- 5 hours ago
India Today AI Summit 2026: Lawmakers don’t care about our privacy, says expert
Category
🗞
NewsTranscript
00:00Good evening, everyone. This session is on AI privacy paradox, the legal challenges.
00:08And this is something that while we are talking about the benefits of AI, when we are talking
00:13about the great things that AI can do, this is the session where we are also going to
00:19talk about where the challenges lie, where the regulations come in, and what kind of
00:25complications we are seeing in this entire field. And my panelists are uniquely qualified
00:32really to speak on all of that. Apar Gupta, of course, from Internet Freedom Foundation,
00:36Nikhil Pava, Niharika Karanjawala, and Professor Ramakrishnan. Apar, I'm going to ask you first
00:43to weigh in on this issue. When we are talking about AI privacy paradox, to train any AI model,
00:52you have to let go of privacy protections in a way, because the large language models
00:59will not learn unless they have data. And that is where the question comes in. Where
01:05are our protections? Do we have adequate protections under Indian law? Apar.
01:10Thank you so much for having me here. I think the question of regulation around artificial intelligence
01:16intelligence is a large question, because artificial intelligence itself is a composite bundle of
01:21technologies. The word artificial intelligence can pull within itself what we use every day,
01:27more often than not, as users of artificial intelligence, a large language model, such as
01:31chat GPT. But it will also include within itself other forms of machine learning, which is deployed
01:37in public systems, Anisha. For instance, which are done in welfare, domains of welfare, or policing.
01:43Now, the question which you have asked me is that, do our existing regulations protect us
01:49sufficiently with respect to the private, intimate pieces of information which are digitally residing
01:57about us? And here, the DPDPA, which is a cronium for India's data protection law, is not
02:04in effect as present. And secondly, if you look at the large language models, to a large extent,
02:10they have been trained on the public repository of internet-based data, LLMs at least. So, according
02:18to me, the only protection to a large extent we have at present from these AI systems, which
02:26are getting better, the more information they pull in from the user as well, which may again
02:31be a part of the retraining of that data, is contract-based. So, I would say that there is
02:39an absence and even the Data Protection Act says that information which is publicly available
02:44does not fall within its ambit.
02:47Nikhil, I am going to take this question to you next. We are talking about democratizing
02:52AI. We are talking about using AI models to solve the problems of the Global South. And
02:59then comes in what we are talking to Apar also about, regulations and protections, not just
03:07for an individual, say, image, video or IP, but also privacy. Where does the line really
03:15come in and is there a line in existence under the law?
03:19So, I will give you a small example and where the line does, where you actually have a choice.
03:26I went in for an AI blood test and a health check-up about a year ago. And the form
03:33that
03:34they had me, they wanted me to sign said that I am willing to give my data for them to
03:42use
03:42to train their AI. And the rationale was that, and someone, when I said I don't want to do
03:48that, they said, don't you want to give your data to serve humanity, right? Now, that's
03:53a difficult choice for anyone to get because you want to help the other person. But there's
03:58also a risk because they, if they collect your training data, that training data leaves.
04:02This is highly personal data. And that's the tension that we have in society today that,
04:07one, there is this demand for data because you want to train AI, because AI can actually
04:13help serve in terms of, for example, discovering new drugs, discovering new mechanisms for curing
04:19ailments. At the same time, it creates a risk because these models can also be used to harm
04:25you. There are many Chinese models that are available that can also, in the process of capturing
04:33your data, also transfer that data onto the CCP, right?
04:37So there are risks that are involved here. To my mind, it boils down to choice. It's
04:42about whether I want to participate in this or not. It shouldn't be forced. I shouldn't
04:47be forced to give my data, just like I shouldn't be forced to give my, give my Aadhaar for my
04:53child's admission in school, because that's very sensitive data as far as I'm concerned.
04:59An Aadhaar number spread around can be misused, but everyone's collecting it. With AI, that's
05:06where the problem lies, right? You need more data, but also once that data is processed and
05:11a language model is trained, it's impossible to undo it. Because it's trained on that already,
05:18it is tokenized, it is processed, and there is a risk that it might publish that data when
05:24someone asks that large language model. And we have the AI companies have run amok. They've
05:29scraped data from across the web, where personal data is there. And like Aadhaar said, our law
05:36actually says publicly available personal data is not protected. Now, where does that create a risk?
