• Home  
  • Technology & AI  
  • India Must Shape How Artificial Intelligence Is Built And Governed: Snigdha Bhardwaj | Full Transcript
AI

India Must Shape How Artificial Intelligence Is Built And Governed: Snigdha Bhardwaj | Full Transcript

Google’s Director and Global Head of Generative AI, Trust & Safety, Snigdha Bhardwaj, on enabling wider and safer access to AI in India

Text size

India’s opportunity in artificial intelligence lies not just in adopting the technology but in helping shape how it works for its diverse population. In this conversation with Nikhil Chawla, Snigdha Bhardwaj, Director and Global Head of Generative AI, Trust & Safety at Google, explains why building trustworthy AI requires systems that understand local languages and contexts, safeguards that keep pace with evolving risks, and greater participation from Indian institutions. She argues that AI must empower people to access information and make better decisions without replacing human judgement and sees India’s diversity as an opportunity to contribute to the global development of AI.

Nikhil Chawla: Hello and welcome to One Big Podcast. I am Nikhil Chawla. For years, Google Search helped us find information by pointing us towards websites and sources. Now, AI can bring that information together into an answer, hold a conversation and increasingly help people take action. That raises important questions about accuracy, sources and the responsibility a platform carries when it presents an answer directly. Today, we’ll discuss what it takes to build trust in AI search, how these systems can serve people across India’s languages, and what their growing role could mean for online safety and the wider web. Joining me today is Snighda Bhardwaj, Director of Search and Generative AI Trust Strategy at Google. Snighda, welcome to One Big Podcast.

Snighda Bhardwaj: Thank you.

NC: Before we start, I want to ask you something. When was the last time you saw something on Google and said, ‘I’m not really sure about this. Let me fact-check it’?

SB: I do that all the time. I think being in trust and safety makes that inherent in what you do every day, whether you’re thinking from the user’s perspective or for yourself. Yesterday, I was having a conversation with my son, checking which was the most polluted country or city in the world at that point in time. It wasn’t Delhi, so you don’t have to feel ashamed about it. It was Singapore because, at this point in time, Indonesia is clearing its fields and burning the material there.

NC: So, it’s not just a Delhi problem. It’s an international problem.

SB: It is an international problem, and it seems that it’s going to continue like that for Singapore for the next two months. So, yes, you always have to verify the information you get.

NC: I completely agree with that. Your role covers trust in search and generative AI. What does that mean for an average user like me? What does your team actually do to make sure that happens, and what are the most important decisions you take?

SB: To explain it in layman’s terms, you, as a user, should be able to trust the information. It should be relevant, helpful and safe when you encounter a response on Search. On a day-to-day basis, my team does that for all of you.  When we build products, we build them for the world. There is a global reach we need to think about, and there is a local contextualisation element. We want to make sure that every user feels this product is for them.

NC: Of course.

SB: How can users act on the responses they get from the internet? Do those responses provide the fresh context they need to make a decision? That is really important and central to my team’s work.

NC: So, how many people are there in your team? Because they’re doing this all the time.

SB: Actually, you’d be surprised. There aren’t that many people. There isn’t a one-to-X mapping, honestly, because we are now using more and more AI to advance our mission and carry out content moderation, enforcement and pattern identification at scale.

NC: How did you end up in trust and safety? What made you decide, ‘This is what I want to do’?

SB: It wasn’t like that at all. I landed in trust and safety by chance, but I stayed by choice. I have a consulting background, and I was looking to settle into a new area. I always found Google fascinating as a company. It’s a technology leader. When I got this opportunity, trust and safety seemed like a new area, and I was curious to explore it. That’s how I landed in trust and safety. It’s now been 11 years, and I’ve never felt bored with what I do every day.

NC: As you mentioned, it’s been more than a decade. What has changed over the last 10 years? Of course, so much generative AI and AI has come into the picture. Could you tell me about that journey and how it has changed for you?

SB: Specifically in relation to AI, Google has been in this space for more than 20 years — two decades, as we generally say. The way our users consume our products and features has evolved over time. Gen Z and Gen Alpha today want information at their fingertips. They want it fast. They want to ensure they’re getting the information they need to make a decision and achieve an outcome as easily as possible. The helpfulness of that information has to increase. From blue links on Search to structured content through Knowledge Panels or the one-box results at the top, we are now in the era of synthesising information.

NC: That’s true.

SB: That also directly relates to our founders’ vision statement: our work is to organise the world’s information and make it universally accessible and useful. I think generative AI is the next step in that mission.

