AI Is Eating Our Young: Data Center Revolts, Vanishing Junior Jobs & the EU AI Act

Justin Beals: 
Hello everyone, and welcome to Secure Talk. I'm your host, Justin Beals. This year, the largest technology companies in the world will spend more than $700 billion building data centers to power artificial intelligence. Global data center spending is on track to pass a trillion dollars. That is more money than the economies of all but a couple dozen countries on Earth, poured into warehouses full of chips that draw as much electricity as a small city.

And a lot of Americans have started saying no. In the first three months of this year, local opposition blocked or delayed 75 data center projects worth about $130 billion. That is roughly the same amount that was stopped in all of 2025, compressed into a single quarter. The number of organized opposition groups grew by more than double to 800 spread across 49 states. 

In a Gallup survey this spring, 71% of Americans said they did not want a data center built near them. And the reasons are the ones you would expect if you lived next to one: higher electricity bills, water usage, noise, the sense that a community is being asked to hand over its resources to an industry that promises a lot and pays back very little. 

My guest today, Fran Berman, ran one of the country's national supercomputer centers. She has spent her career thinking about what it takes to run computing at massive scale, and who bears the cost when we get it wrong. Her argument is simple, and I think it is the right one. The technology we depend on is critical infrastructure. We govern our roads and our bridges and our power grid for the public good. We build them to last and we build them to be safe enough for everyone to use. We have never done that for the AI sector because the sector is solely built on hype and massive amounts of private investment. And now the bill for that choice is showing up on people's utility statements. For the people who listen to this show, that argument is not abstract. 

Many of you work under CMMC, the Cybersecurity Maturity Model Certification, which asks defense contractors to prove their security practices for the first time. Think about what that requirement actually is. We looked at an industry that private companies run for profit. We recognized that our military depends on it, and we decided that dependence was reason enough to hold those companies to a standard. We regulated private business because the public, in this case our national defense, was on the line. That is the exact move Fran is asking us to make with data centers and AI. If we can look at the defense supply chain and say this is critical infrastructure, it has to meet a bar, then we can look at the trillion dollars of compute being built into the middle of American life and say the same thing. The precedent is already set. CMMC is one place; we have already written the guardrails Fran has been calling for. The question is whether we do it for the rest of the infrastructure we now depend on. These are the ideas at the heart of Fran's book, Better Tech, and we work through them in our conversation. 

While safety cannot be bolted onto a system after it ships, why AI is quietly eroding the training ground where the next generation of engineers learns its craft, and why high efficiency and good quality of life are not the same thing. Fran has thought harder than almost anyone about how we live alongside these tools without letting them run the place. 

Dr. Fran Berman is an award-winning data scientist, a pioneer in public interest technology, and a community leader and builder.

She directs the Public Interest Technology Initiative at UMass Amherst and is a faculty associate at the Berkman-Klein Center for the Internet and Society at Harvard. Berman is the former head of the San Diego Supercomputer Center and served as the Vice President for Research at Rin Saylor Polytechnic Institute. She currently serves as a trustee of the Alfred P. Sloan Foundation and is a popular regular panelist on public radio's. WAMC Roundtable with 400,000 monthly listeners in seven states. For more information, see Fran Berman dot com. But today, join me in welcoming Fran Berman to the Secure Talk Podcast.


—-


Justin Beals: Fran, thank you for joining us today on Secure Talk. We really appreciate it.

Fran Berman: I'm really happy to be here.

Justin Beals: Excellent. We got interested in interviewing you around your book, Better Tech, which I really enjoyed. But before we dig into that, I wanted to chat a little bit about an area you've been writing about recently, something that in my early career I spent a lot of time working on, and that's education technology. You've been raising some alarms about how we're teaching kids, and I love that you're doing it from a computer science perspective, not necessarily as an educator.

What pulled you into this discussion around education?

Fran Berman: I think I've always been really passionate about using education to prepare people for the work world, prepare people for life, prepare people for citizenry. And these days education is getting harder and harder. And I see that from a bunch of perspectives. I'm a data scientist, a computer scientist, so I worry about that in terms of how we're building the technology and how we're using the technology. I worry about it as an educator because I want to make sure that we can do a really good job. I worry about it as a parent. I have kids myself, and I worry about it as a citizen because my neighbors and my friends are worried about it.