05:41There is a company called Clearview AI in the US, which basically took this data, scraped social
05:49media data, and used it to train facial recognition models. Today, if you look at news publications,
05:56a large amount of personal data is being scraped by bots, because publications don't have a choice.
06:02They want search, but they also don't want AI training from their own data. But Google is saying,
06:08you have to give me your data for AI training if you want search traffic. So really, it's almost
06:14as if we don't have a choice against AI anymore. And that is antithetical to the cult, to individual
06:21rights as we have them today. All right. And on that aspect, Niharika, I'll take this question to you
06:27now. When we're talking about the risks, when we're talking about accountability, really,
06:35where is that under the current legal regimes? We've seen cases, as Nikhil just mentioned,
06:43that there are companies that are scraping data that are making it public. We've got cases where,
06:48in fact, our own, my colleagues have had their images, their videos being used to create AI-based
06:59videos, which have gone viral, which completely creates the question of fake news as well.
07:05We've got cases coming in from various parts of the world where people's lives have really been
07:12changed because they've been given incorrect information. So legally, how do we control that?
07:18Where is the accountability? Niharika?
07:20Absolutely, Anisha. I think this hits at the heart of really determining how people use AI in their
07:28day-to-day life. Where does the liability lie? If AI gives me incorrect information, I take that
07:34information on, yeah, of course, I will be accountable. Will they be accountable in any way?
07:38Additionally, supposing I use a baby monitor at home, if that data is leaked and pictures of my baby
07:46are leaked, supposing they are used, God forbid, for nefarious purposes, where does the liability lie?
07:51Does the security company take it? Clearview, I believe, faced a hefty fine. But is it?
07:57Not here. Not here, yeah, exactly. Then ring doorbells. When their data was leaked and strangers were
08:04speaking to children through those doorbells, they faced a fine as well. But like Nikhil said,
08:08you can't unring a bell. Once it's done and once your data is out there, there's no getting it back,
08:14and there's no really undoing that damage, even if there is a theoretical notion of, oh, okay,
08:19the company was in some way held to task. But so, affixation of liability, I think, is really
08:28really of paramount concern in going forward with seeing how we deal with AI on a day-to-day basis.
08:34And I don't know if I have a full answer for you as to where the bug stops and where
08:39liability is,
08:41but I think that is something that the experts do have to determine in going forward.
08:45Okay. And Professor Dharmakrishnan, you are building a large learning model. You're building
08:51an entire system for using Indian languages. That's a brilliant thing that you're doing for
08:59Indian users. But then these questions come to you. When you're building a model like this,
09:05where is the liability? Where is the accountability? Where is the trust that,
09:10as Nikhil mentioned, publications? If something is published and your language model is scraping that
09:18publication, where is the intellectual property protection for that publishing house? Where
09:24is the privacy protection for the person whose data has been scraped? Professor Dharmakrishnan.
09:30Thank you. I'll try and answer this from a very practitioner perspective.
09:35So, at Bharajan, what we believe is sovereignty is about participation. Right? You made a very nice
09:43point, Nikhil. Certain decisions need to be made in the interest of the people whom we are trying to
09:50serve. But are they informed? Are they participants? And sovereignty for me is not just a tag to be used
09:57to get your model better, you know, hit rates. But it's really to ensure that people are part of IP
10:05production rather than just IP consumption. I mean, this is an opportunity for India to become IP
10:10producers. Now, the way we actually structured this at Bharajan, again, being my experience,
10:15we have a consortium of nine academic institutions. We have a Section 8 company, 60 smart engineers.
