EP 90: How To Tackle AI Privacy and Governance

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.

Data privacy and AI aren't always two things that go hand in hand. It's, it's something that I think so many businesses and individuals are, are worried about when it comes to using different generative AI tools is what happens with my data. We're going to tackle that today with an expert in AI privacy who I'm very excited to have on the show. So if you're joining us live, please take part in the conversation, ask some questions. I think it's going to be exciting, a very informative conversation. So my name is Jordan Wilson and this is Everyday AI. It's a daily live stream, podcast and free daily newsletter helping everyday people like me, like you, not just make sense of what's going on in the world of AI, but how we can actually make it work for us. All right, so before we get in our discussion on data privacy and AI governance, let's go ahead and go over what's happening in the world of AI news. So we actually have a couple of things related to privacy and governance. So starting off, Google just announced a new way that they are going to be watermarking AI images. So they obviously announced this a couple months back at their annual conference, but it was just released now, you know, within the last couple of hours, that DeepMind, which is essentially Google's AI specialty arm, so to speak. So this is a DeepMind product and it's called Synth ID and it will watermark and label images that have been created with AI. So it's going to be interesting to see how that one plays out. All right, next, the Microsoft president just said that AI needs human control to be safe. All right, so in an exclusive interview with cnbc, Microsoft president Brad Smith warned, and I quote, that AI has the potential to become both a tool and a weapon. So this is obviously something we've heard about before, but if you wanted to check out more on what Brad Smith said, make sure to check out that section in the newsletter. All right, and our third news story of the day, definitely it might be last but not the least. So Chat GPT Enterprise has been released. So this was announced months ago and we covered it on the show when it was announced. But OpenAI has released finally the enterprise version of ChatGPT, a much more locked down version, kind of geared more toward data privacy. So this is something that OpenAI has been working on for many months. I don't even know what all of these terms mean. It's, it's SOC 2 compliant. It's a certain level of compliancy with data, I believe. So a couple other things, details about this chat GPT enterprise. Not everything's been released, so at least right now there's no price tag on it yet. But we do know it will be faster with no caps or limits on GPT4 and longer context windows as well. So exciting news on the enterprise front because I know that so many companies are a little hesitant with their data and they don't necessarily want to hand it over to OpenAI or some of these larger companies. So what a great transition piece for our experts. For today we're going to be talking AI privacy and governance. So I'm extremely excited to bring on our guests for today. So please welcome to the show Katharina Corner from the Tech Diplomacy Network. Katharina, thank you for joining us.

Thank you for inviting me. Great to finish out.

Absolutely. This is going to be a fun one. This is such a topic that I think a lot of people are talking about. But Katharina, if you could just start us off very general, kind of break down this complex world of AI governance and privacy just kind of so the everyday person can understand what does AI governance and privacy even mean to the rest of us?

Thank you, thank you, thank you for that question. Super important question. So I thought I would start with mentioning that we have some globally accepted privacy principles and which are embedded in many, many privacy and data protection regulations around the world which are by the way, popping up here and everywhere like just getting more and more. And of course, all of those privacy principles also apply to AI systems in case they process personal data. So which are those? So first of all, we have data quality, for example, that means making sure the data used by AI systems is accurate and reliable. So for example, if you're teaching a robot to recognize different kinds of fruits, if it learns, if this data is, if data it learns from is full of mistakes, for example, calling apples oranges, that wouldn't be very helpful. So second one is data collection limitations. So data minimization is a very important principle. Always use just the data you really need to collect for a specific task and do not gather extra information. So for example, if you have a fitness app on your phone, if it's asking for a location, but it doesn't need that to count your steps, it's not following this principle, then we have very important principle of purpose specification. That means the data you collect and process as an organization should only be and only be used for the purpose it was collected for. So let's say you sign up for a gaming app. So the email you used for that gaming app to send you game related stuff should of course not be used to advertise other products. Then we have use limitation. This is meant in that way that you should not use data for anything else other than the original purpose without permission. You have photos on your social media platform and all of a sudden those photos are used for something else. That would be a breach of this principle. Then what is super important, and I think a lot of people are very aware of this, is transparency. This means being open about what data is collected, why it's collected, how it's used. Let's say we order something in a restaurant or we have a recipe, of course we want to know which ingredients are used. Maybe I'm allergic to something. So you trust that meal more if you know what was used. And then lastly, I want to mention accountability. If something goes wrong, there should be someone responsible. So those are some privacy principles that also apply to AI systems. And in of course, as expected, there are some particular issues on top of that when it comes to AI. So for example, a big issue is the origin of the data that large language models use the basis for Services Just as ChatGPT, as you mentioned here. Usually the data is crawled from the Internet and we have different regulations around the world. It's pretty complex, the world of data scraping, I mean the regulations about web scraping. So in, in the United States in general, the law says if information is freely available in the public, businesses can use it. So if you have, I don't know if you wrote something on a poster on a public bulletin board, anyone can read it, anyone can use it. Except for when the website says you can take the information or if it's behind a paywall or if it's behind your credentials, then you shouldn't. I mean you're not allowed to. But in Europe on the other hand, any information, if it's public or not, is considered personal information. It is protected by the gdpr. So if someone wants to use personal data, they need a good reason and a permission. Like you borrow a book from the library, if you don't ask, it's not allowed. So that's something that is very, very relevant in the context of AI and LLMs. And should I, do you want to? I mean, I could go on and on and on.

