Industry experts stated that India should develop its own artificial intelligence (AI) safety framework that aligns with its linguistic characteristics, scale, and application-oriented ecosystem, rather than simply copying the voluntary agreement signed last week by US President Donald Trump and leading technology companies.
They also called for mandatory commitments involving all stakeholders aimed at addressing risks such as deepfakes and cyberattacks, despite disagreements over the sufficiency of voluntary measures alone.
These comments came after the adoption of the White House Agreement on Superintelligence—a voluntary AI safety pact signed on September 29 by Trump and top executives from Google, OpenAI, Anthropic, Meta, Nvidia, and Elon Musk's xAI.
Under this agreement, companies committed to implementing internal control mechanisms to monitor their AI models, establishing specialized oversight groups, conducting independent external audits, and ensuring board-level governance. Furthermore, they agreed to meet regularly to establish safety standards.
Nasscom Chairman and Fractal Group CEO Shrikant Velamakanni noted that India should not blindly copy the US model. Instead, it should use internal commitments as a foundation for achieving global understanding of AI safety issues.
Velamakanni emphasized: 'As models become more capable and agentic, we grant them greater access to systems and more autonomy in decision-making. AI safety must be significantly enhanced and must outpace capabilities to prevent catastrophic risks.'
He added that the American agreement is largely a call for self-regulation by companies developing advanced AI. In contrast, the European Union adopted a mandatory, risk-based approach, while China follows a state-governance model.
According to the Nasscom chairman, India can leverage its role in the Global Partnership on AI (GPAI) and the momentum gained from the February AI Influence Summit to avoid worst-case scenarios without denying citizens of India and the Global South access to advanced models.
Soumen Mandal, Senior Analyst at Counterpoint Research, stated that India does not need to copy the American model but can draw inspiration from it to accelerate efforts to build a robust AI ecosystem. He specified that such a commitment should cover pre-deployment testing of AI systems, independent auditing, serious issue reporting, and assigning accountability for AI-related risks.
Amit Khanna, Partner and Automation Ecosystem Head at Grant Thornton Bharat, believes that any commitment must suit the conditions of the Indian market. He noted that India's focus is on creating AI applications, not developing large foundational models.
Khanna stressed: 'The industry commitment in India must ensure reliable AI performance in local languages and unique cultural contexts and must not exploit knowledge asymmetry in the population.'
Ganesh Gopalan, Co-founder and CEO of Gnani.ai, stated that any framework must ensure a level playing field. He insisted that global advanced models must adhere to the same rules as Indian ones, since Indian AI companies compete with global firms for the same clients.
Experts were divided on whether self-regulation is sufficient. Velamakanni stated that 'Voluntary, self-regulatory approaches are necessary but extremely insufficient.' He cited examples from pharmaceuticals and aviation where strong, technically competent regulators enhanced safety and public trust without halting innovation.
He concluded that India needs mandatory rules regarding transparency, incident reporting, and accountability in case of serious harm. He also proposed strengthening the IndiaAI Safety Institute to test new models, but without creating a bottleneck by requiring prior certification of every model.
Mandal supported a hybrid approach. Mandatory rules should apply to potential harm in defense, medical, banking, and government sectors, while voluntary standards should cover rapidly evolving technical areas where regulators may lack technical expertise.
Khanna also advocated for a hybrid 'techno-legal' approach, combining a strong voluntary layer with mandatory mandates, such as provisions on synthetically generated information in the IT Amendment Rules of 2026. He noted that voluntary compliance leaves significant gaps, especially concerning civil rights and public safety.
Gopalan, however, preferred self-regulation. 'Voluntary self-regulation is crucial given the novelty of the technology. The presence of mandatory regulation could stifle competition and raise concerns about preferential treatment for certain companies,' he stated. He added that India needs more advanced AI labs to prevent colonization, which is inevitable if the country lags in the AI race.
The Indian government is in the process of developing its own AI regulation. Earlier this year, IT Secretary C Krishnan noted that while existing legal norms are currently adequate to address initial issues like deepfakes and synthetic content, 'additional regulation or legislation may be required.'
He added that 'this is a conversation that has started, and my minister (IT Minister Ashwini Vaishnaw) and I previously stated that we would consider AI regulation when the time comes, and it seems the time is coming, and we will start considering it.'
Vaishnaw also said that current IT legislation was developed long before the rapid emergence of AI, and a new legal framework may be needed to address the changing landscape.
The industry agrees that any pact must extend beyond just AI companies. Velamakanni stated: 'It cannot be an agreement only between AI companies.' He stressed that the roles of the government, advanced model developers, major users, including enterprises and government agencies, academia, and independent safety researchers must be defined. Workers, users, and civil society must have official channels to report harm.
He also added that safety assessments of advanced models should be conducted by independent expert institutions and published, except in narrow cases where disclosure poses a security threat.
Mandal believed the agreement should encompass the entire AI ecosystem: from AI companies and their business users to regulators, universities, researchers, workers, and users. Khanna argued it should include governmental and public entities that are major AI users, as well as civil society, workers' unions, and technology companies. Gopalan, meanwhile, called for broader self-regulation across the entire AI ecosystem to ensure impartial access to limited computational resources.
Velamakanni stated that the primary concern must be preventing catastrophic and systemic harm. He listed large-scale cyberattacks, election interference, biological or chemical weapons, attacks on critical infrastructure, and loss of control over highly autonomous systems. Personal harms, such as deepfakes, fraud, impersonation, privacy leaks, and self-harm, follow these.
He emphasized: 'Autonomy must lag behind safeguards to provide a safety buffer. Since AI systems are gaining the ability to act in the world (using tools, moving money, altering systems, making important decisions), evidence of their safety and controllability must grow faster than the autonomy we grant them.'
Gopalan noted that given India's large population and democratic system, deepfakes and cyberattacks targeting financial institutions are more serious problems.
Khanna also named deepfakes as a priority issue. He described disinformation, deepfakes, and political manipulation as 'existential threats to the social fabric of India.' He pointed out that liability rules in case of AI system failures, such as misdiagnosis of a patient or erroneous denial of benefits, remain a critical gap.
Mandal prioritized independent auditing, followed by accountability, data protection, deepfakes, and employment impact. He stated: 'Without independent testing and auditing, the system will not be complete, as companies can sometimes overlook AI-related issues.'
