internalise their environmental costs.

Thus, this principle becomes particularly relevant for artificial intelligence. AI systems promise substantial economic and social benefits, yet their supporting infrastructure requires humungous quantities of electrical power, water, and computing hardware. The challenge, therefore, is not whether to develop AI but whether the environmental costs of doing so should be externalised onto communities and ecosystems. Viewed through the lens of sustainable development, environmental disclosure, impact assessment, and resource-efficiency obligations are not obstacles to innovation but the conditions under which innovation remains ecologically legitimate.

Now, applying absolute liability principles to AI infrastructure comes with both opportunities and challenges. Contrary to the nature of hazardous industries considered in M.C. Mehta v. Union of India, hyperscale data centres apparently do not pose risks of sudden industrial accidents, toxic releases, or catastrophic physical harm. However, their environmental impact arises from continued and extensive consumption of electricity, water, and hardware resources, generating cumulative externalities through carbon emissions, resource depletion, and electronic waste.

Still, the rationale underlying the doctrine of absolute liability stays relevant. Large-scale AI infrastructure operators possess substantial control over environmentally consequential activities, while reaping significant economic fruits from them, and are best positioned as monitors and mitigators of the resulting environmental impacts. As environmental harm cascades through patterns of intensive resource consumption rather than discrete pollution events, there is a plausible case to re-examine whether the principles developed in the Oleum Gas case should significantly evolve to deal with the advanced threats of emerging environmentally hazardous activity. The question is not whether AI infrastructure is hazardous in the traditional industrial sense, but rather whether enterprises generating foreseeable and substantial environmental externalities need to be required to bear the costs of those externalities on a no-fault basis.

3.4. Public Trust Doctrine and Its Relevance in the Emerging AI Era

While the Polluter Pays Principle can only be applied when pollution has occurred, resulting in quantifiable harm and the polluter is known. However, the principle falls short in dealing with AI-related environmental conflicts, as the harm is cumulative, water consumption is ongoing, and resource depletion usually precedes the occurrence of any obvious environmental damage.

The Public Trust Doctrine, having its source in Roman law and developed through American jurisprudence, holds that certain natural and environmental resources are preserved for public utilities. Such resources may be held in trust of the State and include waterbodies, land, forest and wildlife, etc.

Coming to Indian environmental jurisprudence, this doctrine was first applied in the landmark case of M.C. Mehta v. Kamal Nath (1997), wherein the Supreme Court held that natural resources like rivers and forests, being public property, make the State their trustee and liable for protecting them for public use. The same principle was reiterated in another landmark decision in M.I. Builders Pvt. Ltd. v. Radhey Shyam Sahu (1999), reinforcing the judicial commitment towards protecting environmental resources and preserving public utilities.

Having already explained previously and throughout this article about the extensive level of water consumption by AI data centres, the environmental challenge resulting from artificial intelligence is not only about pollution but resource allocation as well. In a country like India, where droughts and water shortage issues are quite common in many regions, the allocation of substantial quantities of freshwater for operating privately owned AI infrastructure raises the constitutional question of whether the State is performing its obligations in the AI era as the trustee as required by the doctrine?

3.5 The National Green Tribunal: Bridge or Destination?

The most immediately available vehicle for environmental accountability is the National Green Tribunal. Section 14 of the National Green Tribunal Act, 2010 empowers the Green Tribunal with jurisdiction over all civil cases involving a substantial question relating to the environment, a deliberately broad jurisdictional hook. The Tribunal has developed an activist jurisprudence by asserting suo motu powers, granting expansive standing, and awarding compensation under a polluter pays rationale. A public interest petition before the Tribunal, framed around the environmental compliance of data centres and grounded in the constitutional right to a clean environment, is the first practical step available today, without waiting for Parliament to bring forth special legislation or any amendment to the existing piece.

Yet the Tribunal is simply a bridge, not a destination. It can fill gaps case by case and compel disclosure or operational conditions in a particular dispute, but cannot legislate a comprehensive liability regime, set standardised metrics across the sector, or build the administrative architecture that systematic regulation requires. Its proper role is to hold the line and build a factual and legal record while the legislature acts, not to substitute indefinitely for legislation that has not yet been written.

