You're listening to On Shifting Ground from Commonwealth Club World Affairs and KQED. I'm Ray Suarez. In May, Senate Majority Leader Chuck Schumer presented a sprawling roadmap for regulating artificial intelligence. And throughout 2024, California has been working on a new bill, SB 1047, to prevent AI from causing critical harms to humanity, harms like the creation of weapons of mass destruction or cyber attacks that would cripple commerce. SB 1047's rules would only apply to the world's largest AI models, like OpenAI. At the time of this recording, we're just a few days before California's state assembly will decide once and for all whether the bill will become law. But either way, SB 1047 has already sparked a debate in Silicon Valley, with tech leaders like Elon Musk coming out in support of the legislation, while others are saying it's too early to regulate AI's evolution. OpenAI chief strategy officer Jason Kwan says California's AI safety bill could slow progress and force companies to leave the state.
Execs at many tech giants want to assure us, the American public, that we should trust them. But we've seen that social media platforms didn't have a handle on threats to democracy, and the early results indicate AI may be next. Companies like Meta and Alphabet built ethics and safety teams to keep online misinformation in check, but just as AI started to boom last year, tens of thousands of those roles were cut. In full disclosure, I was one of them. For four years, Josh Lawson was the head of electoral risk at Meta. In many ways, civil society has benefited because a lot of us have gone out with tech expertise, with knowledge of how things actually work on the inside, and we're able to make more actionable recommendations on what the platform should be doing or what the policymakers should be writing. Nowadays, he's the director of AI and democracy at the Aspen Institute. He spends a lot of time talking to integrity teams about issues like local voter suppression, deep fakes, and all the other ways AI could sway an election's outcome.
Now, I've been around tech long enough to understand that the integrity teams are often aligned in the right place, but that they have their work cut out for them whenever they go back inside the company and they have to get the right decision maker to invest the right way. Josh's experience at Meta doesn't make him distrustful of other tech superpowers like OpenAI or NVIDIA. In a lot of ways, he empathizes with their position. It can't just be drumming them over the head all the time because then they think that it's never going to be good enough, right? And that it's not worth investing because they're never going to get a reward in public opinion. But he also feels they still have a lot to prove. The public should be insisting on trust and safety investment across the companies because that saw a hit and the companies say that they're going to be able to replace a lot of those things with AI. Until we have proof that works, we should keep pushing all of the major platforms to invest in trust and safety, to not continue laying off their trust and safety teams, and to really move the needle on the way that they think about their role in the world.
And that's how they prove it to us, is by investing in these tools and investing in the people that are enforcing integrity. Josh Lawson, welcome to Unshifting Ground. Thanks for having me, Ray. As we get closer, now we're inside six months. What do citizens have to watch for in this so-called AI election? You know, these scammers are not rocket scientists. They're making fancy spam. And I think that what voters need to understand is that you decide whether that works or not.
And so we need to understand that fancy spam is going to crop up in places that maybe we didn't experience it before. That can look like different misinformation and disinformation in chat threads on WhatsApp. It can look like someone sharing an image that looks right and credible. So expect it where it hasn't been before. But this is something that voters can confront and they can navigate.
Do we have to deal with not just the ability, not just the reach of these new technologies, but the speed at which they're coming at us? That's absolutely the case. We are in a period of transition, a period of adaptation. Over time, we're going to be just fine. But I think it's important that this first generation of us who are going through this really pivotal change in the information ecosystem develop the internal barometer of how we're going to approach truth and factuality. The real risk is that the mere availability of these tools makes us all throw up our hands.
and say the truth's not knowable, and therefore we can only vote based on vibes, based on gut instincts. And that has been a powerful tool, especially in the former Soviet bloc, for authoritarians to take control of democracy. I think you're making a really, really deep point there, and I want to make sure it doesn't just go by the way sometimes broadcasting does. There was something already happening to us before AI was widely available. In a lot of countries, and in the United States, was the truth, factually verifiable truth, already becoming optional instead of imperative.
That's right. So meaningful civic participation, what it actually means to be a democracy, not just to have votes happening on a given day, relies on factuality because citizens have to be informed enough to make vital, crucial direction given decisions for the democracy that they're participating in. And so we were in a period of adaptation. The risk is that bad actors try to contaminate the information environment further so that those of us who do have something to offer the community by way of how we vote, how we think, how we speak, throw up our hands and say, we don't know what to believe. We can't engage with this. It's too messy. But we get a say in whether or not.
That's our future. And we've seen a lot of other countries going that direction. And that's something that we should pass on and say, no, not today. The bad actors need you to succeed. You are the measure of their success. You worked with Meta.
