Why India’s upcoming AI law must regulate autonomous agents, not just deepfakes - ET Government
An autonomous AI agent is software given authority to take actions with legal, financial or physical consequences without a human approving each action. That authority is what makes it different from a tool that only responds to questions. An agent can affect a person’s money, data or contracts, not just what they read or believe. For example, a person might ask a chatbot how to file a return and then decide what to do. An agent might file the return itself, approve a loan or move money. The article highlights this shift from giving information to carrying out actions on someone’s behalf. The person may not press a final button each time. The article says India’s consultation paper should define agents by the authority they have, rather than by the technology used to build them. It also notes that the Contract Act recognizes agency but does not recognize AI as an agent. That leaves open whether an agent’s decision binds the person it acts for.
What is an autonomous AI agent, and how is it different from a chatbot that only gives answers?
An autonomous AI agent is software given authority to take actions with legal, financial or physical consequences without a human approving each action. That authority is what makes it different from a tool that only responds to questions. An agent can affect a person’s money, data or contracts, not just what they read or believe.
For example, a person might ask a chatbot how to file a return and then decide what to do. An agent might file the return itself, approve a loan or move money. The article highlights this shift from giving information to carrying out actions on someone’s behalf. The person may not press a final button each time.
The article says India’s consultation paper should define agents by the authority they have, rather than by the technology used to build them. It also notes that the Contract Act recognizes agency but does not recognize AI as an agent. That leaves open whether an agent’s decision binds the person it acts for.
Why can an AI agent that moves money or signs a contract create different risks from a deepfake?
A deepfake can mislead people about what they see or hear. An autonomous agent can cause a different kind of harm: it can act on someone’s money, data or contracts. The article describes this as the difference between harms to belief and harms caused by actions. The second kind may require its own rules.
For instance, an agent might approve a loan, file a return or move money without a person approving that specific action. If something goes wrong, the effects are not limited to false content; someone may face a financial loss or a dispute over a decision made in their name. The article says current law is silent on whether an agent’s decision is binding on the person it acts for.
The article says India’s planned AI consultation should address agents as well as deepfakes. It proposes graded liability across the people and companies involved, and raises the need for evidence and clear accountability. These issues matter because a harmed person may struggle to show what an agent did or who should answer for it.
How many parties might share responsibility when an AI agent causes harm, and who are they?
The article identifies four parties whose contributions may matter when an AI agent causes harm: the model creator, integrator, deployer and user. It does not say that all four must always share responsibility. Instead, it argues against automatically blaming only the builder, because an agent’s behavior may not be predictable by the person who created the model.
Consider an agent deployed by a company to handle customer transactions. The model creator supplied the underlying model, an integrator connected it to systems, the enterprise deployed it, and a user may have given it instructions or authority. Those roles can differ, so responsibility should reflect what each party contributed and whether it took due care.
Dr. Pavan Duggal proposes this graded approach. The article also says a harmed person may not be able to prove what the agent did. Experts therefore suggest a presumptive burden of proof: after harm is claimed, those who built, integrated and deployed the agent should show they took due care. The article stresses that blaming “the algorithm” should not be enough.
What could happen if an AI agent makes a decision that is not legally recognized as its user's decision?
The article raises a basic legal uncertainty: if an AI agent makes a decision on someone’s behalf, does that decision count as the person’s own? The Contract Act recognizes agency, but the article says it does not recognize AI as an agent. It describes the law as silent on whether an AI agent’s decision binds the person it acts for.
That uncertainty could matter if an agent signs a contract, moves money or takes another action with legal or financial consequences. The person, a business or another party could dispute whether the action was authorized or binding. The article does not say how a court would resolve such a dispute, so the outcome should not be assumed.
The article argues that India’s planned AI regulation should make this area clearer. It also calls for graded liability and better evidence, since a person harmed by an agent may be unable to prove what happened. Clear rules could help identify who must answer for an action and what remedy may be available, while the legal question about binding decisions remains unresolved in the text.
Why do AI agents need tamper-evident action logs, and how could those logs help someone seeking a remedy?
The article says a person harmed by an AI agent usually cannot prove what it did, leaving courts unable to offer a remedy. It does not specifically call for tamper-evident action logs. As a practical measure beyond the article’s express proposals, records that are difficult to alter could help preserve evidence of an agent’s actions.
For example, such a log might record what instruction the agent received, what action it took, when it acted and which system or person authorized it. If a money transfer or data decision is disputed, those records could help show what happened and which parties were involved. A log would not by itself establish liability or prove that an action was lawful.
The article’s proposed response to the evidence problem is a presumptive burden of proof: once harm is claimed, those who built, integrated and deployed the agent should show they took due care. Reliable logs could help them provide that evidence and help an affected person explain a claim. The article’s broader point is that “the algorithm did it” should not be an adequate defense.
Which existing Indian rules could apply when an AI agent mishandles personal data or suffers a serious security incident?
The article identifies several existing Indian rules and frameworks that may matter when enterprises use AI agents. For personal data, it points to the Digital Personal Data Protection Act, or DPDP Act, and says its consent manager framework comes into force that week. It says the Act carries penalties of up to ₹250 crore.
For a serious cyber incident, the article cites CERT-In’s 2022 directions, which require incidents to be reported within six hours. It also says the 2026 amendments to the IT Rules require platforms to label AI-generated audio and video. The RBI’s FREE-AI framework has recommended board-approved AI policies for banks and other regulated entities.
The article warns that an agent may read customer records, combine them with other data and act on the results in ways the original consent did not describe. It argues that the consultation paper should state plainly that the enterprise deploying an agent is the data fiduciary for its actions. These existing measures do not remove the need for clearer agent-specific rules.
What does it mean in law for someone to act as an agent, and why does granting authority create responsibility?
In ordinary legal usage, agency is a relationship in which one party is authorized to act on another’s behalf. The article says India’s Contract Act recognizes agency, but does not recognize AI as an agent. It therefore asks whether a decision an AI makes for someone is binding on that person, and says the law is silent on that question.
Authority matters because an agent may do more than provide information. The article describes software that can take actions with legal, financial or physical consequences without a human approving each action. Examples include approving a loan, filing a return or moving money. If authority is granted, the actions can affect the person or organization that deployed the software and others who deal with it.
The article argues that regulation should focus on the authority given to an agent, not the technology itself. It also recommends graded liability based on each party’s contribution and stresses the need for clear accountability. Those rules could help address who is responsible when an agent acts, while the question of whether its decisions legally bind a user remains open in the article.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web