NC: I agree. You mentioned blue links. Not too long ago, we used to get 10 or 15 blue links when we searched for something. Now, we get answers directly from generative AI. What is the real difference between the two?

SB: One thing I’d like all our users to understand is that AI answers are not a separate product. They’re built on the foundation of the work we’ve been doing on Search for years. The most authoritative and fresh content from those blue links makes it into AI answers. AI answers are our way of encapsulating information in a bite-sized format for our users, but they still link to the blue links. You can verify the information and check it. If you want to continue your journey, you have the option of checking each source and seeing how the answer has been created and formulated through those links. It’s about giving users the information they’re looking for in a way they can digest better.

NC: What does that shift mean for general users like me? Is it better, or can I switch off the generative AI results and just get 10 links? A lot of people don’t want those AI results.

SB: Users still have a choice about what they want to do with our generative AI features. I won’t go into the user controls that exist on our platform right now, but our philosophy behind making that content available reflects how the ecosystem is evolving. Users don’t want to spend a lot of time clicking through multiple links. To me, it’s a way of providing users with what they’re looking for. The way user queries and contexts are changing is also very relevant to how we’ve been adapting our products and the features we’re launching. Over time, query lengths have moved from keywords to more complex queries in which users provide contextual information.

NC: And now they can speak instead of typing it out.

SB: Exactly. If you want to build a solar roof, users aren’t just asking, ‘How do I build a solar roof?’ They’re saying, ‘I have a house that’s this wide and this long. I have only one roof. How can I make it more sustainable, and what options do I have?’ The complexity of the query is increasing, and I don’t think blue links will do justice to providing contextual information for those queries.

NC: Excellent. I want to understand how you reduce incorrect and misleading answers, especially when it comes to health, money or something critical.

SB: Interesting question. When we think from a user’s perspective, why do they come to Search? Because they feel it’s a knowledge agent or an information agent for them. If a user asks, ‘When and where can I watch this movie?’getting that answer wrong carries a very different risk from getting health information wrong. When we think about the information we share with users and the responses our models provide, we are very risk-averse in sensitive categories. Those include health, finance, the integrity of information in an election scenario, and civics. These are the areas in which we thoroughly red-team, or conduct adversarial testing on, all our models and product areas before they are launched to users. Sometimes, you know what to look for and have benchmarks. At other times, you have to venture into the unknown unknowns of what you’re trying to find. We conduct simulation exercises with experts, both internal and external, and we do this in partnership with the wider ecosystem.

NC: Do you think AI can be 100% right at the moment, or will it take some more time?

SB: I can’t comment on the time. AI is not 100% right at the moment. Whenever you’re querying and getting information, irrespective of the platform or chatbot you’re using, you always need to use human judgement. Research is being carried out into generative AI models, their risks, limitations and constraints. Understanding where AI could go wrong is an active topic of discussion across the industry, including among key opinion formers and regulators. Please always pause and verify.

NC: I agree. I also have a series called AI Masterclass, and after every tool I discuss, I mention that you shouldn’t rely on it completely. You have to use your own brain and use it as a tool to help you, without being 100% reliant on it. Scammers are also using AI, and they’re getting cleverer by the day. How do you stop scams, harmful content and misuse?

SB: One thing I do want to say is that scammers, or rather bad actors, have always been on these technology platforms. This isn’t a new problem. AI has shifted the focus because, while bad actors are using AI to cause harm, defenders such as technology platforms are also using AI to catch them faster. AI is helping us improve our speed and scale. It’s also helping us uncover new signals we hadn’t identified before. How can you cluster information faster? How can you intervene quickly? How can you find patterns of systemic failure instead of simply applying a patch to an issue reactively? Our models are getting smarter day by day. How we use those models to benefit users and technology platforms, and mitigate some of these risks, is critical.

NC: The problem is that these platforms are getting smarter by the day, with new models launching every week or month, but scammers and bad actors also have access to them. How can you stay one step ahead when they’re using the same models? How do you combat that?

SB: One principle we always abide by is safety by design. My team does this: we act as bad actors and hijackers.

NC: White-collar hackers.

SB: Yes. We research both the white-hat and black-hat worlds to understand the techniques bad actors could use. These can range from content misinformation to privacy and security vulnerabilities. Jailbreaks are a good example. You need to prompt the model in a way that attempts to break it as a bad actor would, or attempt a prompt injection. How do you identify those vulnerabilities in the system? When you think about bad actors, you always want to be proactive and build your defences in a way that prevents them from finding loopholes.