And Mark Zuckerberg has been a frequent guest of members of Congress, compelled to testify about the rapidly changing world of social media. And looking at it strictly from the outside, I assumed that some of the assurances he made to members of Congress were meant to avoid being regulated. Once you got inside, once you crossed the aisle from government to private business, did how you looked at this world change? Did you take as sincere the assurances from tech moguls? to try to fix this without oversight. You know, what changed most whenever I got on the inside was an understanding of how global these companies really are. And that very often when they're dealing with governments, they're not dealing with our government. They're dealing with governments often in much more authoritarian countries where regulation is being used as a cudgel to try to change decision-making at these major platforms. Some cases to create leverage if they have employees in those jurisdictions. I think of Russia where they ended up essentially labeling Meta a terrorist organization.
And we had to work really hard to protect some of our Russian employees. So do I think that there is room for regulation? Absolutely, there is. There are really good use cases. But I think what changed the most whenever I got inside the company is understanding that there is a regulated internet. But regulation isn't always good on its own. There needs to be a democratic type of regulation. We're seeing real steps and gains being made in Europe on that front. And I think U.S. regulators, absolutely, with so many people in our democracy calling for regulation, should be thinking hard about it. But they shouldn't be thinking regulation for regulation's sake. They should be thinking very strategically about how to do this. It's really hard to write smart rules.
For instance, with AI, it's very different than social media. Social media was a platform for content distribution. AI is a probabilistic model that creates different types of content that you may or may not post that may just be for your own benefit. And that's a very different world than something where you have billions of users sharing information, videos, content. And regulation in one instance may look very different than smart regulation in the other. I think the other point that makes it difficult to write smart regulation is that people can assume that if there's regulation, then everything's fine. Reality is that you need both a speed limit and you need a radar gun.
You need to be able to understand that there are going to be people who are going to not care at all that there's a rule in place and going to do the bad stuff anyway. This can be a matter of life and death. Ask any Rohingya who's fled into Bangladesh from Myanmar. Did Facebook during your time there do enough, certainly compared to Google and Twitter, in the 2020 cycle in the United States to rein in political advertising? Did it do enough? Did it give us a preview of what to worry about by not being tougher? Yeah, and I should clarify to make sure I'm not at meta anymore, but I can say during the 2020 election, and I was there during that period, there were substantial.
Thank you. Meta kept political advertising, but they put in place a whole bunch of rules. You couldn't do political advertising if you were not within the country where the political advertising was taking place. There was going to be real strict identity-based requirements that would verify whether or not you are the person that you say you are. And there was labeling so people understood that these were political ads and research tools so that researchers could go behind and see who was posting what. But it's always an area that we need to understand is evolving. Bad actors understand that it's evolving, and they push it.
And it's not just about ads. It's all about the organic content, the unpaid stuff, the stuff that goes viral, not because there's dollars behind it, but because there's eyeballs behind it. And that is mostly the playground where bad actors play. And that's where the integrity tools have to be deployed. Looking back at the 2020 election, are there things that you can point to at Meta that really worked? And are there things that you wish you would have done more of?
Yeah, yeah. And this is all just in my own personal capacity, not speaking for the company that I'm not at anymore. But I was really proud of our decision not to recommend political groups. So in Facebook, you can join groups and there were political ones. And for a long time, the algorithm would tee up for you what groups it thought you might like. The problem with that is that algorithms are always just trying to optimize, right? They're just trying to find a shortcut to figure out the best way to play somebody. And that meant that over long term, you could be nudged. You may be like center right, it might push you further right. You may be center left, it might push you further left. And the recommendations process is a way that that could happen. And so prior to the election,
We were able to turn off recommendations for groups that were political and that worked so well that we kept it on. And that ended up being what is currently the case over there. As far as things that I wish that we could do better, the United States benefits tremendously from the fact that these are United States companies. And a lot of us who work there are Americans and understand the world in terms of America. But...
We are global companies. And so, for instance, when Brazil had its presidential election in 2022, and we had a lot of integrity measures in place getting through that election because there was a lot of abuses going on on the platforms, got through, new president was sworn in. And we thought, OK, that's kind of like the United States. We're probably at a place where we don't have to keep such a close eye on it, because once you get the president in place, like in the United States after January 20th, you're in the clear. But we saw that that was not the experience in Brazil. They had their equivalent of January 6th after the president was already sworn seated. And I think that it's important for these companies to understand to continue to build up the teams.
