India’s Proposed Frontier AI Law: What Could the Regulatory Framework Look Like?

India’s Proposed Frontier AI Law: What Could the Regulatory Framework Look Like?

Article by

AI Law

Introduction 

India may be moving towards a dedicated law for artificial intelligence after several years of relying on existing statutes, sectoral regulators and non-binding governance principles. In October 2026, the government confirmed that it intends to release a consultation paper on a new AI law, while recent reporting suggests that frontier AI models, agentic systems, liability and pre-release safety testing may form part of that discussion. 

No draft bill has yet been published, so the exact framework remains unsettled. What is clear is that the policy discussion is moving towards a more differentiated approach, where ordinary AI applications may continue to operate under existing laws while highly capable systems face additional safeguards because of the scale of harm they could potentially cause. 

A Shift from India’s Existing Approach 

India’s current AI governance framework is deliberately flexible. The India AI Governance Guidelines, released in November 2025, favour a principle-based, risk-based and proportionate approach, with sectoral regulators remaining responsible for enforcement within their existing mandates. The government has also stated that the guidelines do not contemplate unrestricted deployment of high-risk AI systems. 

The guidelines originally suggested a view that a new AI law was not immediately necessary. The present discussion suggests that the government may now be considering whether advanced AI capabilities create gaps that existing legislation cannot comfortably address. A frontier AI regime could preserve India’s relatively light-touch approach for ordinary systems while creating stronger controls at the higher end of the risk spectrum. 

What Should Count as Frontier AI? 

The first challenge is definitional. A law cannot impose special obligations on frontier models unless developers and regulators can identify which systems fall within that category. 

One option is to rely on technical thresholds, such as the amount of computing power used during training. That gives regulators a measurable trigger. Secondly, A capability-based approach may also be explored where the regulators could examine whether a model can autonomously identify software vulnerabilities, operate critical systems or perform other tasks capable of producing significant harm.  

A practical framework may need both approaches. Technical thresholds could identify systems requiring further scrutiny, while capability evaluations determine whether stronger regulatory obligations are justified. 

Pre-Release Testing Could Become Central 

One of the more significant ideas under discussion is the possibility of pre-release evaluation for highly capable models. Officials are reportedly examining whether some systems should undergo controlled testing or sandbox-based assessments before wider deployment. 

Such evaluations could test whether a model bypasses safety controls, produces dangerous outputs, exploits technical vulnerabilities or behaves unpredictably when connected to external tools. The value of this approach is that it moves part of the regulatory process to a stage before harm occurs, rather than relying entirely on enforcement after deployment. 

India has already begun building the institutional infrastructure required for such testing. The IndiaAI Safety Institute is intended to undertake safety research, develop standards, test AI systems and evaluate emerging risks. A future law could give this testing ecosystem a clearer regulatory role while reserving mandatory evaluations for systems that cross defined risk or capability thresholds. 

Agentic AI Raises a Different Set of Risks 

Frontier AI regulation will also have to account for agentic systems, which can plan tasks, use external tools and take actions with limited human supervision. These systems create risks that are different from those posed by a conventional chatbot or content generator.

Government discussions are reportedly examining limits on agent autonomy and possible safeguards for high-risk applications. An agent connected to banking systems, email accounts, cloud infrastructure or enterprise software can cause consequences beyond producing a wrong answer. It may send information, initiate transactions, alter records or interact with other systems before a person has reviewed the action. 

A proportionate framework could require human approval for specified high-impact actions, restrict access to critical tools or sensitive datasets and require reliable mechanisms for stopping unsafe behaviour. These obligations would make most sense where the agent has meaningful operational autonomy rather than applying to every AI-assisted workflow. 

Liability Across the AI Value Chain 

Liability is likely to be one of the most difficult issues in a future AI law because modern AI systems rarely involve a single actor. A foundation-model developer may build the underlying system, another company may fine-tune it, a platform may provide access, and an organisation may deploy it for a specific purpose. 

A workable framework would need to allocate responsibility according to control. A developer should not automatically bear responsibility for every downstream misuse, but deployers should also not be able to attribute foreseeable model-level failures entirely to the developer. 

India’s AI Governance Guidelines already place considerable emphasis on accountability and on identifying responsibility across the AI ecosystem. A statutory framework could develop this further by distinguishing failures in model design, inadequate deployment safeguards, poor monitoring and deliberate misuse by users. 

Clear allocation of responsibility would also make enforcement easier. Where serious harm occurs, regulators and affected individuals should be able to identify which organisation controlled the relevant decision and which safeguards it was expected to maintain. 

Who Would Regulate Frontier AI? 

India may not need to establish an entirely new regulator. The institutional framework has already begun to take shape through the AI Governance and Economic Group, the Technology and Policy Expert Committee and the IndiaAI Safety Institute. 

AI Governance and Economic Group can coordinate policy across ministries, while technical bodies can support evaluations and standards. Existing regulators such as RBI, SEBI, TRAI, CERT-In and the Data Protection Board can continue to deal with AI-related risks falling within their sectors. 

This model also avoids centralising every AI issue in one authority. A frontier-model developer may need technical safety oversight at the model level, while a bank deploying that model may remain subject to RBI requirements for its specific use. 

Avoiding Over-Regulation of Ordinary AI 

The success of any frontier AI framework will depend heavily on proportionality. A company training a highly capable foundation model does not create the same regulatory risk as a small business using an existing model to summarise documents or automate routine tasks. 

India’s current governance approach repeatedly emphasises innovation over restraint and the need to match regulation to identifiable harm. Extending frontier-level obligations to every business using AI would undermine that principle and create significant compliance costs without necessarily addressing the systems capable of causing the greatest harm. 

What Should Developers Prepare for? 

The direction of policy makes some preparation sensible.

  • Track model capabilities: Developers should know how capabilities change between model versions and whether systems can access external tools or infrastructure. 


  • Strengthen testing: Red-teaming and controlled evaluations can identify risks before deployment. 


  • Document safety controls: Human oversight, access restrictions and model safeguards should be recorded clearly. 


  • Allocate responsibility: Contracts should distinguish the roles of model developers, deployers and service providers. 


  • Prepare for serious incidents: Advanced AI providers should have mechanisms to identify and escalate major safety failures. 

These practices also fit the government’s existing emphasis on safety, accountability and technical risk management, even before any dedicated frontier AI legislation is enacted. 

Want to Stay Ahead? 

Reach out to the experts at GoTrust today. 


 

Found this useful? Share it.

WhatsAppFacebookXLinkedIn
WhatsAppFacebookXLinkedIn

Automate AI & Privacy Compliance Assessments

Managing AI and privacy compliance can be complex as regulations continue to evolve. GoTrust helps simplify this process by connecting AI systems with ISO/IEC 42001 and NIST AI RMF requirements. It also supports privacy automation and cookie consent management, helping businesses identify compliance gaps, manage assessments, and stay prepared for audits.

Automate AI & Privacy Compliance Assessments

Managing AI and privacy compliance can be complex as regulations continue to evolve. GoTrust helps simplify this process by connecting AI systems with ISO/IEC 42001 and NIST AI RMF requirements. It also supports privacy automation and cookie consent management, helping businesses identify compliance gaps, manage assessments, and stay prepared for audits.

Automate AI & Privacy Compliance Assessments

Managing AI and privacy compliance can be complex as regulations continue to evolve. GoTrust helps simplify this process by connecting AI systems with ISO/IEC 42001 and NIST AI RMF requirements. It also supports privacy automation and cookie consent management, helping businesses identify compliance gaps, manage assessments, and stay prepared for audits.

GoTrust