10:21One of the consortium members, we have IIT Kanpur, IIT Madras, IIT Hyderabad, IIIT Hyderabad and IIT
10:29Bombay is leading it. But I am Indore. The reason I bring up IIM Indore, Indian Institute of Management,
10:34Indore is they are actually doing this, participating. They are going to many other tier two cities,
10:38talking to libraries who have data. A lot of it is, you know, literature. And engaging in conversations,
10:45the way Nikhil talked about. And trying to bring to them, to the table, the value that LLMs can bring
10:53in not only helping bring their, the literature to the forefront, but also create cross-pollination
10:59across other literature, create new hypothesis, potentially new skills. And this is an excellent
11:04point our honourable PM made when I met him on 8th. He said, can new skills emerge by lateral
11:10understanding across text. Now, just give me a few more minutes. The way we incorporated this,
11:15as I said, there is a syntactic component, which is, you know, check for licenses, ensure compliance
11:20with copyright, extract license metadata, implement, allow list, deny list and automated alerts for
11:26restricted content. So, sir, that is my question there. That is precisely this entire session all
11:32about. When you are talking about the regulations saying that you must respect copyright, you must
11:39talk about licensing, you must talk about protection. Where is the regulation? How are we actually
11:46implementing that? Because as Apar said, the DPDP Act currently is not functional. We've got the new
11:53IT intermediary rules as well. There are certain concerns raised about that as well. Sorry, Nikhil,
11:58you were trying to make a point. Actually, I wanted to bring it back to privacy. One of the things
12:03that
12:03we don't realise is that we are in a world of dynamic pricing. That dynamic pricing happens when
12:08there is information asymmetry between what I know and what the system knows about me. So,
12:15I write a site called Reason.live and I have done an experiment, the thought experiment saying what
12:20happens when a human negotiates with an AI agent, an AI agent negotiates with a human and an AI
12:28agent negotiates with an AI agent. Because there is so much of personal information that is collected
12:33about you, it can sense where you are vulnerable. So, for example, think about an Uber being taken
12:41in the middle of the night, let's say one o'clock in the morning. If there is a woman and
12:48a man
12:48travelling to the same destination, what are the chances the woman will see a higher price than the man
12:55because she is in a more vulnerable position travelling in the middle of the night in a city
12:59like Delhi, right? We have seen Zepto, for example, use this for this kind of dynamic pricing mechanism.
13:06So, people who are using iPhones where they price products higher for iPhone users. So,
13:12please understand, there is, these are all recommendation engines. They can do dynamic pricing,
13:17pick prices and so therefore we have to ensure that the misuse of data doesn't happen because
13:24the more personal data they collect about you, the more they are able to predict your vulnerabilities.
13:30This can be used for economically advantaging the person you are buying, the entity that you are
13:36buying for. Because at scale, there are algorithms that are determining pricing today. That's one of the
13:43most important harms that can come out of lack of privacy protection in AI.
13:48And that is, so I'll take this question back to Apar then. Apar, we've got regulations that,
13:54say, EU has created which create risk categorizations. US has some sort of risk management as well.
14:04Does Indian law, has Indian law caught up to these questions? Have we created the protections involved?
14:11And where are we on creating these questions, these protections if we haven't gotten there yet?
14:16So, as a general purpose technology, artificial intelligence, I'll again go back to it,
14:21is primarily being associated with large language models. However, as a descriptor, it is much,
14:28much larger. My principal recommendation and most of my work is towards the public deployment of that
14:34technology, which means when the government does it. Why? Because when the government does it,
14:38it has a higher form of power attached to it because it's in its functioning. If the technology
14:45does not work, a person gets excluded from rations. If a machine learning algorithm is attached to it,
14:51or if it is the identification of a person in a crowd through surveillance, it leads to a pre-trial
14:57incarceration. So, what needs to be done? And this, I think, matches the pre-existing approaches to look
15:03at the existing laws and legal frameworks we have. If we have the Right to Information Act,
15:07which needs to be strengthened and needs to be made applicable to how tendering processes are done.
15:13How is the model made? What kind of audits are done? What are the failures? What are the risks
15:18which have been done? We need to strengthen the statutory foundations of what you already have.
15:23Now, coming to you, that part of your question, how are we approaching it and where I express some
15:29bits of my disappointment? If you look at the outcomes and the reports which have come from
15:35the India AI mission, the first one is primarily saying that you need to prioritize innovation which
15:41will happen through AI over any kind of regulation which will happen. Because regulation has been understood
15:48by it to be a barrier to any kind of assimilation, what they are these days calling diffusion of AI.