global understanding of where the same tools, same app looks the same one country in another, totally different contexts. Is this like an arms race for every protective or preventive strategy that's either pushed in from the outside or developed inside these companies? Is there going to be somebody who's diabolically trying to figure out how to defeat it and raise the ante? It's an adversarial space, right? For every good guy, there's the bad guy and they're trying to one up each other. And in that race, understand that it's not neutral, that the bad guy doesn't go away. We can decide whether or not we give the good folks tools and they have the capabilities and power to match what the bad guy's doing and then beat what the bad guy's doing. But we don't get to just unilaterally say, well, the bad guys get to stop, please stop, right? You don't plead with the bad guys. You have to beat them. And we beat them right now, mostly through these tech platforms. Regulators have a really hard time catching and enforcing.
important laws against bad actors. You just look at the end of the 2016 election. Okay, so we name a couple people in indictments, many of whom, Russian actors, many of whom will never face trial. We can't extradite them, right? The real solution is on the inside of these companies.
All right, well, I need a little sliver of hope here. Is there also an upside for these AI models and their tremendously increased ability and reach to improve civic engagement, drive turnout, increase voter education, leave us in November and then perhaps again in January with a feeling that we got it right? It's an all of the above. You have increased ability to give you a little bit more of a understanding of a full issue. So you might say, hey, I want to understand things like the Electoral College better. And it won't just be, oh, this is the worst anti-democratic thing ever. And it won't just be, oh, this is great because it's a simmering pot and lets everybody cool off. I think we're going to see tremendous gains across information integrity, reliability of integrity, because AI can help us sort through when bad actors are trying to make something go viral and we can get ahead of it and stop it.
Isn't that the exact opposite of what algorithms do, what you're describing? The algorithms have been working very hard for a very long time to make sure I see a lot of one thing and not so much of another. That's right. So the incentive structures for content distributors like social media for a long time have been engagement. And engagement means that you see the types of stuff that either makes you mad or makes you happy and not a lot in between. And reality is that for an informed citizenry, most of what you get is going to be in between. It's not going to be something that evokes an immediate gut reaction. It's going to be something that requires your actual thought, right? You think about it, molo on it, maybe talk.
to some folks in your community about it. And that's not social media. But here we have an opportunity in a new era to try to create standards internally for information so that you get both sides. The incentives are not the same way. They don't actually win whenever you're engaging all the time. That's actually more expensive. It draws more energy. It draws more compute time. They want to give you the right answer right at the beginning. And there are ways that that can be abused and there's ways that that can be helpful. And we need to make sure that we're pushing that the way that it's helpful is the way that we go. Josh Lawson is the director of AI and Democracy at the Aspen Institute. Thanks for joining us on Shifting Ground. Really appreciate you having me. Thanks.
You're listening to On Shifting Ground, produced by Commonwealth Club World Affairs in San Francisco. After the break, is it too late to put the AI genie back in the bottle? U.S. Congressman Ted Lieu on why regulation is a necessary challenge before November. If you missed any part of this episode or want to catch up on other international stories, download our podcast on Apple, Spotify, or wherever you get your podcasts. This is On Shifting Ground. I'm Ray Suarez.
In mid-May, Senate Majority Leader Chuck Schumer presented a sprawling 31-page roadmap for regulating artificial intelligence. The report called for $32 billion in spending to put guardrails on the rapidly evolving technology to protect American jobs, safeguard our elections, and ensure our national security against threats from China and Russia. But the criticism of the report came swiftly. Evan Greer, director of the advocacy group Fight for the Future, said the roadmap leads to a dead end. He called it pathetic.
For many tech experts, regulating AI may be impossible because we don't know how it'll evolve and what it's fully capable of. So why is Washington pushing so hard to contain it? Ted Lieu joins us now. He represents the 36th District of California in the United States Congress. He's a co-chair of the Bipartisan Task Force on Artificial Intelligence and a member of the Subcommittee on Courts, Intellectual Property, and the Internet. So you could say AI is definitely in his wheelhouse. Congressman Lieu, welcome to Unshifting Ground. Thank you, Ray. Honored to be on your show. We went from zero to 100 on AI. It wasn't being talked about at all. And then suddenly the term was ubiquitous without really putting much meat on the bones. The big...
forces in this new industry say, oh, don't regulate us, congressmen. We'll police ourselves and we'll keep churning out these great things that make learning and human adaptation better and easier and faster and cheaper. They've used that argument a lot in a lot of technological achievements over the last couple of decades. What's the answer from Congress? When they say, don't regulate us, you're going to stifle innovation. It depends completely on a use case, in my view. AI is used in millions of applications. And I think it is sort of meaningless to say, are you going to regulate AI or not regulate AI? A much better question would be, in what specific use case do you think we should regulate AI? So look, if your smart toaster has a preference for wheat toast over bagels, we don't care, right? But if you're using AI system for hiring people and discriminates based on race or gender or some other particular class, then we do care.
