Thread by @medianama

LIVE | MediaNama is discussing the impact of the Personal Data Protection Bill and the Non-Personal Data Framework on Data and Artificial Intelligence #DataPoliciesandAI

https://twitter.com/medianama/status/1354712024324870146

Current algorithms do push us to the limit of trade-offs. However, ideas such as differential privacy can help up in navigating some of those trade-offs - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI #NAMA

If you want to find the average compensation in a company. One way is to give the sum total and that doesn't allow to identify an individual. Federated learning is another example where your data stays with you, raw data doesn't go into the cloud - @RahulAPanicker, @WadhwaniAI

At the moment, we can't deploy homomorphic encryption practically - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI #NAMA

We are far away from the tech front being deployed in large scale public good applications - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

Should AI be regulated? I don't think it is the right way to look at it. We should be looking at regulating applications. Do we regulate Chemistry, or do we regulate the Pharma industry which is an application area? - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

The regulations in the financial sector will have to be different in health. It can't all be under the same criteria...bringing in checks and balances is also important-- access control etc. - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

One of the common ways privacy is violated is by data breaches. And AI is usually used to prevent these breaches. There are algorithmic bases to mitigate some of these. We can even mitigate biases using certain techniques so that algorithms provide make more unbiased decisions...

...what standards do we hold algorithms to? @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

Do we need more data for better AI? Naive answer is yes, more data always helps. But the more sophisticated answer is that there are diminishing returns -- most interesting information comes from the initial samples compared to the later samples - @RahulAPanicker, @WadhwaniAI

Deep learning algorithms are extremely data hungry. A child doesn't need to be shown 1,000 cats to recognise a cat, so we know that algorithms today are data inefficient - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

We are in the early days of this new age of AI, and as tech progresses, we'll reach far more efficient processes - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

This discussion is being live-streamed. You can follow it here: pscp.tv/w/1OyKAEvYMnWKb

In Informed consent, the tricky bit is the "informed". How do you inform people of all potential issues? - @RahulAPanicker, @WadhwaniAI

In medical device regulation it's not just the hardware or software, but also the literature around the use of these things. All this comes from the perspective of risk mitigation. Some of the risk will come from algorithms, some from data, and some from outsiders-@RahulAPanicker

There are many domains where low risk applications are allowed to be governed under self regulations. What that self-regulation entails is also well defined. Self-regulations is not the same as no regulation and there is perhaps a large space...

within AI applications where there is a possibility of self-regulation - @RahulAPanicker, @WadhwaniAI

The amount of diversity in data is necessary so that the claims you make of the application can be held up. If I say I have an algorithm that is an automated radiology reading algorithm. Then I have to make sure that the test set for that has enough diversity - @RahulAPanicker

It is also on the developers to anticipate unexpected consequences of algorithms. Technologists bear responsibility - @RahulAPanicker, @WadhwaniAI

This discussion is being live-streamed. You can follow it here:
pscp.tv/w/1OyKAEvYMnWKb

We have to accept from history that we can not be fully ready for all adverse consequences of technology. - @RahulAPanicker, @WadhwaniAI #DataPoliciesandAI

LIVE | We have started with our first panel for the day, where we are discussing Data and Utility. Experts from the lending industry will speak on how they collect data, and use AI to generate insights about borrowers.

pscp.tv/w/1OyKAEvYMnWKb

Nothing in the public domain is personal information. If we have to help out customers and use data that is not in the public domain, then we follow a consent mechanism. What we do offer tech to help lenders in assessing a loan application - @meghnaskumar, @crediwatch

When we collect information not in the public domain, then we only offer our clients the technology to collect it. We don't collect it ourselves. We also give them the tech to purge this data - @meghnaskumar, @crediwatch #DATAPoliciesandAI

I don't think AI per say can be regulated. Use cases need to be regulated. Can you take data of a person to use for an illegitimate purpose? The answer is no. We need to come out with a privacy law. After that look at use cases of AI...

...and then if use cases are illegal, that's how you regulate those use cases - @meghnaskumar, @crediwatch #DataPoliciesandAI

Bloomberg takes a lot of data from the public market and creates a lot of insights using those, and they have ownership over that. - @meghnaskumar, @crediwatch #DataPoliciesandAI

This discussion is being live-streamed. You can follow it here:
pscp.tv/w/1OyKAEvYMnWKb

At the end of the day, what lenders really want to understand is the ability of people to pay returns, and their willingness to do so - Balakrishnan Narayanan, @Early_Salary #DatapoliciesandAI

Personal data and NPD is situational and context based. If I went to a doctor, I might have to revel what I ate, who I met etc. If I go to a banker, he might need much more information...