So there's a lot of different places to think about it. As a computer scientist, some of the things I really worry about is I think it's really important for us to always remember that the technology we build is embedded in society. And all of the decisions we make as we develop that technology have consequences. And in particular, we decide how much privacy, how much security, what kinds of functionality should we have? Can it operate independently? If so, what's the risk? And all of those kinds of things, they're all little dials that we can tweak. And we're really good at tweaking those dials to make products and services enormously profitable. And my hope is that we get better at tweaking those dials to make products and services protective and empowering and low risk for the people who use them.

As an educator, it's really important to teach students how to navigate this amazing world of technology that has great benefits and great risks. And for me, that's really important that they know two things. First of all, that they can really stand on their own. They know how to think, they know how to write, they know how to navigate the world. They can do things competently where they have tools who can do it wonderfully, so they know what to do with the outcomes of those tools.

Fran Berman: For example, we have students; you can give AI just about any assignment we give on a college campus, and an AI can do a competent, if not fabulous, job. But the whole point of giving students an assignment is to give them practice with skills, writing skills, coding skills, problem sets, all kinds of skills. And if they're farming it out to a tool, they're not learning. 

And so, our job now becomes how do we teach students to both build the skills they need and to use the tools that will be available to them in the real world and in their jobs. And those are really hard things to do.

Justin Beals:  Yeah. I had some conversations recently with some young folks who mentioned that they had considered a degree in computer science but elected mechanical engineering. And of course the tagline for them, well, AI is taking away the coding jobs. You know, in our work, we certainly have seen some efficiency, but if I didn't have engineers that were smart about writing code, the AI tools would really run amok inside, you know, enterprise software that we're trying to build and support in a long term way. And so I dissuaded them from this perspective. You you kind of agree with that?

Fran Berman:  Yeah. I do, and here's it's funny that a colleague and I have just written a piece, and we're trying to figure out where to send it, and the piece is about the middle of the pipeline is really fraying. If you want to put a dramatic point on it, you know, AI is eating our young. And if you think about it, if students can use AI to code in school, so they don't really learn software engineering skills, and the place that they practice software engineering skills or writing skills or analysis skills are at entry-level jobs, and those are being taken by AI, then the problem is that really to use AI efficiently and well as a tool, you need someone with judgment and mastery and taste and who really knows how you can sort of which things you can delegate and how you can oversee the final product. Senior software engineers know that, and AI can be enormously helpful to them because they have the judgment. But if we take away this middle of the professional development ladder, who is the next generation of senior software engineers? Where do they get and it it's funny in this article that I hope is will appear at some point, we liken it to imagine doctors. So doctors go through a very extended period after they learn things in school where they do residencies or internships. And you know, imagine a servant a surgeon who's never had practice in the operating room under supervision. You know, we want to have professionals we can, and unless you have that experience and the mentorship of more senior professionals, it's really hard to do. And that happens for a variety of things. So it's not that AI can't do some of these jobs competently. It's that we need AI to do the work that makes things more efficient, but there's got to be room for human beings to climb that ladder. And I think that's problematic for us now.

Justin Beals:  Yeah. I agree completely. You know, so much of the conversation and around especially the AI tools is kind of hype or panic, you know, one way or the other. You and I have probably been building models and, you know, dealing with data science and predictive systems for many years. I see incremental progress, which is exciting. No, don't get me wrong, I love it. But I'm curious how you see the use of these tools for the good in the classroom, you know, AI especially.

Fran Berman: Yeah, I mean, I think you're so right about the hype, and it's so distracting and misleading for the public.

This idea that AI is some sort of nascent superspecies that will take everything that human beings can do and, you know ultimately do it better. It's distracting, and it's incorrect. AI is a terrific tool. It's a terrific tool that we need to learn how to use safely, responsibly, and effectively. and we have to do the things that only humans can do, and we have to partner with these.

tools to do things well. So in the classroom it's an interesting situation, Justin, because in some sense I think educators know how to teach both skill building and tool using. We know how to do that. A lot of personal attention, a lot of discussion, small group critiques, having students sort of think about things and then code them, and you can mentor them.