And also, it's sort of interesting, you made a statement about folks thinking, oh, we don't like this. And I guess what I would say is it doesn't matter because you won't have a choice. It's coming. Because you're not going to, the way 99% of people can experience AI, it's going to come to them in a product. And you won't have a choice to say, I want this product without AI. That won't exist. You're going to get a product and either the product is going to be helpful to you or it's not. You're not really going to care if it's powered by AI. So, for example, I don't think anyone cares that the navigation system is partly powered by AI. If they find it useful, they're going to use it. If they don't, they're not going to use it. But should there be?
transparency. If you are a consumer who wants to know, am I interacting with learning machines? Should you be able to find out? Yes. And the thing I want folks to know is really when people talk about AI, they're talking about large language models. And what people need to understand is that
It is never, ever 100% correct. You can, for example, on ChatGPT or Microsoft's Bing or Google's Gemini, you can ask it a question and it can give you an entire paragraph that is pleasing, maybe even beautifully written, grammatically correct, that is totally bonkers false. There's a term that industry uses for this. It's called hallucination. You cannot...
stop hallucination from happening. And so you can imagine use cases where if you're creating art or maybe let's say you're asking a question, you want to submit an essay. Okay, if it makes an error here and there, we can live with that. But you cannot have AI hallucinate if you're having a field medical prescriptions. It cannot every now and then give the wrong dosage or put in the wrong drug because it just feels like doing that. Controlling nuclear weapons, right? We can never, ever let an AI system ever hallucinate. So I have a bill that requires a human in a loop for the launch of any nuclear weapon. It's never going to let AI launch a nuclear weapon by itself. And so I think something that we might want to think about is...
What cases are there where we might want to require a human in a loop if AI is going to be used? One of the things that was an occasional problem for those of us covering this world was that the people who were the gatekeepers, members of the United States Senate and the House of Representatives, often seemed very soft on the aspect of technology that they were talking about. not fully briefed, let's say, on the possibilities of this thing. And here they were, they were interviewing the Zuckerbergs and the Jack Dorseys of the world. And I thought, I'm not sure I want that person being my custodian on this one because they seem to not fully understand this technology. Recently, Senator Schumer from New York set out what they called a roadmap for moving ahead in this area. Is it a good place to start? And are you satisfied with what's in that roadmap?
Those are great questions. My colleagues elected me as vice chair of the House Democratic Caucus, and I've convened a number of briefings to have members come in and learn about AI in different fields, such as civil rights, such as the labor force and so on. And so we are trying to educate members on the technology, what it can do, what are some potential risks, what are some potential benefits. I commend the Senate for tackling this issue, and I think the roadmap is a great first step. It is somewhat general, and I think that was not surprising because it's really hard to go from...
trying to mitigate a problem or a risk, and then actually coming up with legal language in a law that would address that risk. And it is one thing to say, hey, we think this could be a problem. It's quite another to then write an act of Congress. And we do have to figure out how to bridge that gap. Well, he's talking about, Senator Schumer, that is, bills hitting the floor in the next Congress. Is he being a little too optimistic there in your view?
No, I think the next Congress is definitely durable. I think we're going to see hopefully some laws being passed this Congress. So I don't think that we're going to have some huge 5,000-page bill that covers 67 AI topics and say we're done with it. I think it's going to be much more iterative. It'll be over many years. I think you'll have bills every year that try to address discrete areas in AI. I think that's how it's going to play out. The leaders of the big companies in this space are not of one mind about regulation. Some, anticipating public blowback and anticipating that if they step wrong, they will be heavily regulated, have said yes.
write us some guidelines and we will work with you. While others have said, absolutely hands off. Have we learned some lessons that give us some guidance going forward about how to handle this new thing? Yes. So...
First of all, I think there is bipartisan support for trying to mitigate the risks of AI. And I think many members of Congress and senators are not pleased with how the whole social media thing turned out. I don't think any of us want to replicate that again. So I think people want to be ahead of this and try to mitigate the risk of AI before it more often is something that we're unable to control or to stop and it does really bad things. At the same time, it is, again, very dependent on a use case. So if you just sort of tell me, hey, we should regulate this AI model, I need to regulate it to prevent it from doing what exactly? And I think...