...However, the problem is when that doctor and the banker start sharing information - Balakrishnan Narayanan, @Early_Salary

We are also looking at the impact that a person's "maturity level" and greying of hair might have on a person's delinquency - Balakrishnan Narayanan, @Early_Salary pscp.tv/w/1OyKAEvYMnWKb

We ask customers to submit a selfie to asses their intention to pay a loan back - Balakrishnan Narayanan, @Early_Salary

But how do you asses that by looking at a selfie?: @nixxin

Narayanan: This project is currently under pilot, we haven't launched it yet.

As a lender, image of KYC is a part of the loan giving process. We do have images of customers and use it to understand the system on our own level - Balakrishnan Narayanan, @Early_Salary #DataPoliciesandAI

You need to have some data to test your models, but I think that should happen under a regulatory sandbox. - @Lyciast, @tattlemade, #DataPoliciesandAI
pscp.tv/w/1OyKAEvYMnWKb

The spike in also and credit scoring, should make us think a little about RBI's lending guidelines for NBFCs. We also need to ask why the PDP Bill makes an exemption for credit scoring - @Lyciast, @tattlemade #DataPoliciesandAI

For a class of algorithms, the way current systems are set up, do need a lot of data. But the kind of algorithms that the credit scoring industry really needs, there isn't a lot of need of massive data collection - @Lyciast, @tattlemade #DataPoliciesandAI

I don't know how much value non personal data even adds to the credit scoring function - @Lyciast, @tattlemade
#DataPoliciesandAI

In 2018, the Supreme Court said that we have a right to privacy even in public spaces. Personal data doesn't cease to be personal data just because it is placed in the public domain. - @VidushiMarda, #DataPoliciesandAI

The problem is that we're building tech based on extremely problematic science. And instead of thinking about foundational issues with that, we end up talking about optimising operational processes. - @VidushiMarda, #DataPoliciesandAI

There is an assumption that you can have robust self-regulation norms and that should be enough of protection. I'm sceptical of that. Self regulation contemplates people in power, deciding how they would act. - - @VidushiMarda, #DataPoliciesandAI pscp.tv/w/1OyKAEvYMnWKb

If you have 20,000 CCTVs on the road, and the talk about data protection etc., I think by that time, we're already late to the party. Design and deployment of these tools is as important as operationalise itself - @VidushiMarda, #DataPoliciesandAI

If you have to regulate a use case, then you can do that. If you look at the data, we have 60 million small biz in India and 85% of them don't have access to credit because there isn't enough data available with the lender on them. - @meghnaskumar, @crediwatch

The finance industry has always been a data driven industry. Now we're just replacing that with tech-driven solutions to collect that data. The fintech industry exists today to increase financial inclusion - @meghnaskumar, @crediwatch #DataPoliciesandAI

The pushback is because people are asking to look at the costs. What if to give loans to more people, we are targeting a certain segment of the society? You can disproportionately benefit a certain section. You can give more loans to men for instance - @Lyciast, @tattlemade

Alternate lending does concentrate power. You're rejecting or accepting people's loan applications on the basis of something that they can never contest with you on -
@VidushiMarda
#DataPoliciesandAI pscp.tv/w/1OyKAEvYMnWKb

How do you check your system for bias? While we create the model, we expose it to more than 3000 variables. At some point in time, more variables might not be the right way to look at - Balakrishnan Narayanan, @Early_Salary

When there's consent, I don't think we should be debating what is right or what is not. We live in a world where FB and Google are making billions off of people's data. In fintech, people will benefit from this data - @meghnaskumar, @crediwatch pscp.tv/w/1OyKAEvYMnWKb

With that, we now move on to our second panel for the day, where we will focus on the privacy implications of using AI in the context of the Personal Data Protection Bill, and the Non Personal Data Framework.

pscp.tv/w/1OyKAEvYMnWKb

The vast majority of people have absolutely no idea of what is happening to their data. Zero. The reason is that the way data is being collected for advertising purpose is so complex. This brings me to the more frustrating part of GDPR - a massive enforcement gap. @F_Kaltheuner

In a world that's made of data, data protection authorities are left to regulate everything which really exceeds their capacity - @F_Kaltheuner #DataPoliciesandAI

GDPR hasn't changed the dominance of big tech companies. - @F_Kaltheuner #DataPoliciesandAI

From a policy perspective, it does not make sense of AI to talk about AI in abstract. That's because if we say is AI good or bad for privacy? I'd say it depends. Use cases range from inherently dangerous to absolutely wonderful things for science...