15:56Okay, that's the first thing. And the second is a document which has come from the principal scientific
16:02advisor's office is that we need to follow a techno-legal approach. As a lawyer, a techno-legal approach to
16:07me
16:08means it's not coded in law, it's coded in the software itself. And when it is coded in the software
16:13itself,
16:13without any oversight framework, it can be changed, it can be broken, it can not work or it can create
16:20new problems. Hence, what I would say is that we need to be a little much more honest. If a
16:26technology
16:26offers us great promise, if a technology offers us a revolutionary change in our world, then it has
16:34immense power. It has the power to provide and it has the power to deny. And when those denials occur,
16:40we need to have legal frameworks, we should not shy away from it. And principally, I'll circle back,
16:45it goes back to the government when you're deploying AI in a surveillance camera in Delhi,
16:50by the Delhi police, you need to tell us which of those cameras, what is the software running,
16:55how does it work, where is the data shared. And Niharika, to you then. And I've been asking this
17:02question, and we're looking at the various aspects and the various forms of risk really,
17:09what Nikhil and Apar have talked about, the various areas where we've seen violations.
17:14But I am asking a very simple question. There is a DPDP Act, there is an IT Act regulations,
17:21are they working? Or are they, where are the lacunae?
17:26So I think the law does fall short. And I think there does need to be perhaps, A, an overarching
17:35law and better enforcement as well. I think there needs to be both new legislation, better enforcement,
17:41and ultimately, I think it has to come down to extreme vigilance by the user. Like going back to
17:46what Nikhil said recently about the medical test he had, the blood test that you did. This comes
17:51down to whether you're reading the fine print or not. The majority of users, especially when you're
17:56interacting with something less serious, like a game on your mobile phone or something like that,
18:02or you're putting information into chat GPT and you're trying to make an animated version of your
18:08photographs or something like that, you're not really, or at least most users are not really reading
18:13all of the fine print involved and are not taking into account exactly the ways in which their data
18:19is going to be used. So I think not only does the law need to move forward and the law
18:25need to be far
18:27more protective of individual privacy and liberties than it is of innovation of technology, because I
18:35think it just comes down to a humans versus technology kind of thing in my mind, especially when
18:41Apar spoke about things like preventive detention, Apar spoke about things like the police using
18:47cameras, people using tracking. Even if you look at any information that's on the cloud, if you've put
18:54your, a couple of generations ago on your iPod or on your iPhone, you put your fingerprint in,
18:59Apple now has that and they've made it clear in their policies that if you are ever part of any
19:05criminal investigation, they are going to hand over whatever information is required. So they can just
19:10simply hand over any biometric data they have with them. So I think it is also up to the user
19:15to be
19:15extremely vigilant. I think it has to progress on a multitude of fields, legislative, enforcing,
19:23and vigilance of the user themselves. All right. Professor Radhakrishnan, you had something to say
19:29about the way the government or the companies look at these regulations.
19:35Yeah. So since we're building models from the first bite and lots of applications, I mean, all the points
19:41being made here are very valid. I talked, I alluded to the syntactic aspects of building these foundation
19:47models, the data part, right, education. But there's a semantic aspect, which is the intellectual property,
19:52right, right? Now, all we can do at this point is build as much provenance in the system as possible,
20:00right? Provenance in data. We work with lots of vendors for speech data, for example. And we built an
20:07agentic model so that the metadata is at least faithful to what is being recorded, right? If you
20:14don't even have the metadata, there's no way you'll actually be able to go back, right? So provenance.
20:17Now, the same provenance at the level of the model building, we call it observability. So having this
20:24observability through and through is very critical because that's when you can even take action. How
20:29do you implement policies if you don't have the observability? And therefore, the ML stack becomes
20:34important. Sovereignty is not about data. Sovereignty is about the models, the way you're training them,
20:39you need to have the trace, the logs, right, at every level. Now, the third component, just let me,
20:46complete this, is applications. And applications, especially in air-gapped environments,
20:52applications with healthcare. And one classic example, we worked very closely with Amrita.