It's sort of meaningless when people from either industry or the private sector or civil society say, we should regulate AI. I mean, that's great, but you need to tell me exactly what is the use case that you want me to regulate and what is the harm you're trying to have me prevent. And so just sort of generally say, well, just regulate it. I'm like, okay, that's great, but we do need some more specifics. In the visual sphere, when you ask an AI program to come up with an image and you describe the image in language and it comes up with an image, we have to remember an AI program is dumb. It didn't find anything out. It didn't learn anything. It went and scraped the internet.
for an already existing artistry and already existing products of the human mind to come up with something that they sort of quickly run down the aisles of a grocery store and just shoplift off the shelves. I wonder if there's sort of a moral dimension to this that we're so rushing headlong into this world that we haven't completely considered. That's a great point. AI can be incredibly intelligent and incredibly stupid simultaneously. And one thing to know is that when AI makes mistakes, there are going to tend to be mistakes that a human being would never make. So there's this famous experiment where AI researchers were looking to see if they could tell from simply having AI look at an X-ray scan whether the person had tuberculosis.
So when they did the program and the research, it all seemed to be working well. They deployed it and it failed spectacularly. And they went back and they figured out that what the AI was actually doing is simply recognizing was looking at a scan from a newer x-ray machine or an older one. It turned out that people who had scans from newer x-ray machines tend to be in richer areas, wealthier countries, places where there was very little tuberculosis. And the folks that were scanned by older x-ray machines tend to be in areas where there was some tuberculosis.
So the reason that that AI failed is, again, two things that people should know. One is not only does AI every now and then hallucinate, second is we don't really know what it's doing. The first word in the name of your AI task force is bipartisan. Is there, generally speaking, a different attitude or different shading of attitude in the two parties when they look at this new area of technology? I'm encouraged that the task force has been quite bipartisan in our meetings and in the questions that members ask. I think members of both sides of the aisle are trying to get information, learn about some of the subjects that they may not be as familiar with and really listening to their experts and witnesses. I think another way to think about AI is it is a tool. Tools are not partisan. And so I think we can get bipartisan consensus in a number of areas to address potential bad things that this tool could be used for.
You know, a couple of months ago, you wrote an op-ed in the New York Times, and I know that you don't write the headlines, but the headline was, I think, pretty good. I'm a congressman who codes. AI freaks me out. How do we bring to this the proper level of concern and care and caution while still allowing the tremendous possibilities of this technology to go forward? Is the sweet spot in something that's sometimes a blunt instrument like legislation possible to find?
So let me tell you how I view that question. To me, the best analogy is two bodies of water. One is a large ocean of AI. The other is a small lake of AI. So in this large ocean of AI, it's going to be all the AI we don't care about. So if your smart refrigerator messes up making ice cubes, we don't care about that. In a small lake of AI is AI we might want to care about. And to me, there's at least three buckets. The first is AI that can kill you. So again, we can never have AI hallucinate if it's controlling nuclear weapons. We don't really want these frontier models that if they get very, very advanced to be able to very easily let people make.
a weapon mass destruction or to help you all to create deadly pathogens and give them step-by-step directions on how to do it. We need to have laws and regulations on AI that can literally destroy the world or kill lots of people. The second bucket to me is AI that isn't going to destroy the world but could kill you individually. So when your cell phone malfunctions, it's not going 40 miles per hour. It turns out there's a lot of AI moving objects, planes, trains, automobiles. And I think we need many more regulators who are trained in the unique aspects of AI, folks who understand that AI can never be 100% correct if it's the large language model, and folks who...
might know how to use regulations to better protect the public. And the last bucket is sort of the hardest, which is AI that's not going to kill you or destroy the world, but could still have negative effects in society, such as bias in facial recognition systems or in credit risk systems. And there's also, I think, a timeliness to certain issues. So clearly elections are coming, deep fakes are a problem. And I think it'd be great if we could try to mitigate it with some laws or regulations. But if we don't, then I think the best thing to tell the public is don't trust everything you see on the internet.
You introduced the first ever piece of federal legislation written by AI. What was it? How did it work? Were you impressed with the work product? And could we replace you with one of these products one day? So it was impressive in the sense that I wrote a prompt. Now, it did take me several tries writing different prompts for this at work. But I finally did write a prompt where the AI algorithm generated essentially a bill that I submitted without edits. Now, at the same time, I have to tell you, it was pretty mediocre. So we're not quite there yet.
But we're also at the beginning stages of this. What does ChatGPT version 10 look like five, six years from now? I have no idea. To me, it's both amazing and alarming at the same time. Well, you know how hard it is to write code and...
People who do it say there's a lot of drudgery involved. And AI's utopian prognosticators say, hey, we are going to free you from the drudgery of repetitive intellectual work, repetitive physical work. You'll see. When we move as far as we can go with these things, you'll have more free time. You'll spend less time on nonsense and more time on the core of the work that you do. Don't worry. Are they right?