No, no, I mean, I think just right there we hit on so many different, different points that are so important. You know, making sure data is accurate and reliable, only using it for the correct purposes, which I think is super important. But also, you know, transparency and, and you know, understanding that different countries, you know, have different rules and regulations. I think it's important because I think, you know, even here on the show, a lot of times we're talking about how things impact the U.S. but you know, as, as you see here, you know, May Brit is joining us from, from Europe. You know, Val right here with a comment is, is joining us from South America. So we do have people tuning in from all over, you know, Bronwyn joining us from South Africa. So thank you all for joining and thank you for your comments. But one thing, Katharina, that I wanted to talk a little bit more about is kind of how, you know, companies or even individuals or just us as society, how do we strike a balance between leveraging AI's capabilities because we always hear about it and you know, all of the things that generative AI can do or make is so exciting. So how do we strike that balance between AI's capabilities and, and making sure data kind of remains how it should, which is, which is private or only the data that we are wanting a company to collect for the purposes. So how do we strike that balance?

Are you still running in circles trying to figure out how to actually grow.

Your business with AI?

Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on gen AI. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead. And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to use Gen AI. So whether you're looking for ChatGPT training for thousands or just need help building your front end AI strategy, you, you can partner with us too.

Just like some of the biggest companies in the world do.

Go to your everydayai.com partner to get in contact with our team or you can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on gen.

Well, that's a very, very good question to ask and I don't know if that's too goes too far from the main topic of your show, but I have done a lot of work and research in privacy enhancing technologies. So that's usually the answer I give. So a whole new field has emerged with a vibrant startup community. Google is using it, Microsoft is implementing it in their products. Something that is called privacy preserving machine learning and a huge research ecosystem as well. So because it's true that with traditional privacy protections, often utility of the data decreases, it might even decrease the utility of the data while not even protecting the data very well, which was the first intention. So one example is the protection of personal health information under hipaa. It's a classic example. So that's the US Health Insurance Portability Accountability Act. So there are a lot of entities covered under hipaa. So if you process personal health information, you have to comply with this law. It's a federal law in the United States and under hipaa, personal health information should be de identified for protecting patient privacy. So to support research, reducing risk promoting data sharing between hospitals so that we can have better insights in this personal health information. And one accepted method for this is so called, it's a safe harbor method and that involves that you remove 18 direct identifiers from the data like name, Social Security number, medical record numbers, et cetera. So that law was from, from the 90s and they really put this into law because usually law is tech neutral. So they formulate it in a phrase, it in a way that you know, it can go with time. But HIPAA really said strip those 18 identifiers from personal health information, then you're more or less good to go. But in fact you strip so many identifiers for the reason of privacy protection that it is that you lose a lot of information. Plus it is not even good privacy protection because there are many re identification attacks that can be very successfully conducted on this HIPAA protected health data. And so now here is this new ecosystem, these new technologies and approaches that have emerged over the last couple of years that can unlock the valuable insights from the data while protecting privacy. And those are just to mention a few of them. Differential privacy. So differential privacy is kind of, you randomize responses in the, in a data set. So you cannot say 100% is your data in this data set or not. You cannot, you can mathematically prove that it's not possible to tell if your data was in the data set and contributed to the output, yes or no. Then we have synthetic data. So you build a, you build a new data set that mimics the patterns from the original data set and then you can use it or there's something new Called, I think it's a very poetic name, homomorphic encryption. It's like it's on a maturity curve, it's going up, but it means that you encrypt data and you can still process and have like, you know, conduct computation on this encrypted data without decrypting it. Pretty similar is trusted execution environments. That's a hardware solution where you also work, you have like a secure enclave and the encrypted data goes in only inside it's decrypted and processed and computed upon and then it leaves the enclave in an encrypted way. Or we have secure multiparty computation. So meaning you can have a common computation and output but you will never know what actually what was the input. Actually a classic example is often you can conduct an analysis of incomes, what's the average income for example, or what's the disparity between female and male incomes. But you will not see what the actual input was. So that, that those are some of those privacy enhancing technologies and they are politically supported globally. So we have the US National Strategy to Advance Privacy Preserving Data Sharing Analytics. Earlier this year I wrote an article somewhere, maybe we can post it or whatever. Like there's so many, there's so much policy support for those pets because they're so promising. Because some of them, as I mentioned, they even protect data in use. So we do know that, you know, protection of data addressed, so securing data when it's stored, that's you know, state of the art when it's sent. So you protect data in transit. But some of those new technologies can also protect data while in use. And that's a pretty big thing. So we will get there and of course, like, yeah, please.