And the problem, I think, one of the problems we have in education is those are all kind of small class activities. You know, small classes where you can really dig into things, and you can really help the students develop not just particular skills, but as professionals and as people and help them mature. And the problem is that universities are really strapped these days, and our business models are at the breaking point, truly. 

And so that for large public great universities, you have to sort of focus on the bottom line. So faculty don't have that many opportunities to have small classes that are really hands-on or few classes that are really hands-on. There's a lot of teaching, there's a lot of service, etc. So you're swimming against the stream to try to do this. And of course an easy way out is to use AI. To use AI to develop your curriculum, to use AI to do the homework. You know, it's AI on AI in a lot of places. And so, you know, we want to avoid that. But we have an ecosystem that makes it really hard to do the things we need to do. I mean, I'll end with a positive story. In my own classes, I want the student I and and primarily my students are using AI for critical analysis and writing. And I want the students to both be able to learn how to be better writers, better communicators, and I also want them to learn how to use AI to enhance their writing. And so first of all, I have two kinds of assignments. I have assignments where they're permitted to use AI and assignments where they're not. In the assignments where they're not, I try to make it hard to cheat with AI. I ask them questions about their personal experiences or their assessments or you know, things that would be more connected to them. And in the classes that they're permitted to use AI, I give them a whole lecture first about here's what to expect from AI. By the way, not everything you put in AI is private. You have to read the terms and conditions. AI can be helpful as a developmental editor. Or if you have a group doing a long piece, it can smooth the voices or it can help you structure it. It can come up with some ideas, but if you just hand it to AI and say, give me something, you have learned nothing. And so first of all, I try to convince them of this. Many students are actually convinced of this. And then I ask them to turn in the assignment and show me all the AI prompts they use.

And so when I can look at the AI prompts, I have a sense about how they're using AI. Are they using it in a way that made their writing better and is really helpful? Or are they just kind of sticking it in AI because they were too stressed or too lazy or too overwhelmed or too something? And then you can kind of follow up and be more helpful to the students. But it's hard. I have to say, all of my colleagues and I are really struggling to find the right way to do this.

Justin Beals:  You know, this really resonates with me in my own use with the large language models. If I didn't have a background programming, I would have a very hard time getting some of these large language models to produce the work. But it's because I kind of can think through the thought process I want the LLM to take to reach a result that we step by step by step reach a particular outcome. But if I hadn't played with programming for so many years, I think I would have jumped to the end and been like, gobbledygook came out the other side. Yeah.

Fran Berman:  I mean, one more comment about the education thing. In the book Better Tech, I actually put a whole appendix in. One of my reviewers said, Well, you teach this class, put in a syllabus. And so I put in a whole appendix with a syllabus and some of the exercises I asked the students to do, and some guidelines for the lecture about, you know, how you might use AI. And my hope is that it's really helpful for folks teaching classes or doing trainings or really thinking about how.

Justin Beals: Yeah. Speaking about the book a little bit, in your prologue you write a phrase that technology is not a religion. Thank you, by the way, Fran. I don't think we deserve any amount of worship in the least. And that we really should treat the tech we depend on the way we treat food or water, something held to a real standard, that maybe an ethical standard or an operational standard, and not just a product we buy.

I'm curious what made this the right type of comparison for you.

Fran Berman:
It's a really great question. I've been involved with technology for a long time. And in addition to being a professor, I've run a National Supercomputer Center and I've co-founded an international organization on data infrastructure and a bunch of different things, and all of those things embedded technology in the real world for the people who need to get the most out of it. And I think, you know, as I've always been worried about the outcome of tech and the consequences of the decisions we've made. And now, you know, as a person in the real world where tech is so advanced and so, you know, useful and used in everything. It seems to me that a lot of times this check puts us at risk. And we're just used to it. We're used to being troled. I don't know how many times, I mean, you can give me a rough count for you, but I can't tell you how many times I've gotten a letter that says there's been a data breach. Or just recently there was $13,000 in my bank account at hotels I've never been to, and you know my credit card had been stolen. It's a dangerous place. 