I think you're right in the sense that both technology and AI will replace more and more human tasks that are repetitive. Like automation has already replaced a number of repetitive human tasks. And with AI, you're going to start replacing repetitive human tasks that require some cognition. Now, should we worry? Well, yeah, I think people should worry because it's going to have disruptions in the workforce. But it's also very unclear where that's going to be. The actual Ted Lieu, not an AI-generated...
program or an avatar of some sort is sitting across the table from me in Washington, D.C. He's co-chair of the bipartisan task force on artificial intelligence. Thanks for joining us on Shifting Ground. Thank you, Ray. Almost a quarter century ago, Bill Joy, the co-founder of Sun Microsystems, wrote an essay that got people talking. The title, Why the Future Doesn't Need Us. He talked about learning adaptive self-guided technology as one of the dangers of the 21st century, which was at that point just a few weeks old. He warned us that we ought to be careful about inventing a future that operates without human interaction, that technology was moving really fast, and we ought to be mindful about what we set free into the marketplace. I've thought about Joy's warning a lot in recent months.
as artificial intelligence, AI, zoomed from the fringes of the national consciousness to something closer to the center, and people rhapsodized and tore their hair out over its possibilities. It is in my nature to wonder if both the utopians and the doomsayers are wrong. Dr. Gary Marcus joins us now. He's a scientist, a writer, and a serial entrepreneur, the founder of Robust AI and Geometric AI, deeply immersed in understanding the promise and limitations of artificial intelligence. Welcome to Unshifting Ground. Thanks for having me. Did all of this move faster than even its biggest boosters expected? You know, it just feels like things are ramping up.
One thing moved really fast and another didn't. So on the technical side, we still have major problems. There are these head-scratching errors that these systems make. But what has moved faster is the public adoption of this stuff. Nobody in the field really thought that 200 million people were going to be using these things we call large language models. OpenAI didn't invent them. If anybody did, Google did. And Google had them. Facebook or Meta had them. OpenAI had them. And everybody was kind of playing around with them. Nobody thought they were going to take off and become this public play thing. It was a very good media campaign. They made them feel human-like by typing words out one at a time. And people just got hooked. Nobody saw that coming.
You are called a critic of this technology. Do you gladly wear that label? And what are the cautions that you would have as we move ahead? Well, first of all, I would say skeptic is maybe a better word than critic. And I would say that...
I actually love AI. I want AI to succeed. I've been working on it since I was a kid. So I'm not a critic of the notion of AI, but I'm deeply skeptical that the current technology is the right one. I think that the current technology is morally and technically inadequate. So it is morally inadequate because these systems are biased. They don't know what they're talking about. They can do harmful things. They can easily be misused for misinformation and deep fakes and so forth. And they're technically inadequate because they lie. They're unreliable. Actually, I shouldn't say that they lie because they don't know what they're talking about. You never know what you're getting. So on a technical side, they're very unreliable tools. So I think that the current AI we have is really problematic. And to the extent that a lot of people have fallen in love with them, I'm saying, hey,
Let's look more closely. They're not really doing the things that you want. And I also don't like the power structures around it. It's almost like the ring in The Lord of the Rings. People get close to AI or think they get close to AI and to money and the power, and it corrupts them. And we've seen a lot of that. People that talk about responsible AI suddenly are a whole lot less responsible. I got to tell you, I'm a little reassured every time a third hand pops into an ordered image or a completely ludicrous sentence turn up in a language-based AI prompt answer. I'm wondering, though, whether that's just a question of time.
Is it just until the speed and comprehensiveness can ramp up in such a way that we won't be projects, shall we say, some morally responsible and some less so, that you can imagine with AI. So AI is not magic. It's a bunch of tools, a bunch of different technologies. Some of them work well for some problems, some for others. Some of them don't really work reliably at all. One of those sets of tools is making deep fakes of other people doing stuff. Those tools are unfortunately already really good. Another set of tools would be AI that can solve physics or solve medicine and figure out how to cure Alzheimer's or something like that. We don't really know how to do that yet. So there are different pieces of AI that have different trajectories. Unfortunately, the kind of negative uses of AI are moving faster than the positive ones.
At least to my eyes, that's the case. And I think that's because the misuses of AI often don't require reliability. Whereas if you're building a new drug, you have to be super reliable. So that actor can create misinformation with these tools. And I don't care that a lot of what it says isn't true. That's almost like part of the mission.
