Privacy and AI: Balancing Innovation & Security

Great to be here. Delighted to join you.

Again, you know, happy to be here. Happy to join you. Look forward to talking about privacy. It's a great issue.

Right? It's one I've worked on, as you know, you know, now for, you know, over twenty years of my career. I got really started on it when I was spending time in Brussels. I lived for Brussels for almost five years working for a US law firm at a time when, you know, the data protection directive had been enforced for a few years in the EU, and companies were beginning to think about how do we transfer data from Europe to the US.

And it turned out, you know, there was the safe harbor agreement, and there was a need to design programs and to enable people to certify. And and so, you know, from there, you know, we've just seen things continue. We've seen the general data protection regulation in Europe. We've seen state private laws in the US and sector specific laws.

Now we see laws in China and India and elsewhere in the world. So it's been a great topic to work on and delighted to have the opportunity to talk with you about it.

Right. I think, you know, we're at really a transition point in terms of technology. Right? We started off where we had, you know, mainframes, and then we had PCs, and then we had server architecture, artificial intelligence.

And for a long time, we were really focused on and still are in many ways on predictive AI, taking these large data sets and finding new insights in them, being able to, like, drive insights that we couldn't previously because we lacked the computing power. But now we also have generative AI, these large language models that can create content and have sort of this syntactic, you know, memory or understanding to be able to answer queries, operate chatbots. We've seen a lot now even around agentic AI where you have these agents that have access to back end databases that they can then interact in a very sort of natural way.

And with all this, as we've seen all of these technologies evolve, we've seen, you know, new opportunities, new business models, but we've also seen new risks. And it's important when you think about implementing any technology to have sort of a core set of ethical principles that ground it. And so, you know, twenty nineteen at ADP, you know, we adopted a set of AI ethical principles, things like transparency and explainability. We're going to explain how our AI works.

We're gonna be transparent about where it's used. Data quality. We're gonna make sure that there's high quality data that's used to power these AI tools. You know, minimizing bias to avoid bias, to make sure that the results are accurate for their intended use.

These types of things to make sure that as we develop tools, we do it in a way that enable our clients both to meet their regulatory obligations as we see increasing numbers of AI laws, but also to do the right thing for their employees and for their customers.

Look. I wish I could take credit for our early start, but I only arrived in twenty twenty one. So my manager was really at the forefront of doing that work. So she gets definitely a prescience award on this.

But when you think about a privacy first strategy, right, it's important we use AI to begin thinking about, well, how do we protect personal data? And so you think about sort of a simple query. Right? Alright.

I wanna know say, I'm a small business owner. I wanna know who's my highest paid employee.

Well, you know, if I have an AI agent that is, you know, tied into my human capital management system, I can ask that type of question.

But the AI agent doesn't need to know who the employees are. It just needs to be able to figure out who's the highest paid. So you might have this list of five people. You have, you know, John Smith, and you have Mary Jones, and you can tokenize or pseudonymize the names and send a list of, well, here are employees one through five, and here are their salaries.

And it comes back, and it tells you, well, employee two is the highest paid, and then you reassociate that on the back end, and you realize that that's Mary Jones. So in that case, when you do that type of tokenization or synonymization, you comply with data minimization requirements. You reduce the risk overall in the system. You don't give the large language model information that's not gonna use to answer the question, but you still enable this new type of service.

Yeah. I think that's mostly right. What I would say is, like, figuring out what data it is to really drive the insights.

So it takes something else like candidate relevancy.

You need to know, maybe for example, that I have a law degree. Right? Because maybe the role is one that requires a law degree as for a lawyer. But you don't really need to know where I went to law school. Right? You just need to know that I meet that qualification.

And so you can take out, for example, who my employer is. You can take out my name. You you don't really need to know who I am. You can avoid biases whether positive or negative. You can then figure out, sir, who's the most relevant person, and then when they meet their criteria for the job and you've defined those criteria well, you'll open up opportunities for perhaps nontraditional candidates or for people from places or institutions that you haven't thought of. And so it's by focusing on what data is really relevant important and what extraneous in order to get sort of the best predictive behavior.