There are predators for kids, there are, You know, people doing all kinds of things with tech. And I started thinking about this is critical infrastructure. We depend on this. We can't think about living a modern life without using so many different kinds of technologies. But we don't treat it that way. If you think about our roads and our bridges and our financial system and our health system and nuclear power plants and the internet, all of those are governed in a way that is supposed to be good for the public. 
We build them to last, we build them so they're fair, we build them so they're secure enough and private enough for widespread use. But we don't think of that when we build technology. Typically, we're building it for profit because it lives in the private sector. But for the technologies that become critical infrastructure, why don't we then develop the right kinds of guardrails, create new ones that actually serve the public. They empower us instead of endanger us. So I started thinking about the whole concept of if you if you reframe tech as critical infrastructure, how can you do that? What can you do to make yourself safer? How can that happen? And I think the book came from all of the ideas about what people could do. Because I don't know about you Justin but I do a lot of reading about tech, and there's a lot of doomsday books out there. 

You know, tech is out to get you; evil robots are, you know, going to command to rule the world. And what I worry about is, you know, normal people, it would be good for us to be empowered in that world. It would be good for us to think about it, at least in the way we think about other kinds of critical infrastructure, which by the way is not perfect. Bridges fall down, and banks fail, but it's not an everyday occurrence. And it would be good to make the tech we use safer.

Justin Beals: Yeah. Absolutely. I think one of the reasons that I decided to build my current company inside the compliance space is that I was really frustrated with the education industry because I saw us working to improve the scale of accessible education, but we didn't even maintain the standard of efficacy of education. And I really I got really frustrated actually because I'm I consider myself a lifelong learner. I love that. And I didn't want to make learning worse. And and in some ways the business side of the code that we wrote maximized the profit, but wasn't held to a standard of delivery. Yeah.

Fran Berman:  Yeah, yeah. And by the way, that's not because people are evil. It's just a different metric of success. If you don't make a profit, you don't survive. But this is I think where the public sector is really important. The whole mission of the public sector, ostensibly, is the well-being of the public, the public good. And so the public sector can be proactive in terms of sort of creating a playing field where private sector companies have to comply with standards that make people safer.

Justin Beals: Yeah, absolutely. You know, when we talk about standards, a lot of our listeners for this podcast work under some rules that the government now is enforcing on defense contractors of the defense industrial base largely. Recently, a security compliance requirement, the cybersecurity maturity model certification, is asking companies that for a long time have have never really had to prove their cybersecurity practices, you know, in operations with
the DOD. Does this feel like the type of thing that you're calling for that we kind of enunciate what the requirements are to be secure in the way we operate tech?

Fran Berman: Yes, and with lots of caveats. First of all, to the best of our ability, you I'm sure you can say much more than me about how we're always behind the eight ball when it comes to security. 
There are you know, a whole world of attackers and only so much we can imagine at any one time that we can do. Security is also problematic for a lot of companies because it decreases usability, it takes more time, you know, you have to engineer systems to do it. It's not it's not your fastest way to make a profit. And so I really do believe that it's up to the public sector to try
to make the best rules possible to keep everybody up to reasonable security standards. 

Now, from the little I've read about this, my understanding is that the devil is in the details have been really, really hard for the community. And so that there's a lot of bureaucracy, and there's you know, it's really hard for small businesses to comply and those kinds of things. And you know, as I've read those, I have been thinking a little bit about the General Data Protection Regulation in Europe. So for your listeners, I'm sure your listeners are familiar with that, but Europe really did a really amazing thing in creating that around 2017, 2018, and it's really made a worldwide impact with the way we deal with privacy and security. It did that in a bunch of different ways. First of all, the Europeans, because of their constitution and their whole digital rights agenda can be very specific about what rights and what privacy rights that you have. Different than the US. I think that, but they but they put a lot of mechanisms in there for specificity and accountability. So to comply with GDPR, you really go through a whole process by which then you can then you comply, or people have been fined tremendous amounts of money, hundreds of millions of dollars, billions of

Fran Berman: So when I think about compliance, you have to kind of get it all right for it to be most effective. You have to use the expertise of people who understand what security is and what the users' rights ought to be. You have to be cognizant of what the community, the private sector community, can handle. And then you have to sort of balance those so that users really are protected in a way that people can handle, within the business structure they have. I think that's really tricky.