NC: Always one step ahead.

SB: Yes. We also run external reward programmes and trusted tester programmes.

NC: The bounties.

SB: Exactly, the bug bounty programme. We also have priority flaggers. From an ecosystem perspective, technology platforms need to see where they can get signals, because you can’t do everything by yourself. You can do a lot of monitoring and preparation in terms of defences, but it takes what we usually call a whole-of-society approach. You need those partnerships to make the product better.

NC: Of course. India has millions of first-time internet users. Do they need special care, and how can Google’s safety measures help?

SB: I fully agree with that. I also feel very proud to be Indian because I believe India can lead the world in defining AI and the frameworks that govern it. There are a few areas that are particularly unique to India. Language, and the diversity in how people have conversations with AI, is very different here. That language diversity is key to understanding the context of a query and making sure we respond in the right way. We’ve been engaging in several partnerships and programmes. We’ve been working with MLCommons and IITs to develop benchmarks for what a good safety benchmark would look like for Hindi, for example. We have safety benchmarks for English, but we need to make this work well in India, so we have to adapt. As we train our models, we are also working with IISc Bengaluru and Bhashini on something called Project Vaani, which aims to help teach models the languages spoken across 773 districts in India. That’s the plan, and we’re making progress, but it’s a multi-year programme. While we have these partnerships, please keep the user feedback coming, because that’s something else we continue to monitor.

NC: You mentioned languages specifically. Does AI work equally well in Hindi, Marathi, Tamil and other Indian languages? What are the trust challenges when people converse in their own regional languages?

SB: I cannot confidently say there is parity. In fact, I’d say there isn’t enough parity across languages. That’s also because the internet, and the way people have been using it in India, is new. Earlier, people used English to interact with the internet, and then they moved to Hinglish. Now, if you ask a query in North India, you can see different tabs, such as a Hindi tab and an English tab. As things evolve in the generative AI world, we want models to understand colloquial expressions and slang. When we look beyond the adult population to younger users, we want models to understand Gen Z or Gen Alpha lingo. Sometimes, that differs across languages in India. That’s our main challenge at the moment. How do we make sure AI is responsive and feels like your product, rather than a product based in the US that’s simply adapting or translating information for you?

NC: How do you know what the source is? If I search for something and generative AI replies, or if I’m having a conversation with it, do you mention the source explicitly?

SB: We do, and we do it in different formats. If you go to Search or ask a query in AI Mode or AI Overviews, you see explicit links with each statement. Those indicate that a particular snippet comes from a particular source. In the Gemini app, there’s a way to double-check the information so you can go deeper. It will take you to Search and also fact-check the information for you. If you’re retrieving an image through AI Mode, AI Overviews or Search, you can click on the three dots and check ‘About this result’ or ‘About this image’. That will show you the source and its background. There are different tools and techniques available. Another distinction users are particularly keen to understand is whether something is AI-generated. Our imperceptible watermarking technology, SynthID, created by Google DeepMind a couple of years ago, is very useful here. We’ve expanded our partnerships. We started with two partners last year, NVIDIA being one of them. This year, we’ve expanded to three more partners, with OpenAI becoming—

NC: Is the label enough? I’m sorry to say this, but a lot of people don’t realise that content is AI-generated. I’ve especially noticed this with content creation, such as things you see on YouTube Shorts or Instagram Reels. A lot of people think it’s a real person, and they don’t recognise the symbol or know what it means. Do you think that symbol, logo or watermark is enough?

SB: No, I don’t think so. Whenever you have a fact in front of you, you should be curious about where it’s coming from. From Google’s perspective, when we launched our ads impersonation policy, we said that if you’re advertising and your advert is AI-generated, you have to declare it explicitly. Some of the reels I see also say, ‘Generated by AI.’ But these standards aren’t fully adopted by every platform today, so you need to verify whatever you see on the internet. SynthID functionality is now available in Gemini or Nano Banana. You can upload an image and check whether it is AI-generated. But that’s only one aspect of it, because very few industry leaders have adopted the technology at this point. SynthID will work for a first-party image generated by Google or OpenAI at this point, but that isn’t true for every other image you find on the internet. That’s an active problem. Everybody needs to be curious and understand that critical thinking and human judgement are required.

NC: Snighda, we talked about how these models are updated every week or 10 days. How do you make sure everything is done before launch and that a model is safe to launch? How do you identify problems beforehand?

