Building bigger AI models sans safety systems worrying: WSAI head

Yes. We’re moving towards systems where one AI agent can set a task for another AI programme or agent, communicate with it and potentially interact with the outside world. That gives AI much broader capabilities, but also creates the possibility of unintended consequences that we may not be able to control. Even one life lost because of an AI system’s unintended action would be too high a cost. That’s why we need strong guardrails and safety engineering. AI harnesses and narrow, custom-built solutions can help control what these systems are allowed to do.

There are important developments beyond the race to build larger models. Work around understanding the human genome and mutations across genes, including what has been described as the Genome Atlas, could have significant implications for healthcare. At the same time, AI does not always have to be enormous to be useful. Small models can already perform specific tasks effectively. We use AI in everyday systems such as voice assistants, navigation and logistics. The important question is increasingly how safely and effectively we apply AI, rather than simply how large we can make it.

The problem existed even before generative AI. Fake messages circulated on WhatsApp and people believed them because they aligned with what they already believed. AI has made the problem more powerful by making convincing audio, video and images easier to create. People need to develop a habit of verification. If a news story sounds too good to be true, it probably needs to be verified. Industry should provide clear labels or watermarks for AI-generated and AI-modified content. Malicious actors may remove them, so important information should always be independently verified.

Certainly, but the same technology can strengthen cyber defence. It’s becoming an arms race between attackers and defenders. AI can help automate attacks and identify vulnerabilities, but it can also help organisations detect unusual behaviour and respond faster. The advantage will partly depend on resources and expertise. India already has strong academic capabilities in cyber security, including at institutions such as IIT Madras and IIT Kanpur. That capability needs to keep developing.

In many ways, yes. Building AI capability is easier than building reliable safeguards around it. The consequences of a mistake are not always immediately visible, making safety engineering particularly difficult.

Governments across the world are still developing appropriate norms and regulations. Larger AI companies also have a responsibility to self-police, disclose significant incidents and be honest about the limitations of their systems. Voluntary commitments can help, but responsible industry behaviour and public safeguards are both necessary.

A job is usually a bundle of skills and tasks. AI may replace some tasks while leaving other parts of the job intact. For example, in a BPO role involving OTP verification, customer interaction and routine information processing, some activities can become autonomous while others remain assisted or supervised. So every job is likely to change. Some entry-level roles could be significantly affected, but AI will also create new roles, including forward-deployed engineers and data annotators. I would not give a precise number for jobs AI will create or eliminate because the technology is changing too quickly. We should instead prepare people for changing skill requirements.

India should continue frontier research because we need technical depth, expertise and capability for strategic and national-security applications. But there is also an enormous opportunity in AI diffusion — taking frontier AI to people and enterprises at the right price and making it work reliably in real-world conditions. The last mile has not been completely solved. Some companies have even rolled back AI pipelines because of errors or costs. India can build robust applications that solve practical problems at an affordable price. That is a major opportunity.

We need to develop and deploy useful AI ourselves rather than waiting for solutions from elsewhere. That requires investment in talent, research, applications and the wider ecosystem. We should pursue frontier research while simultaneously focusing on diffusion — taking AI into Indian enterprises, public services and everyday applications at the right cost. Combining deep technical capability with large-scale deployment can help India become a creator rather than merely a consumer.

India is doing excellent work in multilingual AI. IIT Madras, Sarvam, AI4Bharat and IIT Hyderabad are contributing significantly, and some of the work is world-class. But language alone may not remain a permanent moat. Global companies are also investing in Indian-language data and models. Our broader advantage is diversity — social, cultural and economic. India has highly developed environments alongside resource-constrained regions, many languages and very different deployment conditions. If we learn to build AI that works across that diversity and under resource constraints, that could become a much broader global advantage.

Tamil Nadu should not wait for AI solutions to arrive from elsewhere. Healthcare, manufacturing, automobiles, GCCs and fintech all offer opportunities to develop indigenous AI applications. The State has data, domain expertise and people who understand these industries. We need to bring AI talent together with that domain knowledge and invest in developing solutions centred here. The opportunity is not merely to use AI but to build products and capabilities that can be taken to other markets.

Technology can provide solutions, but implementation is the bigger challenge. IIT-Madras has expertise in urban water management and water treatment, including wastewater reclamation. Experts have also worked on flooding-related problems and interacted with cities facing similar challenges.

But having a solution and implementing it on the ground are different things. If storm water drains are blocked by construction debris, technology alone cannot prevent flooding. Traffic is similar. AI can improve routing, signals and traffic management, but there are physical limits to road capacity. Public transport and infrastructure also have to be strengthened. Not every urban problem is purely technological.

We have brilliant students, researchers and entrepreneurs with excellent ideas, but there is often a gap between having an idea and turning it into a product. A student from a Tier-II college may understand an agricultural problem extremely well but may not know how to build a product, find customers or provide technical support. AI Studio aims to develop that product mindset. The other major challenge is patient capital. Deep-tech companies may need 2-3 years before their potential becomes clear. India needs more investors willing to take that long-term risk. With the backing of IIT Madras, AI Studio can improve the chances of promising ideas becoming viable companies.