Wow.

No, I'm just, I'm just saying, like I am y', all, I, I talk about AI every single day and I don't know about you all listening and tuning in. I am getting a first class education right now from Katharina. Like I have so many notes and if you're like me and if you can't keep up with everything, you know, I'm trying to Google things, I'm writing things down. Don't worry, we are going to be sharing everything in the newsletter. But I do have, you know, one more big question I think Katharina, before maybe we can get to a comment or two. So if you do have a question, make sure to get it in now. But I want to talk about, you know, because you kind of talked about how data privacy and AI governance is so much different in different countries. But one thing that I think is really going to be on people's minds, especially here in the US, is when we talk about AI's impact on democracy. Right. Because I think especially here in the us, we've already seen it different political groups using AI in probably ways it was not intended to, to make it seem a political opponent maybe said something or did something they did not do. So I, I, I mean, can you just talk just quickly about kind of the, the risks of AI just on democracy and how it might influence, you know, public opinion, elections and even just how we understand democracy?

Yeah, I mean, that's a big concern for sure. And with the help of AI or machine learning, you for sure can aim or try to influence public opinion and elections. So first of all, with election campaigns, for example, you can target specific groups with tailored messages. I mean, that's nothing new. But you can identify, because you can analyze this vast amount of data in such a great way, you can identify potential swing voters or craft messages to really influence the decisions. Micro targeting. So that all leads to manipulation. Manipulation. Manipulation can also be achieved by, you know, AI generated deep fake videos. We had examples where like, people were already like, you know, videos were made as if they had made those videos and that was not those people. You have filter bubbles, more or less. So AI algorithms can personalize content, showing individual individuals information that only aligns with the existing beliefs. We have, of course, security risks, that's a big thing. So AI can be used by malicious actors to hack into election systems, manipulate voter registration data, disrupt the voting process itself and transparency and accountability. I think that's also a big topic also in this regard. So the use of AI in elections and public discourse raises the questions, who is responsible for the AI algorithms that influence public opinion? How can we ensure also in this context that AI is used ethically and responsibly in, responsibly in the political sphere. And this is why in general also with like, you know, I'm sure you talked about hallucinations by LLMs or services like ChatGPT, like wrong outputs. I think that education is really so important, I don't know exactly how that works, that AI education will find its way into, you know, our public education system system. But I recently started an initiative just here for the Bay Area. I would so any school in the Bay Area that wants me and explain AI to them, like to the, to the kids, I'm happy to do that because I think that's just so important. Just as we have like not every single book just because a book, it's a good book. Right. It's not always like write what it's in that book just because it's a book. And the same is true with actually anything that is on the Internet and any output by any system you can never 100% trust. You should always, you know, use your own, I don't know, common sense or prove it with different sources. So we have to raise this awareness that we're still self responsible for content that we get from the Internet and from really super cool applications like ChatGPT, etc.

Yeah, absolutely. And I couldn't agree more with you, Katharina, about the need for better AI education in the school systems. Absolutely. Because it seems like so many schools, I just had an episode on this recently. So many schools are, you know, shutting it down or not allowing, you know, students to use it when personally I think that's not a good idea. But okay, so we've covered so many things, but I do hope we can get, get a question in there too from our audience. So Nancy saying good morning, glad to be back on the live show. Great to have you, Nancy. Cecilia saying data is indeed one of the most important assets to manage right after our relationships and strategies. Couldn't agree more. So a question here if you wanted to take this one. So Ben saying thanks for the comments and he's asking is it realistic to expect AI models to be fully transparent and explainable? I think that's a great question. What would you respond to on that one, Katharina?