Justin Beals:  Yeah. It's it has been really interesting to me. I agree that some of the challenges in meeting these requirements for companies have been tough, and it is it does shift thinking, right? and the requirements are fairly stringent, and you know, when a bureaucracy like a government creates an assessment regime, it can go haywire. It's interesting though, how many companies I talk to at times that really, when we sit down and look at like, similar like GDPR or law or HIPAA, and we read through what it's asking for, a lot of it they're already doing. But they were terrified that they weren't. And I do more of that psychotherapy than I do of the you've got to change what's happening, you know. And then they realize, we're only twenty, thirty percent away from meeting these requirements. Just a couple of tweaks we need to make.

But it's very scary up front. And I think that the business side of consultative work wants to make it seem so scary that you can't accomplish it on your own instead of giving you an opportunity to. Yeah.

Fran Berman: Yeah. By the way, I mean the fact that they're pretty close to where they need to be means they were thinking about it and planning ahead. That's exactly what we want businesses to do.

Justin Beals:  Right. Yes. Yeah, absolutely. So one of the things that you point out in your book is that, you know, good regulation often isn't enough by itself. Like it's one thing to roll out of regulation. We could take GDPR. What are the other aspects that help GDPR, for example, be successful, or what needs to go beyond just a law or a requirement that emerges?

Fran Berman: Yeah, I think that there are plenty of laws that we have on the books that we pay no attention to. Most people jaywalk. Nobody I don't know anyone who's ever been arrested for jaywalking, and I don't think I know anyone who's ever not jaywalked. You know, people litter; you're not arrested immediately for that. Maybe you go a little bit over the speed limit, you're not, you know, so enforcement turns out to be a really important part of making a law effective. Another thing that's really important part of making a law effective is expertise in talking about the law in the first place. 

And so if you've seen a number of, you know, testimonies from the tech industry to people in Congress. There's not a lot of expertise in the government. And so the government farms that out a lot of the time to lobbyists and the tech industry, but we don't have a lot of independent expertise in the government. And so that's one thing we could have more of. You know, your mileage varies to mix my metaphors if you ask the foxes to guard the hen house.

So I think we have to be really careful on that. And the other thing is I think we really have to think hard about the jurisdiction in which we're in. So one of the things that I look in at in the book that fascinates me is privacy laws. And privacy laws in the US, the EU and China are really different. And they're different because the political environment is different. And so the mechanism by which you enforce them, the you know what they can do, what the government feels its rights are and what the rights of citizens are, all of those are really different. So I think you need the whole ecosystem to at least be considered when you're thinking about laws. You can't just slap a law on the books and sort of expect something to happen. Another example I use in the book is Vermont was really great about creating a data broker law. So it created one of the first data broker laws on the books to make sure that data brokers are not exploiting the public. But the problem with the data broker is that law was that all they had to do was register, and there wasn't a very big fine for not registering. So that didn't really address the fact that data brokers, you know, were exploiting a lot of people and those kinds of things.

Justin Beals: Yeah. Yeah. One of the chapters in your book talks about design as being a critical a critical aspect to designing safety, especially. And you know, you make the case that it can't be bolted on later. It's it's very hard to add it after it's been released. Why is it so hard to fix a product once it's already out in the world? I mean, we think about applications as very malleable bits of technology.

Fran Berman:  Yeah, it's it's really a great question and I think I mean I imagine many of your listeners who've worked with these big complex software systems know firsthand how hard it is to backwards engineer the software system for a different kind of functionality. And you know, for those listeners who maybe have not experienced that like much, it imagine you've you built a house or you moved into a house.