3.6 Corporate Disclosure Frameworks: Why They Fall Short

Corporate disclosure law offers a partial entry point. The Securities and Exchange Board of India’s Business Responsibility and Sustainability Reporting framework has, since the financial year 2022–23, required the thousand largest listed companies by market capitalisation to disclose data on energy, water, and emissions. The framework is detailed, but it contains no metric tailored to computational activity, no measure of energy per training run or water per query, and no mechanism to capture the harm of unlisted operators or foreign developers. The corporate social responsibility regime under Section 135 and Schedule VII of the Companies Act, 2013 permits environmental expenditure as a qualifying activity, but it is voluntary spending, not liability or accountability. The Digital Personal Data Protection Act, 2023, which governs the very data that AI systems process, embedded no environmental obligation at all, a conspicuous missed opportunity to align data regulation with environmental accountability.

4. The Global Landscape: Nobody Has Solved This

4.1 The EU AI Act: Promised Much, Delivered Little

The European Union’s Artificial Intelligence Act is the most ambitious attempt anywhere to regulate AI, and its environmental provisions are its most disappointing feature. Article 40 directs European standardisation bodies to develop deliverables on reducing the energy and resource consumption of high-risk systems and on the energy-efficient development of general-purpose models, and the technical documentation of such models must include their computational and estimated energy consumption. What the European Parliament originally pressed for was more demanding: mandatory lifecycle assessment and binding sustainability requirements. What survived the trilogue negotiations was a regime of voluntary codes of conduct, deferred standard-setting, and self-assessment whose results need not be published.

The Corporate Sustainability Due Diligence Directive complements the Act by imposing diligence obligations along corporate value chains, and may capture some environmental harm indirectly. Even read together, however, the European framework relies on the goodwill of policymakers to make voluntary standards exacting and on the goodwill of firms to comply with standards that are largely unenforced. It is the high-water mark of global ambition, and it remains structurally inadequate on the environment.

4.2 United States and China: Contrasting Approaches

The United States and China illustrate opposite ends of the regulatory spectrum. In the United States, the Artificial Intelligence Environmental Impacts Act of 2024 would have directed the Environmental Protection Agency to study AI’s environmental impacts and the National Institute of Standards and Technology to develop measurement standards and a voluntary reporting system, but the bill never advanced beyond committee. The most instructive precedent, however, lies not in the United States but in Singapore, which imposed a complete moratorium on new data centre construction from 2019 to 2022 specifically because of energy and land constraints, and restarted approvals only under a regulated, sustainability-conditioned framework. Singapore’s experience demonstrates that binding environmental conditions on data centre development are not merely theoretically achievable but have been successfully implemented by a jurisdiction with comparable levels of digital infrastructure ambition.

China, by contrast, operates the most prescriptive existing framework. Under its public-procurement and efficiency standards, the power usage effectiveness of a data centre must fall below 1.4 from June 2023 and below 1.3 from 2025, and the ratio of annual water consumption to IT power consumption must remain under 2.5 litres per kilowatt-hour. Whatever one makes of the political system that produced them, these are binding numerical standards for AI infrastructure. The lesson is unambiguous: enforceable standards are technically and legally achievable. The variable that separates China from the United States is not about capacity but different political will, and China’s tiered efficiency classification offers a concrete template that a country like India could adapt to its own constitutional and administrative setting.

4.3 The Comparative Lesson for India

Three conclusions follow from the comparison. First, no jurisdiction has comprehensively solved the problem; India would not be playing catch-up so much as entering an open field. Second, India’s constitutional environmental jurisprudence is more expansive, and its National Green Tribunal more activist, than the comparable institutions of the European Union, the United States, or China, which means India begins with stronger doctrinal foundations than any of them. Third, India’s domestic data centre boom makes action urgent rather than academic: the facilities are being built now, in water-stressed cities, on a coal-heavy grid. India is, on this analysis, uniquely positioned to lead, possessing the constitutional tradition, the judicial machinery, and the pressing domestic need at the same moment.

PART III: BUILDING THE ANSWER

5. The Case for a Purpose-Built Framework

5.1 Why Existing Law Cannot Be Stretched Far Enough

The argument for a purpose-built framework rests on four gaps that judicial creativity alone cannot close. The first is definitional: the concepts of pollutant, point source, and identifiable harm, on which existing environmental law depends, do not map cleanly onto a harm that is consumptive, distributed, and embodied. The second is jurisdictional: the largest AI developers operate across borders, and national law has limited reach over conduct and supply chains located in other jurisdictions. The third is evidential: the transparency of corporate environmental data, particularly on water, makes liability difficult to establish even where doctrine would otherwise apply. The fourth is institutional: no Indian regulator currently holds a specific mandate over AI’s environmental footprint, leaving authority fragmented and partial. Each gap can be narrowed by litigation, but none can be fully bridged without legislation. Incremental application of existing law is necessary, and it is insufficient.