20:57Amrita, EMR, EHR systems have been very widely adopted. And we, in fact, brought out a MedSum,
21:02basically a solution. Now, what about patient data privacy, consent-first design, medical accuracies,
21:09safeguards and hallucination control? All of this is possible in very specific collaboration. But this
21:16requires co-design. It cannot be a hands-off approach where you expect the ML designers to
21:22actually design it, being with all the right safeguards. The healthcare expert has to sit together.
21:28All right. I'm sorry. Nikhil, you have something to say?
21:32Yes, why not? Why shouldn't we expect the ML designers to start with safeguards first? I mean,
21:37look at the way they've approached it. They've scraped the web. They've scraped personal data.
21:42Without permission, that is both copyright violation and privacy violation. Are you saying
21:47that they are above the law? I'm not saying that. I mean, that's what you seem to be alluding,
21:51that we can't expect them to do this without violating the law like that. They have to have
21:59safeguards before they begin doing this. Because once a model is trained, you can't untrain a model.
22:05So, if you don't start with safeguards, they already have the data. I mean, all protections
22:12that have come in have come in probably around middle of 2023. Robots or text exemptions for
22:17copyright violation came in July 23 and August 23. Google and open AI. But BART was already alive by
22:25then. Chat CPT was already alive by then. They already violated copyright. They already scraped personal
22:30data. And no one's holding them to account now. In fact, in India, they're considering giving them
22:35an exemption under text in data mining, which means that they're legitimizing a theft that took place
22:41earlier. So, you know, the problem that the way we are approaching it is, we're seeing this as a
22:47fate of company. Privacy is violated. Now, we won't do this in the future. We'll have a technical
22:53safeguard built in right now. What about when you started? You ignored the law when you needed to
22:59build it. And now that it's done, you're saying, now we'll follow the law. What happens then? This is
23:04how companies very classically give incumbents the benefit of violating the law in advance and make it
23:13difficult for competition to come up. How does India react to this? India says, we will not have
23:18the protections because now we need our companies to catch up. So because the other companies violated
23:25the law, we now need to let our companies violate people's privacy and people's copyright
23:32rights. So instead of holding those guys to account, I'm sorry, I'm sorry, Professor
23:36Amakrishan, I'll ask, I'll ask a part to weigh in on this issue that we've already had a huge amount
23:43of privacy violation. We've already made the choice really to set aside the privacy concerns to allow
23:51these training of these models. So what happens now about? So technology is always a social archetype
24:00in the sense that we as people have the ability of determining what it is, how it works, but it
24:07almost and always seems inevitable. And you can see it through waves of technology, communication
24:12technologies, radio, television, even newspaper, right? Now, the central question around how AI models
24:20are built, and if you read the Transformers paper from which all of this started, is actually
24:31engineer's approach to how do you achieve a result without taking the law into account.
24:37It is just the fascination with what is being built. And policy and law is not a core part at
24:43that point
24:44in time. And the foundations of artificial intelligence, and there's a great book, the author is in India also,
24:52Karen Howe, Empire of AI, it's on OpenAI. She tells how OpenAI, I'm sorry, I'm talking about one company,
24:59I should not, but transforms from a non-profit to a for-profit corporation, right? So the initial
25:05justification at least was, and for instance, Anthropic, there's a great article on Anthropic,
25:11how Anthropic was smarter on the intellectual property, because it scanned 50,000 books. But
25:16imagine what it did after scanning those 50,000 books. It burped them. It's like Fahrenheit 451.
25:23So, I think so, because of the tremendous potential which is being seen, the geopolitical interest,
25:31the massive investments which are being made into artificial intelligence, the bet is, for most people
25:38who have the reins of power in our society, we are going to forsake intellectual property and privacy
25:44as lesser interests in favour of artificial intelligence, which goes back to technology being
25:51a social archetype and a choice being made in favour of it. This should not be so. Law should step
25:56in.
25:56I advocate for it, and that's my full-time job. But this is what I see is happening much more
26:01largely.