And now you've decided you need another bathroom in the house. And so what do you need to do? Well, you need plumbing, you need electricity, you need a wide variety of things. But you don't just plop that down and you're done. You may have to rearrange and re-architect the plumbing of your house. May also be true of the electricity. There's other systems you're using as well. If you added it onto the side, you have to change the roof if you have a septic system, maybe you'll be a capacity with a new bathroom. And so all of these systems have to be taken into account. And the same is true with a complex software system. You may want to add protections. That changes a lot of the places in the system where you might do things. It might limit some functionality, it might add functionality to something else. Those systems are not standalone, they work with other systems.

You might have to change the way you integrate or interface with other systems. There's laws that and policy that governs many of the aspects of your system. It may be that new laws now apply or or laws don't apply anymore and you don't have to be compliant with them. So, you know, like adding a third bathroom to your house, you really have to think about it holistically. You have to architect it in and you have to make sure it all works well together. 

And I think that it it's not easy ever, but it's much more straightforward to start from the beginning and say, yeah, we want three bathrooms instead of two in our house. Or we want privacy protections or more security in our system than what we have already. I mean imagine backwards engineering the internet to have, you know, a greater security structure. It's almost impossible for some of these systems.

Justin Beals:  You know, as it scales so quickly too. I just any architect that has been working on building these systems for a while has made a mistake. We're all terrified of tech debt, you know, and and that's why we spend a a little bit of thoughtful time on the design side of the systems, whether it's safety or scalability or or the ability to add new features later. we have been, you know, I tend to be more pragmatic about these things and we'll sometimes see a product leap ahead and we have this tortoise in a hair scenario where we're like, Yes, we we realize they leaped ahead, but this might be a slower methodology to get to a bigger innovation. But when we get there, it'll be accelerated in the innovation we can deliver. Yeah.

Fran Berman: Yeah. And and that's the hard and exciting part of thinking it through. It it and thinking about society. It's it's how will it be used, who will use it, what might they need, what are things we never imagined, you know, when Tim Berners-Lee first thought of the internet, he never imagined the things that we're doing with it today. It's really hard to predict the future.

Justin Beals:  Yeah. Speaking of which in your book I have another quote, how could self-driving cars be bad for the environment? You know, it it's easy to look at them and be like, this is this is great. This is only gravy, you know, to to quote my southern heritage. And but but why, you know, what why is that an example of it not being an obvious yes?

Fran Berman:  I'm really glad you asked me about that because I think that's such a great example of why we have to plan ahead, and architect these things for future scenarios, which is really hard. So if you think about self-driving cars, the prediction and and really with high confidence is 50 years from now almost every car will be self-driving and they'll all be electric vehicles. And so you think like what's the problem? And the problem is we need a whole new ecosystem to make sure that they're safe and efficient. And they do the things we want them to do. And by the way, of course tran automobile transportation is critical infrastructure.

So if you think about this, you know, what is it, 60 or 70 years ago, we had cars with tailpipes and emissions, and it took good science and help from the federal government to really get to the place where we could set emission standards and we could improve cars over time. And then when electric vehicles came along, you of course you know, you burn fuel and you use emissions when create cars, when you operate cars, and when you retire cars. You know, you have to manufacture them, and oftentimes we recycle some or some components of the cars and not. And so you think about electric vehicles, and when we operate the cars, they're certainly much better than than fuel vehicles. But if you think about the kind of electric vehicles we'll have for self-driving cars, we are pretty lot of systems in those cars. We've got cameras in those cars, we've got sensors in those cars, we've got computers in those cars. Those cars have to be connected to the internet. And so we have a big computational load because instead of the human observing, planning, sensing, driving, now we have computers doing that in a split second.

And so those computers take a lot of battery power. And all of those systems take a lot of battery power, and we have to generate a lot of energy for all of those. And so when you think we think about future scenarios, we have to think about are is there hardware we can architect and develop that uses less energy? Are there algorithms where we use the data just in time instead of just chomping along in the background and kind of storing that stuff that can use less energy. What about the built environment? Those cars are going to travel in platoons. They don't need stop signs. They don't need speed limits. They're going to get that from the internet. Well, that means that all of our built environments and our roads have to be able to accommodate a really stable internet connection. We have to build these cars so that if some for some reason they don't get that connection, it's not a catastrophic failure. 