5.2 The Constitutional Imperative

The action case is ultimately a constitutional one. AI’s environmental harm degrades air, water, and climate, and so engages the right to life under Article 21 as construed in Subhash Kumar. The State’s duty under Article 48A to protect and improve the environment is, in this light, a positive obligation to regulate that harm rather than a discretionary aspiration. India’s Nationally Determined Contributions under the Paris Agreement commit to reducing the emissions intensity of GDP by 45% by 2030 and achieving 50% of cumulative electricity from non-fossil sources by the same year; expanding AI data centre capacity on a coal-heavy grid places the State in direct tension with its own internationally declared obligations. The precautionary principle, embedded in Indian law by Vellore Citizens’ Welfare Forum, instructs that scientific uncertainty about the precise magnitude of harm is not a licence for regulatory inaction. A public interest petition can frame the question directly: whether the State’s failure to regulate the environmental harm of artificial intelligence violates Articles 14, 21, and 48A of the Constitution. So framed, the issue is justiciable today, in the Supreme Court or before the National Green Tribunal.

5.3 The Polluter Pays Principle: A New Formulation for AI

The polluter pays principle must be revamped to fit the structure of AI-generated harm. Its traditional formulation presumes an identifiable polluter, an identifiable harm, and a direct causal chain. AI’s harm has none of these in pure form; it is produced by multiple actors through diffuse pathways across global supply chains. The answer is a regime of proportionate liability, in which an operator’s responsibility is calibrated to objective and measurable factors: the computational intensity of the system it runs, its energy consumption above a renewable-energy threshold, its water consumption in water-stressed locations, and the hardware lifecycle and electronic waste it generates.

This is best achieved by extending the absolute liability doctrine of M.C. Mehta to large-scale AI infrastructure, treating a hyperscale data centre as an inherently resource-intensive enterprise whose operator bears liability for the environmental harm its operation necessarily entails. Where harm is genuinely diffuse and cannot be attributed to a single operator, joint and several liability across the AI supply chain provides the mechanism to ensure that the cost is borne by those who profit from the activity rather than by the public that suffers its consequences. The doctrine need not be invented; it needs to be adapted.

6. Elements of a Proposed Framework

6.1 Mandatory AI Environmental Impact Assessment

The cornerstone of any framework should be a mandatory environmental impact assessment for AI, modelled on the existing Environmental Impact Assessment Notification, 2006. Application should be threshold-based, capturing large language models, generative systems, and data centres above a specified computational capacity, so that the burden falls on facilities whose footprint is material. A pre-deployment assessment would require projection of energy consumption, water usage, and hardware lifecycle. Periodic post-deployment reporting would compare actual against projected consumption, subject to independent verification, and the results would be publicly disclosed in a form accessible to citizens, civil society, and the National Green Tribunal. Disclosure is the precondition for every other form of accountability; without it, liability cannot be established, and enforcement cannot begin.

6.2 Regulatory Architecture

The present regulatory landscape is fragmented across the Ministry of Electronics and Information Technology, the Ministry of Environment, Forest and Climate Change, the securities regulator, and the central bank, each holding partial jurisdiction and none holding a complete mandate. The proposal is a dedicated AI Environmental Division within the Ministry of Environment, Forest and Climate Change, equipped with inspection authority, the power to mandate disclosure, the power to impose penalties, and the power to attach environmental conditions to data centre licences. Such a division would coordinate with the State Pollution Control Boards, which already exercise operational and emissions oversight, and would provide a single locus of responsibility in place of the current diffusion of authority. Clear reporting lines and a defined appeal structure would give the regime both coherence and due process.