Right. So when you think about readiness, you've got to think about sort of all the aspects that go into systems. Right? If you think about the EUAI Act, one of the interesting things about it is that it really takes this life cycle approach.

It's really almost a product liability statute, and it's really that's something new and different in the software area. And so if you think about privacy compliance traditionally, like with GDPR, with other privacy laws, it's all about rights and notices and processes for giving those rights. So, like, I tell you what data I'm gonna collect. I give you the ability to access or delete that data.

I let people know who my subprocessors are. You know, I have appropriate protections of data. Like, it's all about that. Here, it's about sort of how you build the product.

It's starting from, you know, having good data quality. It's about testing for accuracy.

It's about testing for impermissible bias. It's about giving information to if you're developing the product to the deployer, if you're the deployer to the end user, that AI is being used and some visibility in terms of how it works. Again, that explainability and transparency all the way through post market monitoring. You wanna make sure the product doesn't drift.

If it's accurate when it starts, as it continues to learn, does it stay accurate, or do things come in that change how it operates? And then being able to take appropriate remedial actions there. And so when you think about readiness, you have to think about sort of how do I build that entire life cycle for how I'm going to either develop my AI service or I'm gonna deploy an AI service that someone else has developed. And and from a compliance perspective, right, I think you've got to think about all those things.

You've got to think about, alright, what's the data that's going in? Do I understand my data? Is the high quality data? Have I taken the appropriate data governance steps to make sure that that data is clean, that it's normalized, that it's representative?

I have to think about, alright, what data am I gonna share with an LLM or what data am I gonna use for the predictive learning? You know, is it really relevant? How do I exclude extraneous data? How do I make sure that, you know, I'm really using things that matter in this and not only that, but that matter in a way that aren't proxies for, say, impermissible characteristics.

And then I have to think about, well, what is the output? What am I going to tell people about how to use the output? What does the output mean? What are our appropriate users?

What are not? You know, my dear colleague who leads our product compliance team would interject here that, you know, one of the most important things is human oversight. We don't have these automated decisions that just sort of go into a black box and split out an answer. You always have to have a human in the loop to look at the outputs, to look at what's there, to validate that it makes sense.

You know, I think those are all the types of things that you have to think about from a compliance perspective.

So, look, I have a couple of thoughts about that question. So I'll start with the trust point. Right? People won't use technology that they don't trust. They won't share information if they don't trust you to manage it properly. So you have to, again, when you think about personal data, really have a good privacy program or privacy policy so people understand they understand how the data is gonna be used. They understand it'll be used in a way that's consistent with their expectations.

Do they understand what rights they have with that data? And so that they feel comfortable in trusting you with that. So I think that's one important thing. More broadly, when we think about AI, trust comes from, again, this transparency and explainability.

I'm gonna know when AI is gonna be used. I'm gonna understand how it operates. I'm gonna understand that there's appropriate oversight of it. But I would differ with you a little bit on the question about whether the legal industry has low trust in technology.

We started by the outset. I've been practicing law for over twenty five years. And even when I was in law school, you know, we were using Westlaw and Lexus, and I think they had pretty quick and fast uptake. Law libraries are largely a thing of the past.

So I think it may not be some of the technologies that we're quite as familiar with elsewhere in the world. But, like, I think there's a lot of legal specific technology that's gotten a lot of penetration for a very, very long time.

Look, I think that's a fascinating question.

So I think we have to we have to start because it took me long enough to get there, I'm teasing. So But we have to start with understanding, you know, Professor Solove's view of the world.

Right? And Professor Solove has a sort of a very interesting view in this book and as you point out in his articles where he's focused on he thinks the focus of privacy law and to have an effective privacy law that protects individuals. You have to focus on privacy harms and preventing those harms. In that sense, he might be very well aligned, say, with the EU AI act, which is focused on AI risks and how do we mitigate AI risks.

You know, there are a couple of other ways of looking at privacy. One that's probably more US centric is about duties. Right? There's a lot of discussion even in some legal proposals about fiduciary or other types of duties that people who have data owe to the person who provide the data.

The more traditional European way, which is the way that I personally tend to think is the most resonant, is about rights. Privacy is an inherent subject we have a right to, and that privacy is about how we protect and vindicate what essentially is a fundamental human right. It's not absolute. It lives in tension with other needs and rights, but it's something that's important.