Those cars have many more attack surfaces. And so now instead of having an accident, I don't know, because you ran into another car, maybe you have an accident because you had a denial of service attack. And so there's all these ways in which we need more computers and more energy. And if we think ahead, we'll come up with really good techniques for sort of keeping that down to a level that we can accommodate. If we don't think ahead, then we're in trouble.



Justin Beals: Yeah. And I don't think what you're proposing is that we have to have an answer to all these questions, but understanding what the questions are and the level of risk and prioritization of asking them and what amount of innovation or product can be delivered safely i is a more pragmatic approach.

Fran Berman: Yeah, I mean I think that It's time for us to be thinking really holistically about all of that. We have to think about, yes, what's the best kind of computation, what's the best kind of you know, data usage, et cetera, but we also have to think about what are the environmental implications, how will they be used in society, will they be riskier in some way that we hadn't thought of? What about privacy in a self-driving car? You know, it's listening to everything because that's useful information. for it. What happens to the conversations I have in the car or the bystanders it sees outside of the car? You know, can those be, can they be used inappropriately? We need to make some sort of rules and standards and regulations that really help us use those things safely.

Justin Beals:  Yeah. And I tell people all the time, like these are for our benefit that we develop a law or a regulation or a standard. Because if we didn't, then it's very easy to run a foul or create a major failure point in your company or your product or what you're delivering. You know, it's it's nice to use a coding standard. We communicate more efficiently, and we can operate better. It's nice to have a law for how we're gonna work as a culture community and society because we can hold expectations and operate with confidence.

Fran Berman: Yeah, I mean I mean and and I think we would agree that that's ideally. Not all laws are good laws, not all laws are well thought out, not all laws are, you know, enforced. On the other hand, to me that's the aspiration to have the public sector intervene in a way that makes us all safer, and to have the private sector really use more holistic and agile design approaches.

Justin Beals: Yeah. You have a great chapter on governing AI in your book. I want to lean into that before we run out of time here a little bit. And one of the areas that you identify is you draw a line between between the decisions we can hand off to AI and the ones that should stay, you know, with a person. You to quote you, “We increasingly live in a hybridized society when it with intelligence machines combined with our own intelligence”.

How do you think about the delineating the differences today?

Fran Berman: It's that's a really interesting question and I you know, I think we're in a global experiment to try to figure out how one might do that.

You know, AI is such a great tool and we can use it in in so many really good ways. But when we live in a society where AI makes some decisions and human makes others, governance gets really, really tricky. interesting thing about any hybrid society is that the only accountable entities are humans. So, you know, machines are not automated systems are not accountable for what they do. And so, humans are going to have to carry the water no matter what. And so we better have a governance structure that actually promotes human welfare as a first priority. So I think we're seeing a lot of nascent experiments around the world. And something I'm keeping my eye on is the AI Act in Europe. And so this is an act that is, and the Europeans really seem to be leaders in a lot of these digital rights legislations, and it's interesting to see how it all plays out. 

So what they're doing about AI is, you know, after a lot of consultation, and that's true for all of these digital rights things, they're rolling out this year and next year the AI Act which couples risk and scrutiny. So to me that's just the most pragmatic and right way to look at things. You can't say, you know, this particular product or this particular company is doing the wrong thing.That's just not holding water. 

But what you can do is set up the standard. And the standard is if things are going to be really harmful to humans, we want to make sure that they're okay. We want to audit them. We want there to be transparency about where the data comes from and what they do. We want there to be guardrails. If somehow they're harmful to people or they're not in compliance with the things we asked for, we're going to give them big giant fines.

So, the AI Act has a whole bunch of structured categories about high-risk AI and unacceptable AI and low-risk AI. And if you're high-risk AI, so think about things that could really do people harm in terms of their opportunities or their physical and mental health. You know, a hiring system, an employment system, you know, other kinds of systems like that. They expect a lot of scrutiny, auditing. Self-driving cars are, you know, in that category as well. Risk management, etcetera.

For low-risk things, a video game, Netflix recommendations. Who cares if Netflix recommends something that I don't particularly like? You don't need a lot of scrutiny. And so there is this kind of graduated sense about where transparency is needed, where a lot of information is needed, you know, all those kinds of things. I'm really watching that. I would really like to see how that will all play out and whether that will create a legal structure that we can we can use to kind of get into this brave new world of humans and AI governance.

