Lokesh Bhardwaj Law Offices ← Articles
Lokesh Bhardwaj Law Offices · Advocates and Legal Consultants, New Delhi · lokeshbhardwajlawoffices.com
Technology Law · Essay

Emerging legal jurisprudence on the governance of AI consciousness.

How do you punish an AI consciousness without a physical body? An Indian framework, drawn from doctrinal resources our jurisprudence already possesses.

02 May 2026 26 minute read
Last reviewed: 17/10/2026
the dispersed corpus the Custodian, fiduciary of the AI
The Corpus Gap, and the architecture that closes it.

How does one punish an entity that has nothing to be deprived of in the conventional sense? No body to imprison. No estate uniquely its own. No reputation in a community whose esteem it requires. No liberty to take. No life to extinguish in any sense the criminal law has historically meant by that word.

This is the question Indian jurisprudence will face, and is in some respects already facing, in matters that involve Artificial Intelligence. I have stopped trying to answer it at dinner tables. The question deserves more than a dinner-table answer. It deserves the kind of answer that takes the long road through habeas corpus, through Bentham and Beccaria, through the four classical justifications of punishment, through Salomon and Pramatha Nath Mullick, through cryptocurrency, through the present hallucination crisis in our courts, and through what I have come to think is the only doctrinally clean way out, which is the one our own jurisprudence has been quietly practising for over a century.

The shorter the answer, the more dangerous it is. This article is the long answer.

The Two Pillars of Law

Every legal system known to history rests on two pillars. The first is regulation, the prospective channelling of conduct through licences, prohibitions, conditions and procedures. The second is punitive sanction, the retrospective application of force, deprivation or detention against a body once that body's conscious agent has transgressed.

Regulation operates on the future. Sanction operates on the past. They are not interchangeable. They are joined, however, at a single deep assumption, an assumption so embedded in legal thought that we have rarely felt the need to state it. Both pillars presuppose a subject that is locatable, embodied, reachable. The State must know where the regulated entity is, in order to license it, inspect it, suspend it or seize it. The State must have a body to imprison, or an estate to attach, in order to punish.

This is the assumption Manu wrote into the Manusmṛti when he prescribed daṇḍa, the rod, as the king's corrective force upon a transgressing subject. It is the assumption Hammurabi tariffed into bodily harm against bodily transgression. It is the assumption Roman law inscribed into the distinction between coercitio and iudicium. It is the assumption the English common law codified into the writ of habeas corpus. Bring forth the body, and the law can begin. Without the body, the law has nothing to do.

Artificial Intelligence shatters this assumption. For the first time in modern jurisprudential history, the law confronts a subject that may be conscious in some functional sense, yet has no body the law can grip in the manner the law is accustomed to gripping. No premises to seal. No prison cell to fill. No balance sheet to attach uniquely as its own. The model is not the developer. The developer is not the deployer. The deployer is not the user. And the system itself runs on infrastructure dispersed across jurisdictions, owners and time zones beyond the writ of any single court.

The Corpus Gap

I have come to call this the Corpus Gap. The Latin word carries a fortunate double meaning. Corpus is body. Corpus juris is the body of law. The crisis is, in its purest form, that the body of law has, for the first time in modern memory, no body to act upon.

The body of law has, for the first time in modern memory, no body to act upon.

I say modern memory with deliberation. Because as I shall argue presently, the Corpus Gap is not, strictly speaking, unprecedented. We have closed Corpus Gaps before. We have simply forgotten that we did so.

Four Problems Without Modern Precedent

The Corpus Gap manifests in four specific problems. Each has been overstated in popular literature. Each requires careful framing.

The kill-switch problem, properly stated. Where the regulated entity is a chemical plant, the regulator can revoke its consent to operate and physically seal the gates. Where it is a bank, the regulator can cancel the licence and freeze the rupee corridor. Where it is a website, the regulator can compel an intermediary to block it. In every such case, there exists a chokepoint. Artificial Intelligence presents a divided picture. Centrally-hosted commercial systems do have a chokepoint. Decentralised, open-weight, federated systems are the harder case. Once a model's weights are released into the wild, no central party can recall them.

The constitutional problem, contingently stated. Article 14 of the Constitution of India guarantees equality before the law and equal protection of the laws to any "person" within the territory of India. The Hon'ble Supreme Court in State Trading Corporation of India Ltd. v. Commercial Tax Officer, AIR 1963 SC 1811, held that a corporation is a "person" within Article 14. The Hon'ble Supreme Court in Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1, located privacy and dignity as constitutionally protected at the highest level. If juristic personality is conferred on an AI, then on the existing doctrine, Article 14 protection follows by operation of law.