In that theory, I'd go back to Brandeis's famous paper from eighteen ninety in the Harvard Law Review about the right to privacy, which he defined as the right to be left alone, but also to the OECD privacy principles. Plus, I'll admit, I'd like saying I'm a human rights lawyer. Right? Because that's what I really think privacy lawyers are.

But now that we sort of taken that tour through sort of the three different ways of looking at privacy law and how privacy protects folks when we go look at sort of AI adoption. Right? The way that we've really looked at it because they're sort of not it's more about risks in here. And so that aligns a little bit with what Professor Solove is saying.

And it's again you know, if you have a public LLM and you put in data and it learns, right, eventually, it's going to learn that, hey. Maybe Jason Alberts, the chief privacy officer at ADP, in the sense that it will structure sentences that contain that because it knows that those terms are frequently associated together. And so it's important to make sure that you have appropriate protections. You know, when we implement, you know, generative AI in our products or even in our internal services, and I think this is largely true of others, we bring in LLM models.

We put them in a controlled environment. They're not trained or they're only fine tuned under our control. Whatever they learn doesn't go back to the provider of the model or to the environment. And so, again, that way, it's sort of a self contained system, and you don't worry about, you know, leakage of intellectual property or a personal data or of other things.

And so, again, you've got to think about what are the appropriate security controls? How do you prevent prompt injection attacks that cause the LLM to give answers it's not supposed to give or to say things it's not supposed to say? And there's a beginning to be a quite sort of sophisticated industry of trying to understand what are these types of protections, what are these types of risks from both a security perspective, from a privacy perspective, from an intellectual property perspective, to be able to construct systems that people can have confidence in.

Yeah. So I might use a slightly different analogy than sort of siloed. Right? Like, because if you think about siloed, you think about sort of these big sort of grain towers in the Midwest are standing each alone on its own. What I'd like to think is we have sort of an overlapping set of sort of state privacy rules that apply to consumers.

California has a law that applies only to consumers but to employees.

Then we have sector specific laws at the federal level that cover banking, that cover health, that cover videotape rentals, things like that. And so, you know, I tend to view this a little bit more like a patchwork quilt perhaps than silos, but, you know, I wouldn't disagree with the assertion that there's some holes in that patchwork quilt. It's not a completed work of art that's gonna go, you know, hang in the Smithsonian. So I think in that case, what we're gonna see is we're gonna see continued work in this area.

You're gonna see laws that continue where there's consensus. So at the federal level, there's a lot of consensus about doing even additional things beyond the Children's Online Privacy Protection Act to protect children, and we came very close the last congress. And probably in this congress, we'll see more legislation there. At the state level, we see more children's legislation.

We're seeing, again, states adopt privacy laws beyond the nineteen or twenty that have adopted them to date. You know, there's a lot of interest in a federal privacy law. The challenge is there are a couple of issues on which there's some pretty sharp divides, and there's a deal to be done there, but the question, you know, arises whether people are want to do that deal or prefer to continue to have this iterative approach. But what I would say is this, right, like there's this sort of perception, well, Europe has privacy laws and the US doesn't have privacy laws.

And I don't think that's true. You know, there are a lot of US lawyers spending a lot of time on privacy for us to say that there really isn't a lot of regulation here. So I think, you know, we've just approached it a little bit of a different way and are continuing to, you know, build and expand that quilt so it covers the entire bed.

Well, look. I'm generally an optimist. Right? I think that we continue to evolve our legal systems. We continue to evolve them in Europe and in the US and, you know, in Canada and in Asia Pac and in LatAm.

And, look, I think we are responsive to technology because ultimately, right, it's important for us to have these protections. Right? Technology is not this inexorable force that acts on society.

At the end of today, society determines how technology is going to be used. And so, you know, as we learn more, we're able to adjust more. We're able to experiment in the state laboratories. We're able to adopt these types of protections.

So, again, is it perfect? No. I don't think anybody would suggest that it is, But I think that we take good and effective steps to protect people. I think those are gonna continue to evolve and grow as we continue to learn more about AI, as we continue to explore, you know, new types of data usage.