6.3 Liability and Enforcement

The liability regime should impose strict liability on large-scale AI infrastructure, dispensing with any requirement to prove negligence, consistent with the absolute liability tradition. Extended producer responsibility, already a feature of India’s electronic waste regime, should be applied so that hardware manufacturers remain liable for waste across the lifecycle of their products. Carbon pricing should attach to data centre emissions above the national grid average carbon intensity, a mechanism already partially enabled by the Carbon Credit Trading Scheme established under the Energy Conservation (Amendment) Act, 2022, which provides an existing legislative vehicle for AI-specific carbon obligations without requiring entirely new primary legislation and water tariffs should reflect the true scarcity value of water in stressed locations rather than a subsidised rate that conceals the real cost. Penalties should be structured in tiers, combining financial penalties with operational restrictions and, for repeat violations, licence cancellation, and they should be calibrated to computational scale rather than imposed at a flat rate, so that the largest emitters face the largest consequences.

6.4 Transparency and Disclosure Obligations

Transparency obligations should run parallel with liability. Mandatory model cards should include standardised environmental impact data, and large operators should file annual AI environmental impact reports with the regulator on the public record. Metrics should be standardised across the sector, covering energy per query, water per training run, and hardware lifecycle emissions, so that disclosures are comparable rather than idiosyncratic. The application of the Right to Information Act, 2005 to public sector AI deployments should be clarified by legislation, ensuring that citizens can interrogate the environmental cost of AI procured in their name. Employees who disclose environmental violations by AI companies should receive whistleblower protection, recognising that those inside the firm are often the only people with access to the data that opacity otherwise conceals.

6.5 The Interim Role of the NGT

Pending legislation, the National Green Tribunal should act as the interim guardian. Suo motu action on data centre environmental compliance is the most immediate step available, and a carefully framed public interest petition, attentive to standing, forum selection, and the constitutional framing of its prayer, can secure interim directions requiring disclosure and imposing operational conditions while Parliament deliberates. Standing under Section 19 of the National Green Tribunal Act is available to persons likely to sustain injury and to representative bodies with a record of environmental work, making environmental NGOs, affected residents near data centre clusters, and public-spirited legal practitioners all potential petitioners without the narrower threshold that applies before civil courts. The jurisprudence the Tribunal develops in the process would supply the factual and legal record on which durable, comprehensive legislation can later be built. The Tribunal cannot be the final answer, but it can ensure that the years before legislation are not years of complete impunity.

PART IV: CONCLUSION

7. Conclusion and Recommendations

7.1 Summary of Findings

The findings of this article can be stated compactly. AI’s environmental footprint, across energy, water, and waste, is real, rapidly growing, and at present almost entirely unaccounted for in law. Indian environmental law, through the Environment (Protection) Act, the National Green Tribunal Act, the constitutional provisions, and the polluter pays and precautionary principles, supplies a genuine doctrinal foundation but not a complete answer. Global frameworks, including the most ambitious of them in the European Union, have failed to deliver binding environmental accountability for AI. A purpose-built AI environmental liability framework is ultimately necessary; incremental fixes cannot bridge these gaps alone.

7.2 The Central Argument Restated

The law has always, eventually, caught up with technologies that cause harm. It caught up with the lead in petrol, asbestos in buildings, and with chlorofluorocarbons in the atmosphere, in each case after a period of profitable denial during which the harm was known and unaddressed. Artificial intelligence is today’s legally unregulated harm. The only question that matters is whether the law catches up before the damage, to water tables, to a warming climate, to the communities living beside these facilities, becomes irreversible. The history of environmental regulation is a history of belated correction; the opportunity here is to correct early, while correction is still cheap.

7.3 Recommendations

The following measures, ordered by instrument and timeline, would together constitute a credible response.

Recommendation

Instrument

Timeline

Mandatory AI EIA

MoEF Notification under EPA 1986

Immediate

AI-specific BRSR disclosure

SEBI LODR Amendment

Short term

Dedicated AI Environmental Division

Administrative order, MoEF

Short term

Strict liability for AI infrastructure

NGT PIL / Supreme Court

Immediate via litigation

Carbon pricing for data centre emissions

Finance Act Amendment

Medium term

Water tariff reform for AI cooling

State water policy

Short term

Extended producer responsibility

E-Waste Rules Amendment

Short term

Mandatory model cards, environmental data

MeitY Regulation

Short term

RTI clarification for public sector AI

RTI Act Clarification

Immediate

AI Environmental Liability Act

Primary legislation

Long term

7.4 Closing

With artificial intelligence becoming an inherent part of socio-economic lifestyle, its environmental impacts need to be considered through the lens of legal accountability. The challenge for India is not the absence of legal principles, but in adapting the existing environmental jurisprudence to a new form of resource-intensive digital infrastructure.