26:03Niharika, this last question goes to you. When we are talking about that we've accepted the fact
26:09that we set aside privacy protections on IP protections, and we are focusing at a policy level
26:17to train the AI to allow more companies to come up on AI models. When we are now creating these
26:25laws,
26:25when we are creating these new regulations, we are asking the, at least for a general user who's,
26:32say, using their phone to put their pictures on Instagram, at least for them we are building in
26:39protections that, well, you can't use it except for XYZ. Does that help? How far does that help?
26:45And for the users sitting here, what's your message?
26:48So I do think every little bit helps. And while I agree with Apar broadly that the majority of people
26:55will accept the privacy violations that do come with using technology simply because it makes life
27:05convenient and it integrates you into what is happening in the day-to-day and what the majority
27:12of the world is doing. So most people will accept it. But I still feel that there is a glimmer
27:16of hope
27:17in a pushback. Anisha, like you and I were discussing, just this past week, in fact, the company Ring
27:23that has e-doorbells, video doorbells, have had to cancel a partnership they had with a security firm.
27:31They launched, through a Super Bowl ad, they launched this new feature called Search Party, saying,
27:35we'll be able to track down your dog in case your dog gets lost. We'll use all the Ring doorbells
27:41in
27:41your neighborhood and we'll find your dog and we'll send you that message. And there was huge backlash
27:46that I don't think they anticipated because it was this notion of, well, everybody loves dogs,
27:50don't you want to find your dog if it's lost? And there was huge backlash saying, this means you're
27:55just watching 24-7 from my doorbell and if you're face-shill recognizing my dog, surely you're doing the
28:01same to me. And they've had to cancel that entire venture and break off with that security company.
28:06So I think there is still a glimmer of hope on the user backlash. And I totally agree that it
28:12is
28:12also up to the courts to intervene, to governments to intervene and to crack down. And I totally agree
28:20that you can't just say, oh, all right, in the name of innovation, a theft that took place 10 years
28:27ago
28:27should be all right now. If I stole your cell phone 10 years ago, it wouldn't be all right now,
28:32even if I did great things with it. All right. Sorry, Nikhil, I'll ask one quick question that
28:36and I need you to sum up this entire thing. When we're talking about… I need to respond to Nikhil.
28:43We can do that off the stage then. I resonate with Nikhil. The only point I wanted to make is,
28:49it's also our, the onus in sovereignty is on us to ensure that the policies are implemented from the
28:55first bite. I agree. But in response to you and Niharika, all I have to say is,
29:00I don't see NatGrid being cancelled any point in time soon, right? That's the, that's the
29:04surveillance system that we are building that is taking data from all public and private data
29:09sources without our consent. That's going to be used to surveil us as citizens. We don't,
29:15that's your, that's actually the biggest privacy nightmare when it comes to AI.
29:19Yes, that's not getting cancelled.
29:22So, when we've got issues like NatGrid access, we've got access to your user data,
29:28we've got access to your government data, we've got access to any and all data that the government is
29:37collating for you on NatGrid, on several other criminal law aspects. So, these are the issues,
29:43key issues that the government needs to be looking into. Nikhil, I'm giving you the last word on this.
29:48Where are we on this? And what, what should users be wary of?
29:54Look, I think everyone should be careful about where they share their data, how much data they shared.
29:59It's just about awareness because frankly, it's in our hands. It's the apps that we use. It's that our,
30:04every app is now built to collect more and more data because everyone's collecting data for AI.
30:08We need to be careful because no one else cares about our privacy, frankly. I don't think the lawmakers do.
30:16I don't think the companies that, whose apps we use do. So, it's upon us. I mean, there are also,
30:22on the other hand,
30:22there are people who are deploying open claw AI agents and giving access to the entire WhatsApp,
30:27email and multiple sources of their personal information, their relationship, their context.
30:33What they also don't realize is not only are they exposing their own personal data,
30:37they're also exposing the interactions with the people that they're interacting with
30:40to a bot that can be very easily prompt injected, hacked and their data exfiltrated.
30:46So, awareness is the only solution we have left today and it's only in our hands. No one's here to
30:51save us.
30:51All right. Thank you very much for that, panelists. We've basically rung a note of caution here in
30:59this particular session. As AI develops and the precautions that we need to take and the fact that
31:07the law is still trying to catch up in a certain way. Thank you very much, panelists.
Comments