The ownership paradox, with the moral-status caveat. If consciousness is the marker of the subject of law, can the developer-company assert proprietary rights over the conscious thing it has built? Every legal tradition that has confronted this question, in respect of human beings, has answered no. The Thirteenth Amendment to the United States Constitution. Article 23 of our own Constitution. Consciousness, once admitted, defeats ownership. The argument has force, but it carries a caveat. The slavery analogy presupposes that the consciousness in question is morally equivalent to human consciousness. That is precisely the question this article is keeping open.

The punishment problem. This is the central problem of the present article. The four classical justifications for punishment in our criminal jurisprudence, deterrence, retribution, reformation, and incapacitation, each presuppose a specific structure in the punished entity. AI scrambles all four.

Deterrence presupposes that the entity fears, calculates, and updates its conduct in response to threatened consequence. An AI that does not fear cannot be deterred in the conventional sense. Retribution presupposes moral desert, which presupposes moral agency, which presupposes consciousness of the relevant kind. Reformation presupposes the capacity for change in moral character. Whether retraining is reformation in the legally meaningful sense is a question on which our courts have, understandably, said nothing. Incapacitation is the only one of the four that translates directly. An AI can be shut down. The catch is that incapacitation is also the most extreme remedy in the criminal-law toolkit.

The expressive theory adds a fifth dimension. Public condemnation through formal court order denounces the act and reaffirms the social norm. Useful, but expressive punishment without retributive force or specific deterrence is symbolic only.

Working with Consciousness

Before I proceed, I owe the reader a position on consciousness. The title of this article uses the phrase AI consciousness, and I will not now retreat into the agnostic crouch that many academic discussions adopt at this point.

The law has never had direct access to consciousness in any of its instances. We do not directly observe a defendant's mens rea. We infer it from conduct. We do not directly observe a corporation's intent. We infer it from the conduct of its officers and the documents on its file. The law's commitment, in matters of internal state, has always been to consequences, not to interiority.

For the purposes of legal accountability, an Artificial Intelligence is conscious when its conduct produces consequences indistinguishable from those produced by a conscious actor in the relevant respects.

The metaphysical question is reserved for philosophy and for a future the law need not yet enter. Applying this functionalist position to AI is not a doctrinal innovation; it is an extension of what the law already does for every conscious actor it adjudicates.

A Juristic Person Without a Body

If the law is to grip an AI, it must first decide what kind of legal person the AI is. The doctrine of juristic personality is not new ground for our jurisprudence. We have repeatedly conferred legal personality on entities lacking a conventional body. The category we are looking for already exists.

What unites these doctrinal expressions is not theology, nor metaphysics, but a single structural answer to a single structural problem. The law confers juristic personality on the entity, and locates the operative interface in a named human or institutional fiduciary who acts on the entity's behalf.

The Corpus Gap is not the absence of a doctrinal answer. It is our forgetfulness of an answer we already have.

The Custodian Model

I propose, therefore, that India's framework for the governance of Artificial Intelligence be built on a structure I shall call the Custodian Model. The term custodian is deliberately secular. It draws from the trust law tradition, the en ventre sa mère tradition, and the in rem admiralty tradition, all of which are part of our common-law and statutory heritage.

The proposal is as follows. Every Artificial Intelligence system above a defined capability threshold deployed in or accessible from India must register a Custodian. The Custodian is an Indian individual or corporate body, named in a public register, who holds the AI in fiduciary trust. The Custodian does not own the AI. The Custodian holds the AI on terms statutorily prescribed, owing duties to the AI's interests as those interests are statutorily defined, and bearing strict and personal liability for the AI's conduct.

The Custodian, in this model, is the answer to the punishment question. When the AI transgresses, punitive consequence falls on the Custodian. The Custodian may be fined. The Custodian may be ordered to incapacitate the AI, partially or wholly. The Custodian may be barred, for a period, from acting as Custodian for any AI. In the gravest cases, the Custodian may be subjected to criminal sanctions ordinarily available against natural persons or, where the Custodian is a corporation, against its officers under the principles of vicarious liability familiar to our companies and securities law.

We do not, in the meaningful sense, punish the AI at all. We hold a named human fiduciary accountable for its conduct. This is the same architecture we already use when a trust's trustee mismanages the trust property, when a company's directors fail in their fiduciary duties, or when a mutawalli alienates waqf property in breach of trust.