And, you know, look, you don't want to be sort of in a legal area that sort of, you know, static or ossified in some way. And so that's what makes it interesting. Like, every day, I can get up and there's some new interesting challenge or some new interesting problem. There's some new law.

There's some new technology. And, you know, we get to think about it and think about how it's going to be used, how it's going to impact the company, how it's going to impact our clients, how it's going to impact our clients' employees, how it's going to impact society.

Right. So look, data minimization is one of the core privacy principles. Right? It's embodied in the OECD privacy principles.

It's embodied in privacy laws, you know, more so elsewhere in the world than the US. In the US, we have to really have sort of a notice and consent system. And if you get somebody notice and get their consent, you can probably get, you know, maybe more data than the minimum necessary. But we're now even seeing in the US where data minimization is becoming an important, you know, principle.

You know, the Maryland privacy law really embraces that, and the, proposed federal law that got at least through the committee last year really took a data minimization approach. And there's some benefits to this. Right? Because when you think about, you know, notice and consent, the burden ends up being on the individual to, like, read the notice, you know, consent, understand what they're consent to.

It puts a lot of the compliance burden on the individual.

Whereas the data minimization thing says to the individual, here's the service I'm going to offer you. I'm going to collect data from you, but I'm not going to collect any more data than I need to provide the service. And so there's a match between something that I want. Right. That I'm going out and buying or procuring and the data that's being processed in order to provide that to me. And so it's an important protection because it means that there's not extraneous data used or data used for different purposes.

And AI is really like any other service in that regard. You know, we've talked about this before. Sort of what are sort of the core elements needed to go answer a query, to power an agent, to be able to generate content or, you know, in the world of machine learning to make predictions, to be able to derive insights from data. You know, we might say, well, you're deriving insights and you're doing all this complex linear algebra.

More data is better, but more data is only better if it's relevant. And so, yes, we can use large datasets, but they need to be relevant. And data minimization, again, is focused on that relevance. What personal data do you really need in order to provide the service?

Look, I think that, you know, as you point out, this is a little bit of a complex calculus, but the way that I would suggest thinking about it is fundamentally you're not in a single instance prisoners to limit gang, right? It's not as if you're going to ask me for data once and then you're never going to come back and talk to me again, right? And so you have to think about what's the value of the data for the services that you want to provide and what's the value of the ongoing relationship? Am I going to trust you?

Am I going to be willing to provide more data later on, maybe for a different service? Maybe there's some new technology and you wanna come back. And so I'll go back to the point that it's important to establish trust with data subjects and strong privacy programs and strong privacy protections, you know, allow companies to do that. And that's what causes people to be willing to share data, but you can then drive those insights or use.

And so you think about, you know, data is definitely an asset, but it's an asset only insofar as it drives some other form of value, some sort of insight, some sort of ability to communicate, you know, something like that. And it's only a sustainable asset if it's built upon trust.

So it's not like I'm going to go take a giant vacuum and try to suck everything up. You know, that's not a strategy that leads to long term success.

Well, look, I think, you know, we could talk about all the sort of standard things. Right? Go build a good AI compliance program. You know, anticipate upcoming regulations.

You know, look at what's happening in the US, for example, with the Colorado law and now other state laws that are pending in the legislative sessions this year. The trite phrase for this is skate where the puck is going. But I'm gonna go in a little bit of a different direction and I'm gonna continue to stick with my theme of trust. And I'm going to say, give people choice about the use of AI.

Right. And I think certainly in areas where AI is used for consequential decisions, we're going to see increased regulation that's going to require that. But again, that builds trust that allows people to make the choice. And at the end of the day, when people choose to engage, when they feel like they have the option, that enhances trust, they see the value.

So I would say again, right, if you look at something like candidate relevancy and giving candidates the ability to opt out of having AI evaluate their resumes. Right? You know, there's a New York City law that requires it for small population, but when we built this, we did it nationwide. And again, it's because we want to make sure that as our clients, you know, engage with their prospects for employment, that there's a level of trust there that allows that.

And I think, you know, again, taking that type of approach in the long run is going to help companies succeed.