SB: First of all, we’re at a landmark moment. We launched Gemini 4 yesterday. So, it’s a argon era, as we call it now. From a model perspective, we have very high standards for the safety benchmarks we want to meet before launch. One of those is child safety. We have zero tolerance for any child safety violation in the model. If we see the model behaving in a particular way and showing those signs, that’s an absolute no. There are also different harms, which we classify as egregious and non-egregious. The risk tolerance for a model getting something wrong about entertainment tickets is very different from that for providing incorrect information when someone is looking up health symptoms. We have to meet those benchmarks. We test and retest the model until it meets them. We work with engineering teams who sometimes fix things in real time and sometimes continue to adapt different packages on the model to make it safer and more helpful. Apart from the safety benchmarks, we also test for unknown unknowns. We carry out adversarial attacks and identify new issues that could make users feel the model isn’t trustworthy. One important point is that our work doesn’t stop with the model launch. For example, the new model launched yesterday, but the work isn’t done and dusted. We continue with what we call post-launch monitoring. These are generative models, and they adapt very quickly and change their behaviour.

NC: If something goes wrong, how quickly can you respond?

SB: At times, it’s a matter of hours. In most cases — 99% of them — it’s a matter of hours. We call that a reactive escalation fix. But sometimes, if you really want to make a model safer, training it can take a few weeks. We want to make sure we’re not just applying patches. We need to identify patterns and make sure the protections are durable and robust, so they don’t have to change with every new model release. We need to monitor continuously. Users also provide very good signals through feedback when they think responses aren’t going the right way or find something harmful. We act on those signals as well.

NC: You’re a parent yourself. What is the one thing you want to make sure your child doesn’t get access to in terms of AI? How would you make it safer for them?

SB: I have a son who turned 13 this May.

NC: Welcome to the teenage years.

SB: Yes, of course. When we have these conversations, the way children explore AI is very different from how we explored the internet in our time. Access to information is now very broad. A 13- or 14-year-old can hardly imagine what it was like for us. They think I’m from a prehistoric era or the age of dinosaurs: ‘Oh my God, you didn’t have a mobile phone while you were growing up?’ It’s a basic need for them now. There are several ways in which we’re catering to youth users. We have strict guidelines on what we show users under 18 and the tone we use. Specifically, I want to talk about the persona element. We want to make sure our AI models don’t give the impression of sentience or form human-like relationships while students or young people are having conversations with them. AI models should provide information, increase curiosity, enable critical thinking and guide students through their projects. But students shouldn’t simply become passive recipients of the knowledge coming through these tools. We have strict guidelines that distinguish how we build these products for young people. We also don’t want them to avoid using AI simply because they want to stay safe. We want parents, students and young people to feel comfortable with AI, and we want school administrators to come along with us on this journey.

NC: What happens to my information? If I ask Gemini something, where does that information go? How private and safe is it? When we were typing things, we knew what we were entering. But once you get into a conversation, you keep talking. I’ve seen people talking for hours. My own friends talk and discuss things at night. Where does that data go? How safe and secure is it?

SB: There are several types of data we don’t even store. For example, if you upload your own image, we don’t store that data at all. We give users controls and privacy settings so they can choose how personal they want AI to be. With personal context — the personalisation we’ve launched in Gemini and AI Mode — we want to help users along their journey, but they have a choice about whether to see personalised results. Even on Google Search today, if you go down the page, there’s an option to select non-personalised search results. It forgets the search history you would otherwise have.

NC: Incognito.

SB: Exactly. We want users to have control over their information. Their conversations are secure by default, and they’re between the user and the chatbot they’re talking to. But we also want to intervene whenever a user’s conversation enters sensitive territory. We need to understand the query and the context of that conversation, because that’s an important part of how we build safety and security into the process.

NC: I think it’s about more than a child, or anyone else, accessing explicit content. I’m more worried about people thinking, ‘This is my psychologist, my psychiatrist,’ and making life decisions while talking to an AI agent. We’ve seen cases in which people were told they should end their lives. That’s the scary part. How are you working to make this safer, especially when it comes to health and mental health concerns?

