India’s AI Governance System Has a Coordination Problem

On July 2, India’s Supreme Court set aside two insolvency decisions that had relied on six citations that were either nonexistent, unrelated, or accompanied by fabricated passages. Two referred to genuine judgments but included invented passages. One came from an unrelated case and also contained a fabricated passage. Three did not exist.  

The National Company Law Tribunal had used the material in its decision. The National Company Law Appellate Tribunal then upheld that decision without detecting the errors. The problem reached the Supreme Court only after surviving two levels of adjudication. 

The apex court held that any fake or hallucinated material entering judicial reasoning made the result “no decision in the eyes of the law.” It adopted a zero tolerance approach to unverified AI-generated precedents and directed the Bar Council of India to develop preventive guidance and disciplinary consequences. 

India’s response to such risks has accelerated. The government released its national AI Governance Guidelines in November 2025. Binding amendments to the Information Technology Rules took effect in February 2026. The government formed the AI Governance and Economic Group and its expert committee in April. The Supreme Court and the Reserve Bank of India released separate draft frameworks in June. 

These measures do not form a single regulatory code. The IT Rules are binding. The national guidelines are advisory. The court and RBI frameworks remain drafts. Most operational provisions of the Digital Personal Data Protection Act will not begin until May 13, 2027.

The government is now reconsidering whether that structure is enough. On July 3, MeitY Secretary S. Krishnan said discussions on separate AI legislation had begun. He offered no timetable and stopped short of announcing a bill. 

The immediate challenge for companies is therefore not an absence of rules. It is the lack of a common method for translating those rules into evidence, operational responsibility and enforceable controls. 

Existing Law Works Best After Harm Occurs 

India’s current approach is largely technology-neutral. It regulates conduct and consequences rather than trying to write a separate rule for every kind of AI.

A payment obtained through a cloned voice may still constitute fraud. Defamation does not become lawful because a model produced the statement. Consumer, privacy, cybersecurity and intellectual-property law can apply without a comprehensive definition of artificial intelligence.

Prashant Mali, a cyber, privacy and AI lawyer and founder of Cyber Law Consulting, described this as a practical starting point.

“The future is not regulating artificial intelligence,” he said. “The future is regulating artificial harm.”

The approach works most clearly when AI operates inside an already regulated activity.

“An AI-assisted loan decision can still be examined under banking rules,” said Dinesh Jotwani, co-managing partner at Jotwani Associates. “An insurance chatbot that mis-sells a policy can still trigger insurance and consumer-protection obligations.”

Indian courts have applied the same logic to synthetic impersonation. In Anil Kapoor v. Simply Life India, the Delhi High Court restrained unauthorized commercial use of the actor’s name, image, voice and other personal attributes, including through AI and face-morphing tools. In Arijit Singh v. Codible Ventures, the Bombay High Court granted interim protection against AI-enabled misuse of the singer’s voice and persona.

Samarth Luthra, an advocate enrolled with the Bar Council of Delhi and registered as a foreign lawyer in England and Wales, said such orders do “more than recognize privacy or personality rights.” They also indicate how courts expect platforms and service providers to respond to synthetic impersonation, he said.

But these decisions are fact-specific and largely remedial. They usually address visible harm after it has occurred. They do not set economy-wide requirements for evaluating models before deployment.

The Supreme Court’s draft regulations illustrate the distinction. They reserve decisions on law, fact and justice for judicial officers. AI must remain assistive, and approved systems would be subject to human supervision and verification. Those safeguards apply to the judicial system, not to every bank, hospital or employer.

The amended IT Rules are similarly targeted. They define realistic synthetic audio, visual and audiovisual material and extend intermediary duties to that content. They do not establish an equivalent regime for ordinary text generated by chatbots.

Existing law can therefore identify many recognizable wrongs. It is less useful in answering three earlier questions: How should a system have been tested, what evidence should have been retained and which participant should have prevented the failure?

Responsibility Frays Across the AI Supply Chain 

The coordination problem begins with data.

Once the relevant DPDP provisions take effect, personal data will generally require consent or another basis specified in the Act. The law excludes data made public by the individual and information another party is legally required to publish. That exclusion does not resolve copyright, contractual restrictions or other legal duties.

“The internet may be publicly accessible, but it is not a free raw-material warehouse for artificial intelligence,” Mali said.

He advised companies to favor licensed datasets, genuinely anonymized or synthetic information, contractual data-sharing arrangements and clearly applicable statutory grounds.

Malcolm Gomes, chief operating officer at Privy by IDfy, said companies must be able to trace what personal information a system uses, where it came from and which permission covers it. They must also follow the data as it moves through applications, retrieval systems, data lakes and models.

Erasure creates a harder technical problem. Removing a customer record from a conventional database is straightforward. Proving that the information no longer affects a trained or derived model is not. Machine unlearning remains expensive and unreliable at scale, Gomes said.

Copyright adds a separate uncertainty. A DPIIT working paper proposed a hybrid system involving a blanket license and statutory remuneration for copyrighted material used in AI training. It remains a policy proposal, not law.