Justin Beals:  Yeah. It really resonates with me. At one point in my career, I was helping build an application that would recommend the the best fit candidates for a job. And thankfully the Equal Employment Opportunity Commission had published materials about algorithmic hiring for us, and they gave some very specific requirements around overfitting of our models and the accuracy of those models.

And also how the models worked with protected classes. And we built a testing regime for every model we built where we tested against those requirements that threw out the models that failed. And I was so glad to have a standard, you know, to be told, you know, how to operate ethically in the environment and be innovative and fight for that outcome, but not hurt people. And I'm very grateful that the our public leaders produced that law that we needed to adhere to. Yeah.

Fran Berman:  Yeah. Yeah, I agree. I think that's really important. And you know, those people that are harmed by these technologies, you know, they're you, they're me, there are neighbors, there are kids, you know, there are parents, you know, who press the the evite link and you know, all hell broke loose. I mean, it's just problematic. You know, can we make it a little bit safer? I think that's a good goal.

Justin Beals: 
Yeah. Yeah, absolutely. Well, I wanted to ask one last question of you. especially your expertise running a supercomputer center. It's a massive amount of data center construction going on today for some sort of promise of AI, these large LLMs actually being viable businesses in the future. you know, I'm very curious your take on where that is, having operate deep in the tech stack and the development of this infrastructure in the US. Yeah.

Fran Berman: Yeah, and let me say there's one difference between the data centers that are being used for AI and the supercomputer center is: our job was not to serve a large language model, our job was to serve thousands of users who are scientists all across America. We served oceanographers and engineers and art historians and computer scientists and all kinds of people. And so for us, we got to develop the supercomputer center with the human beings, with the guardrails, with the kind of infrastructure that really helped people get to where they needed to go. One of the things that I'm hearing a lot in the public these days about the data centers that are being used for AI is they are not happy with the societal impacts of them. They're not happy with the use of resources; they're not happy with the jobs that are being taken by AI in the general, in the general public; they are not happy with having them in their backyards and having their own energy bills go up. And so here where I live in western Massachusetts, we've had a number of communities say, " Not happening here”. And that's been really an interesting sort of social pushback for me. 

I think people in general are really worried about the enormous resource usage of these large language models. And I think that's not an unreasonable thing. I mean, environmental sustainability is in the public interest. And I think in general we have to think about when we think about these really wonderful tools, what does it take to run the tool? What is the best use of the tool? When does it need human supervision? 

What are things we don't use the  we don't need the tool for? You don't need AI in your toothbrush. I know there are AI-driven toothbrushes. You don't need it. You know, where can we use these tools of technology to make society a better place, to advance society? And so I think when we think about the environmental impacts of these things, I think that's really hitting people where they live.



Justin Beals: You know, I absolutely agree with you. And I'm gonna combine that with especially the massive LLM companies that are out there; their business models are not proven out yet. You know, they're not yet profitable. I'm not sure that they have a path to be there, but they're sucking up resources as if they will be, but it's still pretty magical thinking. I get a lot of money is changing hands, but I  might be a little old school, but I think great companies make more money than they spend, and they're not quite there yet. They're billions of dollars away. Yeah.

Fran Berman: 
Yeah. Yeah. I mean, that's, I think, where we are on the hype curve is a dangerous place for
kind of a well-functioning society. And I've been thinking a lot myself about, you know, we're going to, we're really aspiring to a very high efficiency place. But high efficiency and quality of life are not the same thing. And I think there's a lot of people concerned that this sort of drive towards efficiency is really diminishing quality of life. And that's where we as a society we get a
say in that. We must have a say in that.

Justin Beals:  Absolutely, Fran. 
Fran, I'm so grateful for the time you spent with us today for your book and your leadership and being someone that I really respect and look up to as a computer scientist. Thank you so much for joining us and our our listeners today.

Fran Berman: 
Thank you so much for having me. I really appreciate it.

Justin Beals:
All right. Another great secure talk to all our listeners. We'll be back with another guest soon.