Yes, yes, yes, yes, very good question. So I think for some models or some applications, like white box applications, it is possible. So I mean the majority of AI is machine learning. So if we have a simple, I don't know, recommender system and the model is a decision tree. So I mean it's really decision tree like I don't know, you know, I don't know, apples or oranges. And then does it have a warm or not or whatever. So in the end you have a healthy apple or whatever. This is an explainable model. Even it already gets a little bit more complicated if you have a simple model that is still tricky like with a random forest. So you have hundreds of decision trees and then you, you know, and then you compare the outputs of those or compare them and you find have one model. But of course we have black box models and there are many approaches trying to solve the explainability issue, but I do not see that the issue is solved. So I think in the future or Actually, already right now I think we will see more AI specific applications, I mean sector specific applications of AI so that models are built from the onset on with explainability in mind. So I know like I recently came across one company focusing only on lender decisions and really having this approach from the very beginning on. And this is very natural, a very, you know, foreseeable development I would say, because we have those principles, privacy by design. So you have to build privacy in the whole architecture from the design phase on or we have security by design. So now we have responsible AI by design. So you have to build them from the beginning. I think with a lot of models that have already been built, you will probably not achieve full transparency and explainability. So I'm totally with you here.

Wow, what a great response. Tackling that from all possible angles. I think we have time for hopefully one more question here. Katharina. So May Brit asking because you know, you, you talked already about how different countries are handling privacy differently. So May Britt's question is, do you find a specific country or area more effective in their AI privacy and AI governments? Because obviously, you know, the EU and some other, you know, countries throughout, throughout the world kind of govern their AI privacy and data much differently than here in the usa, which essentially, at least for now is self governance from, from the largest, you know, companies here in the us. So what's your take? Are there certain, you know, in your opinion, in your experience, are there certain areas or countries that are maybe more effective at data governance and AI privacy?

I mean, yes, of course, the first thought is the EU has already this golden standard of data protection with the GDPR and is coming up with the EU Air act, which will also extraterritorial effect because if you will offer your services in the EU, you will also have to comply with the EU AI act, classifying AI systems into different risk categories, etc. But I mean, what is, what is effective? I mean if there's also this concern that those, that the.

That's a good, good point.

Yeah. AI act is so effective that it will actually might, you know, hinder. Hinder. Help me here with the word like, you know, the startup scene, like, you know, make it more difficult.

Yeah. It might stop companies from being able to grow fast.

Yeah, yeah. So I mean, what is effective? And also I just wanted to mention, because you said, you talked about the dos, you mentioned it, we do already have a lot of regulations also in the US which applied to AI. It's not that there is nothing here. So we have a lot of sectoral law. There are a lot of state laws now popping up which tackle AI. A lot of AI issues are also regulated under privacy law. So that it's not nothing. It's just so complex. It's like a patchwork. So more effective. But I think more effective would mean it's, there's legal clarity. So I think if we have one law in the EU and maybe one law or executive order or whatever in the United States on a federal level, that will make it more effective. And I think this will be coming. So this was a kind of. Yeah, not, not, not so clear of an answer. But, you know, some thoughts. It's.

No, I mean, but that's, that's such a good point because, you know, one person's definition of effective AI, privacy and governance might, might mean to someone else, you know, stunting business growth or, you know, keeping companies from fully scaling at the pace to, to which they might want. So we tackled so many important topics on today's show. Katharina, thank you so much for joining us because I know we went all over the horn. Thank you for sharing your expertise and knowledge with us all. Thank you so much for joining.

Thank you so much for having me. Thank you, everyone.

All right, and just as a quick reminder, like Val is saying, looking for the newsletter on the topic, don't worry, go to your everydayai.com or if you're listening on the podcast, just look at the show notes. You'll find a link there because we're going to be sharing some of the articles that Katharina mentioned that she wrote. And we had so much in this episode dealing with AI, privacy and government. So don't worry, Go sign up@youreverydayai.com so thank you for joining us and we hope to see you back again for another edition of Everyday AI. Thanks, y'.

All.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit your everydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.