SB: Mental health concerns are real. These are real risks we’re seeing in today’s world. Whether AI is involved or not, people sometimes feel isolated, depending on their lifestyles, and they may turn to these models and chatbots for companionship. We have strict guidelines on companionship for both children and adults. We do not want AI to be your companion. We do not want it to build a relationship with you. We want to make sure we break the conversation after a certain number of turns. We call that introducing friction into a user’s conversation or journey. That’s important because we want information agents to help you get the information you’re looking for without seeming like a partner. For children, as I said earlier, we have specific persona guidelines. The tone of the response matters. You cannot reinforce certain messages. If a child is looking for, ‘How can I be a thin pro?’, for example — I don’t know whether you know that term. I’m just learning this Gen Z language.

NC: Yes, ‘skibidi’ and so on.

SB: If someone wants to lose 10 kilograms in 10 days, we don’t want to respond to a query like that.

NC: What kind of response would a child or teenager get if they searched for that?

SB: First, we say that losing weight that quickly isn’t good for your health. Then, we try to intervene and steer the conversation towards, ‘What does sustainable health look like for you?’

NC: That’s everything.

SB: Exactly. Even when proposing solutions or recommendations, you always include a disclaimer to consult a health expert. This information or these recommendations shouldn’t be treated as expert advice. I’d say that for both health and finance: always consult a trusted practitioner who is an expert in the area, rather than relying on information available through AI.

NC: Other agents, however, are trying to be as personal as possible. If we talk about Moves and other platforms, they’re trying to make you feel as though you’re talking to a companion. Where do you draw the line? You want to sound helpful and as human as possible, but you don’t want to become someone the user thinks they should rely on. It’s like the film Her, in which he fell in love with an AI chatbot. We feared that day would come, and it has. Where is that fine line between being seen as a very personalised, good friend someone can talk to and get answers from, and recognising that things are going too far?

SB: There’s a difference between personalisation and being human in your interactions. Personalisation can still be objective. For example, if you’re interested in travel, search for travel information and have travel queries in your search history, we can personalise recommendations about destinations you might want to visit, ways to save money on flights, or hotels you might want to consider. To me, personalisation means being objective and factual, providing information, making the user’s journey more aware of their context, and helping them explore possibilities and opportunities. Human-like interaction or companionship is a completely different area of risk, and I think it’s more harmful. AI is there to help you make the right decision, but you need to use your judgement. You cannot get into an emotional trap in which AI becomes your boyfriend, girlfriend, companion or teacher. We use specific language in our responses, such as, ‘I am only a chatbot, so take my recommendations with a pinch of salt,’ without being preachy about it. Because you’re not prescribing a particular solution. You’re providing factual information as it is available on the internet. In sensitive areas, as I described earlier, we want to make sure we’re grounding it in data and linking to the right source.

NC: I understand. What is the one thing that every Indian should be aware of, or wary of, when it comes to generative AI? A lot of people in tier-three and tier-four cities are only now getting access to it and getting the hang of it. What should they know to stay safe?

SB: First, I feel AI has provided a huge opportunity. It helps put users in tier-one and tier-four cities on a par with one another. You might not have access to information through other means, and your phone may be your best source of information. In both places, you’re providing a foundation of standard information, because the response you give to someone in a tier-one city isn’t different from the response you give to someone in a tier-four city. That’s an opportunity, and it has bridged the gap quite a bit. That’s important to me. Another thing we’re seeing in India is that content may not be available in a particular language. You might still need to switch to English or Hindi because you can’t get the information in your local language. Always check, and try asking the question in different ways. I’m hopeful we’ll bridge this gap over time, but that’s the situation today. It’s also because of what we call data voids or content voids: we don’t have that content in so many different languages. The third thing that’s particularly relevant to Indian users is the concept of a shared device, where you’re asking queries alongside your parents.

NC: I was about to ask you the same thing when you were talking about children. How does Gemini, or any AI tool, know that a child is using a parent’s device when the child’s age hasn’t been entered?

SB: First, Gemini is only available to signed-in users, so you would have declared an age when signing in.

NC: Of course.

SB: I’m assuming you mean a situation in which it’s my device and my son is using Gemini on it. First, I’d tell parents to make sure they’re providing parental supervision. That’s very important if you’re letting your child use your phone. You should know what your child is using it for. If you have a different account or device and have created a child account, we offer parental supervision through Family Link. You need to work with your child as they navigate this new world of generative AI. To draw a parallel with the internet era, or even the time before the internet, think about teaching a child how to cross the road. You tell them what to be careful about and lay out detailed steps. But your child still has to use their judgement if you’re not with them. But you wouldn’t let your child cross the road alone if they were four, five or eight. You’d always be there. Think of yourself as a partner to your child as they explore. They’re curious about this world, and you can guide them through its challenges and risks.