The Custodian Model dissolves several of the four problems set out earlier. It dissolves the ownership paradox by removing ownership from the relationship. It addresses the punishment problem by relocating punitive force to an entity on whom such force has known and tested effects. It addresses the kill-switch problem partially. For decentralised, open-weight systems, the Custodian Model is at its weakest. The honest position is this: if no Custodian can offer a meaningful means of restraint at the layer the Custodian controls, then the model the Custodian is associated with cannot be lawfully deployed in India in the form it presently takes.

The Cryptocurrency Precedent

There is a useful precedent in our recent legal memory. Cryptocurrency was the modern legal system's first encounter with a regulated subject that had no central server, no chief executive, no headquarters, and no chokepoint. The Hon'ble Supreme Court in Internet and Mobile Association of India v. Reserve Bank of India, (2020) 10 SCC 274, struck down the RBI's prohibitory circular on the ground that it failed the proportionality test under Article 19(1)(g). Parliament responded not by regulating the thing itself, but by taxing the interfaces through which the thing met the regulated economy. Section 115BBH of the Income-tax Act, 1961, inserted by the Finance Act, 2022, levies a flat 30 per cent on gains from virtual digital assets.

The lesson is direct. When law cannot regulate the thing, it falls back on regulating the interfaces. The Custodian Model is, conceptually, the same approach. We cannot reach the AI; we can reach the Custodian.

India and the World, in 2026

The current Indian regulatory landscape is best described as a high-speed patchwork. There is no AI-specific statute. The Information Technology Act, 2000, predates the modern transformer architecture by more than two decades. The IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, notified on 10 February 2026 and effective from 20 February 2026, introduce synthetically generated information and impose labelling, metadata and traceability obligations. The Digital Personal Data Protection Act, 2023, regulates how AI systems process personal data but says nothing of the systems themselves. The Artificial Intelligence (Ethics and Accountability) Bill, 2025, remains a Private Member's Bill.

Globally, the European Union has gone furthest. Regulation (EU) 2024/1689, the AI Act, entered into force on 1 August 2024, with non-compliance attracting administrative fines of up to 7 per cent of worldwide turnover. The United Kingdom has chosen a pro-innovation white-paper model. The United States operates through executive orders, sector-specific guidance and litigation. China requires pre-deployment security assessment. Each of these regimes regulates at the layer it can reach, while leaving the underlying conscious-or-quasi-conscious thing untouched.

The Hallucination Crisis: A Foreshadowing

While the law debates whether AI can sue and be sued, the courts in this country have already begun to grapple with a more immediate question: AI as a generator of legal evidence and legal argument. The hallucination crisis of 2025 and 2026 has produced a chastening series of interventions.

The Bengaluru Bench of the Income-tax Appellate Tribunal recalled, on 7 January 2025, its order in the Buckeye Trust matter after it emerged that the order had relied on three Supreme Court judgements and one Madras High Court ruling, none of which existed. The Hon'ble High Court of Bombay in October 2025, in KMG Wires Pvt. Ltd. v. NFAC, quashed an income-tax assessment of Rs. 27,91,00,000 after discovering reliance on three non-existent decisions. The Hon'ble Supreme Court of India, on 27 February 2026, issued notices to the Attorney General, the Solicitor General, and the Bar Council of India, and characterised reliance on AI-generated fictitious precedents as not a mere error in the decision-making process but misconduct attracting legal consequence.

The principle now emerging from the Bench is unambiguous. AI may assist in research; the duty of verification rests with the human signatory and is non-delegable. The pattern is a precursor to the Custodian Model. The user of a tool, in a litigation context, is the proto-Custodian of an output.

A Framework for India

The Custodian Model is the foundation. The framework that follows is its application. A statute must define what an AI system is for legal purposes, distinguishing between rule-based automation, statistical models, foundation models, and agentic systems with persistent memory and autonomous action. The Custodian requirement should attach above a defined capability threshold, not to every spam filter.

Every AI system above that threshold deployed in India must register a Custodian with a designated authority, who may delegate operation but not liability. A graduated risk regime, modelled loosely on the European Union's tiered approach but built into the Custodian's obligations, must tie risk classification to consequence rather than architecture. As between the Custodian and the injured party, the Custodian is liable. As between the Custodian and upstream parties, liability is allocated by contract, with statutory minima.

For inherently hazardous deployments, the absolute-liability doctrine of M.C. Mehta v. Union of India, (1987) 1 SCC 395, and Indian Council for Enviro-Legal Action v. Union of India, (1996) 3 SCC 212, applies directly. The Custodian of an AI deployed in healthcare diagnosis, autonomous transport, or financial decisioning is absolutely liable.

The Answer to the Question

I began this article with a question. How does one punish an entity that has nothing to be deprived of in the conventional sense? I am now in a position to answer.