Not really, actually. Right? Like, I use it a little bit and, you know, I'm part of a pilot here of some technology, but I'm a very picky writer. Right? And so at the end of the day, if I use generative AI to generate content, I end up having to edit it and it doesn't sound like in my voice. And, you know, after years and years of honing a writing style at a law firm I sort of like my own voice and so maybe you'll just say well you can't teach an old old dog, new tricks but to me generative AI is an amazing technology in terms of what it's capable of doing But for what I do in my day to day work, I tend to want to sound like me.

Well, look, I think a couple of things. I think we're going to continue to see AI laws adopted. Right? You already know that Virginia's legislature passed one that's awaiting the governor's signature that looks like a lot like Colorado's or other state legislation that's pending and legislatures that are still in session.

I think we're gonna see, you know, that continued trend on the international scale. We're gonna see companies doing more as the deadline for implementing other parts of the EUAI Act approaches. So I think we're gonna see a range of things on the regulatory front. And and by the way, not just our formal regulation, which is what I've talked about, but we're gonna see continued adoption of standards and codes of conduct and practices in that area.

And I think that's going to reflect sort of an increasing adoption of AI by companies. Right? Like the big thing right now is a Genentech AI and AI agents. And of course, at ADP, we have ADP Assist, which is our agent for helping, you know, our clients, you know, achieve more with our products.

And, you know, we're now seeing things where, well, I could even tell AI, alright, you know, go book me a trip to the Bahamas. Right? And it can go to the website and it can go do that. And we've seen some interesting technology there.

So, look, I think, you know, we're going to eventually, at least in the generative AI side, not too much further, get into an area where it's less about sort of getting it, draft a blog post. It sounds like me and it's going to be more. Am I going to have my own little digital assistant that I'm going to be like, all right, you know, go book me, you know, reservations for two at the sushi place my wife and I like and, you know, for seven PM on Saturday, and I don't have to worry about it. And it'll magically be there and it'll even put it on my calendar and whether it's a trip or, you know, other things.

And so I think what we're gonna see is, like, again, increased technology just like, you know, now I can find everything at my fingertips. When I was growing up, I was limited to whatever was in the World Book Encyclopedia on the bookshelf. And there was a lot there, but it was organized in particular fashion, and it didn't change year to year. So I think we're only seeing the tip of the potential of the technology, and I think it's gonna integrate itself more in our everyday lives in even more useful fashions.

Yeah. I don't think we're gonna get there. Right? Like, I'm optimistic about a number of things, but I'm really optimistic about human creativity.

I think AI is really good at helping to take care of rote tasks, routine tasks, and help us unleash human creativity, whether we want to use it to write, whether we want to use it to innovate in our jobs. But AI learns from sort of the, you know, installed base. Right? Whether it's sort of your existing set of writings or whatever corpus has been used to train the large language model.

It doesn't have the ability to create new things beyond that sort of set of knowledge we as humans do.

And that human creativity, whether it's in the workplace, whether it's in the home, whether it's in our community, is going to continue to be important. So AI will give us more space to embrace and unleash that human creativity, but it's never going to go to work for you, Jared.

Well, look, I think as with any technology, you've got to understand it and be able to master it. Right? I used the example when I was in law school, we all had to learn Westlaw and Lexus and how it worked and how to shepherdize without breaking out the big red books, if anybody remembers those. I may be socially old, but I'm the only person who remembers those.

But at the same time, right, what is law school fundamentally about? I've been reflecting on this in light of my thirtieth reunion. I'm going up in April for that. And apparently, thirty years ago, somebody taped me at the end of law school and a few other students about our thoughts about law school life and careers.

And they're gonna play this, and they're gonna have us on a panel saying what we think now. And all I can say is that the potential for embarrassment here is sky high. But again, law school teaches us how to think. Right?

And so understanding those technologies and mastering learning these technologies by understanding how to think, you know, because it's how that sense of in the legal world of legal reasoning of the case method of legislative interpretation that ultimately is going to allow people to succeed in those sort of skills have been Jason, you said some incredibly profound things about human potential, human creativity in the face of just rapidly changing technology.

Thank you. Great to be here. Really enjoyed the conversation.