The bad actors don't really care that the stuff's not that reliable. And the good actors are like, well, this doesn't really work well enough for me to be doing medical diagnosis yet. So there's actually an asymmetry right now. And in 2060, we will look back the way we, in 2024, look back at people who had flip phones. And we're like, yeah, flip phones were kind of fun, but they had to do that crazy T9 stuff to do texting. And what a pain. I can't believe that they survived that stuff and it would a mess. We're going to look back at 2024 AI and be like, yeah, they understood how cool AI could be, but they didn't really have it working yet. They were kidding themselves. When you appeared before the Senate to talk to them about the...
possibilities and problems of AI. What did you tell them? What did you want them to take account of? Well, I remember one thing in particular, as I was writing my remarks, I had this sentence, I said, they're going to use these tools to manipulate our elections and our markets. And I was like, I have good examples for the election stuff, but I don't have anything for the markets. Can I really say this? And it was a week later that people did actually start manipulating markets with AI tools. There was a deep fake picture of the Pentagon on fire and the market fell for a few minutes, like a bunch. And I don't know if that was deliberately manipulating the markets, but I'm sure other people are like, well, I could short the market if I use propaganda generated by these tools. So what I told the Senate is we need to get cracking. These things are moving fast and they are going to be used for disinformation campaigns. And I said, they're going to be used for cybercrime and so forth.
is what's transpired. That was just over a year ago. And what makes me sad is in the room, there was very strong bipartisan support for a lot of the things I was talking about, like having FDA approval process before you roll out a really large model that might have risks for society. I talked about having global AI governance and so forth. The UN has picked up on that. They're trying to do that. The United States Senate has not done anything. The United States Senate, even though there was bipartisan agreement in that room that we need to move fast, that we need to do kinds of things I was talking about, have not actually put anything forward. Individual senators have, but Chuck Schumer, who is the Senate majority leader.
spent nine months or so talking about AI and so forth, only to release a white paper with no actual legislation. And unfortunately, what's going to happen is the Senate's about to go into recess. And when it comes back, everybody's going to be focused on the election. And then who knows what happens next year? Everything will be different one way or another. And so there was an opportunity there with strong bipartisan support to really do something of substance. And it didn't happen. And I find that deeply disappointing. I actually just started and wrote a book because of this. It'll come out in September called Taming Silicon Valley. And part of the point of the book is that...
Congress is not doing what it should be doing. The choices we make in the next couple of years around AI policy are going to have lasting impact, just like the choices we made or didn't make around social media did. And this could be a lot worse. And so I wrote this book to get people to stand up and go to their senators, go to their representatives and say, look, this is really important. We really need to have an AI policy that works for the many and not just the few. The administration has laid out a roadmap for eventually regulating AI. Is this something that's actually harder to do than we've fully taken the measure on?
Well, first thing I'd say is I think the Biden administration has done everything that it can do or almost everything that it can do. I think that they've taken the threat seriously. They've tried to do what they can within the existing structures. But you also have to understand that they're the executive branch and not the legislative branch. And so if they tried to make a law in their executive order, it would get shot down instantly, right? Because it's just not how our country works and it'd be against the constitution for them to do so. So they have done what they can within their power. Unfortunately, the Trump candidacy has threatened to take it all down. If Trump wins, then we're really going to be left vulnerable, I think, to whatever might happen here.
But then really it's up to the Congress because that's how our constitution works is the executive branch does not make law. The legislative branch has to make the law and so forth. They mostly haven't. What are some of the bottom lines for any attempt to regulate this industry? What do you think is necessary to be in there that both protects us from some of the worst possibilities of this technology, but also understands that you can't anticipate every eventuality and you don't want to stifle what could be a tremendously productive industry?
We need an independent agency to make these decisions. And this was very much appreciated in the room when I was there with the Senate. One of the senators, I believe it was Durbin himself, said, we move slowly. You know, our job is not to make decisions in the moment. It's to make lasting decisions, which is why we have agencies like the Food and Drug Agency or the FAA and so forth, right? The FAA can move faster than Congress. You don't want Congress to sit there wondering about a new flight design. You want an agency that has experts that understand flight to make those decisions and to be able to make them quickly. I think an agency, both for the positive uses of AI, the agency should be sending out memos saying, hey, you know, we can do this thing as a federal government that we're not to save money.
move faster, et cetera, and also to issue warnings and so forth. Because, you know, if you have an AI that's rolled out to 100 million people or 300 million people, it can have an effect on the world. It can, for example, bias how people vote in elections, doing it in subtle ways that nobody notices. So you need to have some kind of regulation and really layers of regulation to look at what's going on. Then we need auditing after the fact. So one example is the chat GPT is probably evaluating people for jobs. And we know that it will do that in racist ways to some degree. And there needs to be a way of us investigating that, looking at the logs. How have people actually used it? What are the results?
I want to go back to what you touched on briefly, and that is the international scene. Regulation by Congress has a writ that runs in the 50 states and American territories, yet the internet still aspires to be borderless. It may be less so than it was in an earlier age. How much does a United States Congress member have to say about how this technology operates across the planet? I mean, ultimately, I think we need some global treaties for various reasons. And Congress plays a role in there being a global treaty, but it's a global process and they're complicated. But I'll give you a couple of reasons why we should want to do that. One is...
As the AI gets smarter, and I don't think it's very smart right now, but it will be someday, whether that's in three years or 50 years or somewhere in between, there are risks of runaway AI, sort of like Terminator scenarios. I don't really think AI is going to kill us all, but it is easy to see how AI could cause catastrophes of different sorts, maybe not literally extinction. Some people talk about we need to have information sharing between different countries and ways of pulling a plug if we need to pull the plug and so on. So we need treaties around that. And then secondarily, but important, these systems, we haven't talked about this, but they use up an enormous amount of energy and people want to make them bigger. And so they're going to use even more energy. So imagine you have these things like that take like the amount of.
energy that the state of New Jersey takes for a certain number of hours or whatever. And you have to do that differently everywhere you deploy a model. So you have 193 countries. You have like 10 companies doing this. That's already 1930 models. It's an enormous amount of energy. So you would really actually like international standards. So you can train your model once and you can use it everywhere rather than having to do it 193 times and multiply out those costs. It's not really in the interest of the company. It's not in the interest of the environment. We really need that as well. So I hope that the US Senate and House of Representatives will be supportive of our efforts to do that. I'm glad to see that the UN is taking steps towards having some kind of international AI governance. But any of these things obviously takes a long time to work out.
Countries don't want to give up their sovereignty even a little tiny bit. Sometimes they have to. Is there a tension between the desire for the people who are inventing this future to derive profit from it and own the intellectual property around it and the tendency of these systems to go out and take... other people's intellectual property and repurpose it. Oh, absolutely. The rules for thee and not for me. I mean, you know, all of these companies are basically stealing artists and writers and so forth blind. I mean, at least OpenAI is starting to do some licensing deals, but on the whole, massive numbers of writers, artists, actors, and so forth have been ripped off. ScarJo was right, I think, to protest this, and I'm really glad she did. And at the same time, you have places like MidJourney went after Stable Diffusion for using MidJourney's data. It's like MidJourney is taking everybody's data, but don't take data from us. It's outrageous. Dr. Marcus, before we go, could you touch on what rank-and-file citizens?
should be careful about between now and November and between now and January? Well, first thing is everybody has to realize that misinformation is going to come fast and furious. You have to put your critical thinking skills on. Just because you see something in a video or read something does not mean it's true. You have to cross-reference. You have to look up and see whether anybody has commented on this. Is this thing you're seeing really true? Be very aware that people are going to try to manipulate you with these tools. I'd say that's number one. Number two is...
care about the writers and the artists and so forth that are getting ripped off. You know, right now they're coming for the artists and the writers, some of whom might be you or your friends. Eventually they're coming for everybody. Remember, a lot of companies now log every keystroke that you type. They all want to automate everything. And so you should really stand in solidarity and say, I don't want to use stuff if you're not licensing it properly. If you're just going to take what you need from these folks, I'm going to stand with them. And so I hope we'll actually consider things like boycotts of companies that don't get their act together. And more generally, you have to ask yourself.
Who are the people I'm voting for? What are they going to do about technology policy? Do they care about me or are they just in bed with the technology companies? And I hope people will take that seriously as they make their votes because like what we saw with social media is there was a chance to do something and we didn't. And now we have lots of problems with teenagers and depression that are perhaps tied to social media. We have lots of polarization and public discourse. Like we didn't really get it right the first time and then things get entrenched. We really need to think in the 2024 elections, can we as citizens try to nudge things so that our rights are protected? So we have an AI that works for the masses instead of just one that lines the pocket of some surveillance capitalists in Silicon Valley.
Dr. Gary Marcus, thanks for joining us on Shifting Ground. Thanks a lot for having me. You've been listening to On Shifting Ground, produced in partnership with KQED, with funding from listeners like you, with special support from Mark and Maury Jane Perry, and Anne and Greg Avis. Today's episode was produced and mixed and mastered by Mateo Schimpf. Additional production and engineering were provided by Rob Spade. KQED's Jim Bennett is our technical supervisor. Jared Sport is our executive producer. Philip Yun is co-CEO of Commonwealth Club World Affairs. Our music is from Blue Dot Sessions. I'm Ray Suarez. Thanks for listening.