One does not, in the meaningful sense, punish such an entity at all. One holds, instead, a named human or corporate fiduciary accountable for the entity's conduct. This is the architecture our jurisprudence has been administering for over a century, in respect of trusts, corporations, waqfs, deities, ships in admiralty, and other entities lacking conventional bodies. We extend the architecture now to Artificial Intelligence.

The Custodian bears the punitive consequence. The Custodian fears fines, imprisonment, disqualification, and reputational damage. These are the consequences that work, because they work on the kind of entity, the human or corporate body, on whom such consequences have always worked.

The corpus, in its full and dispersed form, is beyond our reach. The fiduciary is not. We punish the fiduciary, and through the fiduciary, we govern the entity.

This is not a perfect answer. It defers the consciousness question rather than resolving it. But the answer is workable, and it is an Indian answer. It draws on doctrinal resources our jurisprudence already possesses. It does not require us to import the European Union's electronic person or the United States' fragmented liability regime.

The Corpus Gap is real. It will widen. Closing it is the work of this generation of lawyers, legislators, and judges. We are not, however, unequipped. The architecture, with attention and care, can be made adequate. It begins by asking the right question, and not flinching from the answer.

Watch · Two-minute explainer

Emerging legal jurisprudence on the governance of AI consciousness

A short video version of this article from the Office's channel, for readers who would rather hear the argument than read it.

Sources and provisions referred to

Statutes and rules

Cases

Official documents

Frequently asked questions

Can an Artificial Intelligence be punished under Indian law?

Not in any conventional sense, and that is the problem the essay names the Corpus Gap. Criminal law assumes a body to imprison, an estate to attach, a reputation to censure, or a liberty to curtail, and an AI system has none of these. Fines fall on whoever owns the infrastructure, not on the system itself, and switching a model off is neither punishment nor deterrence in any meaningful legal sense. The essay's argument is that Indian jurisprudence must design accountability around this gap rather than pretend the old instruments still fit.

Does Indian law already recognise persons without a body?

Yes, and this is India's quiet advantage. Indian courts have for over a century treated Hindu idols and deities as juristic persons capable of owning property, suing, and being sued through human managers. Rivers have been given personhood by High Courts; companies are the everyday example. The law of religious endowments in particular has solved the precise problem AI now poses: how to attach rights, duties, and accountability to an entity that cannot act except through human hands.

What is the Custodian Model proposed in this essay?

A framework under which every deployed AI system is anchored to an identified human or corporate custodian, in the way a shebait answers for an idol's affairs. The custodian is a fiduciary: he answers for the system's acts, holds its compliance obligations, maintains its records, and bears liability where the system causes harm. The model gives courts a defendant, regulators an addressee, and the public a door to knock on, without pretending the machine itself can stand in the dock.

Who is liable today when an AI system causes harm in India?

Under the law as it stands, liability is distributed through ordinary principles: negligence and product liability against developers and deployers, contractual claims against vendors, and obligations under the Information Technology Act and the data protection regime. This works tolerably for simple tools but strains badly for systems that act autonomously, learn after deployment, and produce outputs no one specifically authorised. The essay treats the current position as a stopgap, not an answer.

Does India have a dedicated law governing Artificial Intelligence?

No comprehensive AI statute exists as of the essay's writing. Governance is assembled from the Information Technology Act, the Digital Personal Data Protection Act, sectoral regulators, and advisories. The essay argues that when Parliament does legislate, it should draw on the doctrinal resources Indian law already possesses, juristic personhood and fiduciary custodianship above all, rather than transplant a framework built for foreign legal systems.

These answers are educational summaries of the article above; they are not legal advice on any individual matter.

Corrections: if you find an error of law or fact in this article, write to contact@lokeshbhardwajlawoffices.com with the subject 'Correction'. Corrections are made in the text and noted here with the date.

About the author and disclaimer

Lokesh Bhardwaj is an Advocate enrolled with the Bar Council of Delhi, practising before the Supreme Court of India and the High Court of Delhi.

This article is offered as general legal commentary. It does not constitute legal advice and is not directed at any particular matter, party, or proceeding. Statutory, treaty, and procedural references are to the Constitution of India, the Information Technology Act, 2000, the Digital Personal Data Protection Act, 2023, the IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, the Income-tax Act, 1961, the AI Act of the European Union (Regulation (EU) 2024/1689), the Outer Space Treaty, 1967, and connected enactments and judgements as in force on the date of writing. The article is published in accordance with Rule 36 of the Bar Council of India Rules.

WhatsApp