...We need to push governments and companies towards the more ethical side of the debate. We also need to talk about market dominance since the biggest AI companies in the world today are the biggest tech companies - @F_Kaltheuner #DataPoliciesandAI
pscp.tv/w/1OyKAEvYMnWKb

We need to find ways to protect people from inaccurate inferences about them. If you incorrectly put someone in a group of 'reckless' people who shouldn't get credit, that has real world consequences - @F_Kaltheuner #DataPoliciesandAI

Lot of companies in the credit industry who say they use AI, actually don't use AI. In India, 90% of lending happens via PSUs or NBFCs. These companies manually look at the files before giving a loan...

...Loans given by fintech companies in India is less than 2% - @StatsAndPushups, @creditvidya #DataPoliciesandAI

From a credit perspective, we are far behind in terms of what AI can achieve. Our algorithms are so premature that they don't men take transactional data into account. - @StatsAndPushups, @creditvidya #DataPoliciesandAI

There are attributes and there are derived attributes. You go to a bank, and show them a bank statement. Then a person would check that to see how financially disciplined you are (number of cheque bounces) before giving a loan...

By using algorithms, what is the average daily balance of this person over the last 30 days. A machine can instantly make that calculation. Does this person look like a guy who could say his loan back? These are derived attributes...

All these attributes go into a scoring model and it'll decide the weightage of a particular attribute over the probability of a person defaulting on their loan - @StatsAndPushups, @creditvidya #DataPoliciesandAI pscp.tv/w/1OyKAEvYMnWKb

A lot of labelling happens in the advertising industry, not so much in the lending industry - @StatsAndPushups, @creditvidya #DataPoliciesandAI

The actual use of AI in India is not some super intelligent computers working. To a large extent it's being used for content moderation on platforms. In areas like facial, emotional recognition it is increasing because these tools are getting cheaper - @divijualsuspect

There are 3 main things impacted by AI tech:

1. Idea of AI's capacity to process, capture and store information
2. The implication of data driven decision on personal autonomy
3. Inferences and group privacy - @divijualsuspect #DataPoliciesandAI

I don't think we should think of regulating AI in general. It really depends on what we think AI does and regulate it from that perspective. - @divijualsuspect #DataPoliciesandAI

With AI, it's simply too risky given the types of harm. We need ex ante regulation - a law that speaks to AI in general before even a use case has come about. - @BasuArindrajit, @cis_india #DataPoliciesandAI pscp.tv/w/1OyKAEvYMnWKb

I would start with the use case of AI because every single use case has different implications on the rights of a person. Usually, there is a huge disproportionality in power differentials between impacted party and the end user. - @BasuArindrajit,

It is a tragedy that as a nation that's trying to get adequacy status, and as a nation that's trying to get some soft power, we are allowing state agencies to have carte blanche access to data. @BasuArindrajit on Clause 35 of PDP Bill. pscp.tv/w/1OyKAEvYMnWKb

In the Indian lending system, there are more biases in humans compared to machines - @StatsAndPushups, @creditvidya

I think a machine is more consistent, and not necessarily more fair - @Lyciast, @tattlemade #DataPoliciesandAI pscp.tv/w/1OyKAEvYMnWKb

If you see CCTNs, that relies upon records that were created during colonial times - how people are rowdies, or history sheeters because they belong to a certain tribe. There are a number of biases that are encoded in these databases - @divijualsuspect #DataPoliciesandAI

When we talk about emotional recognition, we say that it isn't really science. However, at some point it will become exact science someday. Do we even want to use that as an argument as of now? asks Lahar Appiah #DataPoliciesandAI

And that’s a wrap! We’d like to thank Flipkart, Microsoft & Facebook for their support for this discussion. We’d also like to thank our community partners The Centre for Internet and Society for their support. You can watch the discussion here: