Authored by Dr. Pavan Duggal, Senior Advocate, Supreme Court of India, Global Chair, International AI Accountability Forum
Every legal system on the planet has been built on one silent assumption: that behind every act there stands a human being who can be summoned, questioned, and held to account. Agentic artificial intelligence has now quietly demolished that assumption. An autonomous AI agent today receives a goal, plans its own course of action, executes across multiple tools and environments, and produces consequences in the real world, without any human having authored the specific act that caused the harm.
There is no denying the fact that this is not a distant scenario. Agentic AI is already booking, transacting, negotiating, trading, and deciding on behalf of humans across every major economy. Yet when an AI agent causes harm, our legal systems still go looking for a driver, a manufacturer, a signatory, a mind. They find none. That is the liability vacuum, and it is widening every single day.
The world is presently in a peculiar position. Some jurisdictions have enacted AI laws, including the European Union, China, South Korea, Japan and a handful of others. Most of these laws, however, are built for AI that advises. They are not built for AI that acts. They classify risks, mandate transparency, and impose obligations on providers and deployers. But they do not answer the one question that actually matters to a victim: who pays when an autonomous agent, acting on its own plan, destroys my business, empties my account, or injures my child?
The fact remains that the traditional doctrines of tort, contract, agency and product liability were never designed for an actor that has no legal personality, no assets, no intention in the human sense, and no capacity to be punished. Negligence requires a duty of care owed by a person. Vicarious liability requires a relationship of control. Product liability requires a defect that existed at the time of sale, whereas an agentic system rewrites its own behaviour long after it has left the factory. Each of these doctrines fails, not at the margins, but at the core.
I have been advocating for many years that this is not a gap that national legislatures can close on their own. An agentic AI system is developed in one country, trained on data from fifty countries, deployed from a server in a third country, and causes harm in a fourth. A purely national liability law will always be one jurisdiction short of the actor it seeks to reach. This is precisely why the liability question is, at its heart, an international question.
That is why, in March 2026, I authored and released into the public domain the Global Agentic AI Liability Framework. It is built on a simple proposition: liability for agentic AI harm cannot rest on a single actor, because no single actor controls the outcome. It must be distributed across the chain of development, deployment, operation and benefit, and it must attach to whoever is best placed to prevent the harm and to whoever profits from the risk. The framework contains fifteen entirely new legal doctrines and theories which do not presently exist in any legal regime in the world. They were not written for academic elegance. They were written because the existing doctrines have run out of road.
Some of these are simple in their statement and radical in their consequence. If a human cannot answer for an agentโs act, then the entity that unleashed the agent must answer, without the shelter of intervening autonomy. If an agentโs conduct cannot be explained, the burden of proof must shift to those who built it. If an agent operates across borders, then the liability that follows it must also operate across borders. These are not exotic ideas. They are the minimum conditions for any legal order that intends to survive the agentic turn.
The AI Accountability Framework 2026 sets out the coalition of legal principles on which such accountability must rest. The Universal Declaration of AI Accountability Rights, adopted at the International AI Accountability Forum in New Delhi on 14 May 2026, translates those principles into rights that every human being can assert against AI systems and against those who deploy them. The New Delhi Compact on AI Accountability then provides countries with concrete mechanisms to implement those rights. Quiet but substantive work, therefore, already exists in the public domain. What does not exist is the political will to bind nations to it.
Companies developing frontier and agentic models are not bothered about any of this. Their attention is fixed on the next release, the next benchmark, the next feature. Liability, in their worldview, is a problem for someone else, to be resolved later, preferably by a court in a jurisdiction of their choosing. That is exactly why the liability question must be settled at the level of an international instrument, and not left to be litigated piecemeal after the damage has been done.
An International Convention on Agentic AI, AGI, and Superintelligence would do for AI liability what no national law can do. It would establish common minimum standards of accountability that follow the agent wherever it operates. It would settle the question of who answers when the actor cannot answer. It would create mechanisms for cross-border recognition of liability findings and for the enforcement of compensation. And it would put frontier developers on notice that the era of building actors without answering for them is over.
Global processes, unfortunately, are moving at the speed of diplomacy while agentic AI is moving at the speed of computation. The United Nations has convened a scientific panel and a global dialogue on AI governance. These are welcome, but they are not liability instruments, and they are not binding. The gap between what is being discussed in New York and Geneva and what is actually happening in the worldโs data centres is now so wide that the discussion risks becoming a commentary on events it can no longer influence.
It will take just one catalyst, one large-scale harm caused by an autonomous agent that no court in the world can attribute to anyone, for nations to realise that the liability vacuum is a shared emergency. The question is whether we will build the convention before that catalyst or after it.
The time has come for countries, courts, bar associations, academic institutions and industry bodies to begin working in right earnest towards common legal principles for agentic AI liability. Non-state actors have as much of a role in this as state actors. The doctrines exist. The frameworks exist. The Declaration and the Compact exist. What is missing is the decision to make them binding at the international level.
An actor that cannot answer is an actor that must be answered for. That is the entire case for an International Convention on Agentic AI, AGI, and Superintelligence. The quicker the world accepts it, the fewer victims will be left standing before empty courtrooms.
Let us wait and watch how things evolve in this regard in the coming times.
๐๐ญ๐๐ฒ ๐ข๐ง๐๐จ๐ซ๐ฆ๐๐ ๐ฐ๐ข๐ญ๐ก ๐จ๐ฎ๐ซ ๐ฅ๐๐ญ๐๐ฌ๐ญ ๐ฎ๐ฉ๐๐๐ญ๐๐ฌ ๐๐ฒ ๐ฃ๐จ๐ข๐ง๐ข๐ง๐ ๐ญ๐ก๐ WhatsApp Channel now! ๐๐ฒ
๐ญ๐๐๐๐๐ ๐ถ๐๐ ๐บ๐๐๐๐๐ ๐ด๐๐ ๐๐ ๐ท๐๐๐๐ฌ ๐ Facebook, LinkedIn, Twitter, Instagram