Responsibility also becomes unclear when several companies contribute to one product.

The Calcutta High Court confronted that problem in May. It said deciding whether ChatGPT was an intermediary or an originator was a “complicated and vexed” question requiring technical evidence. At the preliminary stage, it treated ChatGPT as an originator because it generated new content rather than merely transmitting third-party material.

The government did not create a separate classification in its July 29 parliamentary response. It said an AI service’s intermediary status and eligibility for safe harbor would depend on the service and the functions it performed.

Finance offers a more developed model. The RBI’s June draft would make a regulated institution accountable for every model it uses, including models bought from outside providers. It would require a board-approved model-risk framework, an inventory, independent validation and continued monitoring.

The draft also requires institutions to validate third-party models themselves, despite assurances supplied by the vendor. Contracts would need to provide technical information, audit rights and exit arrangements. Models with greater autonomy or customer impact would need stronger controls, including the ability to override, suspend or deactivate them.

Companies are beginning to make their own divisions of responsibility.

Mandar Patil, executive vice president at Cyble, said the company conducts provenance checks, documents model behavior and applies stronger human controls to higher-risk uses.

Cyble acts as an application developer and systems integrator. Patil said it accepts responsibility for configuration, access controls, monitoring and application-level safeguards. Foundation-model providers remain “responsible for the underlying model, while customers define business uses and govern their own data,” he said.

That may be a workable commercial arrangement. It is not yet a generally accepted legal division.

Krupesh Bhat, founder and CEO of Melento, formerly SignDesk, said explainability, auditability and human oversight must be designed into products from the beginning.

“We now design every agent assuming its decisions may one day need to be explained to a regulator, defended to a customer and trusted by an employee,” he said.

Bhat distinguished systems that create content from systems that make or execute business decisions. A marketing draft and the automated denial of a loan do not justify the same controls.

These accounts describe company practices. They do not independently prove that the safeguards work.

The burden will also fall unevenly. Large companies can retain specialist lawyers, risk teams and auditors. Smaller businesses often depend on models they cannot inspect and contracts they cannot negotiate. Without common evidence standards, every company can define “risk-based governance” in its own favor.

Governance Requires Three Kinds of Proof 

Leaders do not need to wait for every legal question to be settled. They do need evidence that a system is controlled.

A practical governance model should require three kinds of proof.

Traceability

The company should know which AI systems it uses, which data they touch, which outside providers are involved and which business decisions depend on them.

Testing

The company should have documented evidence covering performance, known limitations, foreseeable failures, bias, security, hallucinations and changes after deployment.

Intervention

A named person or function should have the authority to challenge, override, restrict or suspend the system. Human review should be real rather than a ceremonial button placed beside an automated decision.

Consider an AI system used to assess loan applications. Traceability identifies its data, vendors and downstream decisions. Testing examines errors, bias, drift and performance across customer groups. Intervention gives a qualified employee the evidence and authority needed to reverse or stop a decision.

The three proofs also expose weak vendor arrangements. A provider that offers no usable documentation may prevent the deploying company from validating the product. A contract can transfer financial exposure, but it cannot manufacture evidence that the buyer does not possess.

Implications for Leaders

C Suite

Within 90 days, senior management should assign one executive responsibility for the organization’s AI inventory and governance process.

The inventory should include both approved tools and informal employee use. Each system should be classified by the consequence of its decisions, not by the novelty of its model.

Executives should also decide which risks the company will not accept. A system affecting credit, employment, health, safety or legal rights should not enter production merely because a vendor calls it “enterprise-ready.”

Functional Leaders

Product, data, risk, legal and security leaders should create an evidence pack for each consequential system.

It should record the intended purpose, data sources, permissions, known limitations, validation results, monitoring metrics, outside providers and escalation process.

Vendor reviews should test whether the organization can meet its own duties when the provider controls the model. The answer cannot rest on a contractual promise the buyer is unable to verify.

Boards and Governance

Boards should ask four questions before approving a consequential AI system:

1. What harm could the system cause, and to whom?
2. Which participant can prevent, detect or contain that harm?
3. What evidence shows that the system has been tested and is being monitored?
4. Who can challenge, suspend or override its output?

The board does not need to understand every technical detail. It does need to know whether management can produce credible answers.

India’s new institutions could help standardize those answers. The AI Governance and Economic Group can coordinate policy across ministries. The Technology and Policy Expert Committee can translate technical issues into policy. The IndiaAI Safety Institute is intended to develop testing methods, standards and research capacity.

Their success will depend less on the number of documents they publish than on whether companies know what evidence to keep and who remains responsible when a system fails.

India does not chiefly lack AI rules. It lacks a common way to turn those rules into evidence, authority and accountability.

A dedicated AI law could provide that connective structure. It would need common risk categories, minimum testing standards, incident reporting and clearer responsibility across the supply chain.

Another broad statute that leaves those questions unresolved would add regulation without creating governance.