NC: Students regularly use AI for homework, research and exam preparation. How do we use it as a useful learning tool while making sure children don’t become so dependent on it that they stop thinking for themselves?

SB: That’s an active point of discussion. How do we ensure AI helps children stay curious or improve their critical thinking skills, rather than making them dependent on it? My son and I have this conversation every day. The way they use AI in school is that, after doing some maths problems, they can use it to check their answers or see whether their logic was right. Our tools, and several of the features schools are adopting today, are based on Socratic learning models. When we develop these features, we consult pedagogy experts, clinicians and educators to create products that can help improve students’ critical thinking every day. Take Guided Learning in Gemini. India has the largest youth user base adopting Guided Learning, and it is based on Socratic principles. We want children to explore the different ways of using AI, but we don’t want them to become dependent on it.  There’s also a difference between infinite scrolling and using AI to learn a new skill, or having a conversation with AI about a topic nobody in your school understands. Some children are very advanced in robotics and coding, and not everybody in their class or school is at that level. They may feel isolated and wonder who they can talk to about a technique. I feel AI can provide helpful responses when children are curious about information.

NC: You mentioned children using AI in school and schools allowing them to do so. I wasn’t aware of that. At what age are students given access to AI in schools?

SB: It depends on the country. At the moment, we say our products are safe for users between 13 and 18. They can sign up for Gemini with their declared age and create accounts under parental supervision. We also have Workspace or education accounts through partnerships with schools and higher education institutions, and they’re using our AI tools too.

NC: That was about today. Let’s talk about India in 2047 and the Bharat we’re all aiming for. What does trusted AI look like for more than a billion Indians by then?

SB: Given the pace at which AI is evolving and the ecosystem is changing, I see India becoming very independent. I see people empowered to use AI to gather information, make their jobs easier, automate mundane tasks, or use AI as a supplementary resource or agent to help with their work. I see an empowered India in 2047 that is fearless and can use or adopt this technology to its advantage.

NC: Let’s see what the future holds. If your team had to give AI a warning, like those we see on cigarette packets, what would it look like? What would be the equivalent of ‘Drive safely’ or ‘Be safe’?

SB: The equivalent of ‘Drive safely’, as a recommendation, would be: ‘Think, check and then click.’ You need to pause. That’s really important. As a warning—

NC: Yes, the statutory warning.

SB: Considering how AI behaves today, I’d say: ‘Impressively confident. Occasionally wrong. Human judgement required.’ To me, that’s a critical part of how we’re evolving AI. We feel AI can be a good execution agent, but you need to provide the intelligence and use your judgement as you navigate the generative AI world.

NC: I agree. You look beautiful today. Did you ask Gemini, ‘What should I wear today?’ And especially that nail polish, with the gold polish on one of your toes — was that a Google recommendation?

SB: Not at all. That’s my own creation. Thank you very much for the compliment on the dress. I’ve worn an Indian outfit after a very long time.

NC: You look lovely.

SBj:Thank you.

NC: Great. My last question, which I ask each of my guests, is: what’s your one big dream for India? Your personal dream for India, and your dream for Google India or Google.

SB: Let me talk about India first.

NC: Sure.

SB: And not just in 2047 — I don’t want to wait until then. In the next few years, I want every Indian to be able to use this technology fearlessly. They should feel empowered. Young people should be exploring and becoming more curious about it. Parents should feel more in control and empowered by what technology offers. It should be a scam-free world in which you can trust the information you receive. That trustworthiness and fearlessness as we explore this technology is key. That’s my dream for India. I want India to be fully enabled to use generative AI to scale its impact in the world. I also feel India has an important role in co-creating the future of AI, simply because of our diversity. How can India create a playbook that isn’t narrowly focused, but is built for scale? That’s really important to me.

NC: I’ve already asked you the last question, but your answer made me think of something. India has had its UPI moment — one of the best things we’ve done over the last few years. Do you think we can still have our own LLM and have that UPI moment again: a model from India being used all over the world?

SB: Of course. Why not? We should allow different models to coexist. Depending on how you train a model, some can be very good at health, while others can be very good at coding. We need to create a world in which we’re not afraid of other people developing this technology. We should coexist in a world where we all feel good about users adopting it and making their lives simpler. To me, competition isn’t the goal. Coexistence, shared responsibility and the development of humanity should be our goal.

NC: Thank you so much. It was lovely talking to you.

SB: My pleasure. Thank you very much, and thank you for having me here.