Science Friday - ChatGPT And The Future Of AI, Turkey Earthquakes. February 10, 2023, Part 1

Episode Date: February 10, 2023

How Scientists Predict Where Earthquakes Will Strike Next The pair of earthquakes that hit Turkey and Syria this week left the region grappling with death and destruction. Despite the region being seismically active, this particular area hadn’t seen an earthquake of this size for decades. There are ways of knowing where the next big earthquakes will happen—but not when. Scientists use knowledge of fault lines and historical data to make their predictions, but saving areas from mass casualties often relies on infrastructure policies. Building codes that prioritize strong buildings can save lives, but older structures remain vulnerable. Across the globe, in California, the health impacts of electric vehicles are beginning to be seen. A study published this month finds that for every 20 EVs in a zip code, asthma-related visits to the emergency room drop by 3.2%. This is a striking number for a technology that’s just now becoming more commonplace. Joining Ira to talk about these stories and more is Umair Irfan, staff writer at Vox, based in Washington, D.C.   ChatGPT And Beyond: What’s Behind The AI Boom? The past few months have seen a flurry of new, easy-to-use tools driven by artificial intelligence. It’s getting harder to tell what’s been created by a human: Programs like ChatGPT can construct believable written text, apps like Lensa can generate stylized avatars, while other developments can make pretty believable audio and video deep fakes. Just this week, Google unveiled a new AI-driven chatbot called Bard, and Microsoft announced plans to incorporate ChatGPT within their search engine Bing. What is this new generation of AI good at, and where does it fall short? Ira talks about the state of generative AI and takes listener calls with Dr. Melanie Mitchell, professor at the Santa Fe Institute and author of the book, Artificial Intelligence: A Guide for Thinking Humans. They are joined by Dr. Rumman Chowdhury, founder and CEO of Parity Consulting and responsible AI fellow at the Berkman Klein Center at Harvard University.   Transcripts for each segment will be available the week after the show airs on sciencefriday.com.   Subscribe to this podcast. Plus, to stay updated on all things science, sign up for Science Friday's newsletters.

This is Science Friday. I'm Ira Flato. A bit later in the hour, we're going to be talking about the artificial intelligence boom, you know, tools like chat GPT that can create believable human language in seconds. We'll be answering your questions about it. Our number 844-8255-844-Sai Talk, or you can tweet us at Tsai Fry. But first, that pair of earthquakes that hit Turkey and Syria this week left the region. devastated. And despite being seismically active, these are the largest earthquakes the region has experienced in decades. There are ways to know where the next big earthquake might happen,

but not when. Joining me today to talk about the science of this story and others from the week is Umer Irfan, science writer Ed Vox, based in Washington. Welcome back to the show. Thanks for having me back. Nice to have you. Okay, let's talk about these earthquakes. Was there any warning that they were going to happen? In the short term, no. There wasn't really any sign. People in the region didn't get any kinds of alerts when this happened.

And the earthquake struck very early in the morning. So a lot of people were asleep at home when this happened. And that's part of why the devastating toll has been so high. But this is a region that's known for being seismically active. There's actually two major fault lines that run through Turkey. And this is an area that historically has had major earthquakes. But as you noted, it hasn't had a. major earthquake in this specific region for decades.

And so the challenge here is trying to come up with the probability and actually telling people how to respond and prepare for this. And that's a problem that we face all over the world when it comes to earthquake risks. Because we know where the fault lines are. We just don't know when the fault lines are going to break or move. Right. I mean, there are some very early signs in that you can get over very short term. Like, for instance, we know that earthquake waves travel over a period of time.

And sometimes you can send in some parts of the world text message. alerts to people hundreds of miles away, but that only buys you a few minutes. And if you're near the epicenter, you basically have very little to no warning. And that's where you see some of the worst devastation. And the earthquakes were designated 7.8, 7.7.7. Remind us what these numbers mean. Right. So this is a logarithmic scale that measures the intensity of these earthquakes. The old-fashioned way was using something called the Richter scale, which you may have heard of.

Oh, yeah. This is, that was basically a scale that was calibrated to Southern California, where it was developed. And scientists found that actually it wasn't very good at describing earthquakes in other parts of the world. And so what they developed instead was a scale called moment magnitude, which like the Richter scale is also logarithmic, which means that each number going up represents a tenfold increase. So a magnitude seven is ten times more severe than a magnitude six. But what it does is it also captures different kinds of geology and the different kinds of, kinds of waves that can travel. For instance, in very hard rock, earthquakes can travel for a very long distance very quickly, but in softer soils and in softer geology, that can actually

attenuate. That can actually act as like a shock absorber. And so what this scale does is it allows you to make more apples-to-apples comparisons between earthquakes in different parts of the world. Now, this is still not always that useful as a scale for architects and engineers who are designing buildings and trying to build structures to resist earthquakes. They're often more interesting. interested in peak ground acceleration, which basically measures how fast the ground is moving and given earthquake or displacement, which is basically the total amount the ground can shift during an earthquake. And those are sometimes the more relevant ways of measuring the intensity of these kinds of events.

More useful. Let's move on to your next story, which involved this week's State of the Union address. President Biden mentioned cancer 13 times, which I think may have been surprising to a lot of people, and just a little reminder, a bit over 50 years ago. I think, it was last fall, President Nixon declared a war on cancer, and President Biden mentioned two cancer initiatives. Let's talk about those. Right. So, President Biden has had, you know, a personal interest in cancer. His son died of brain cancer back in 2015. And when he was vice president under President Obama, Biden was put in charge of this cancer moonshot initiative. And the idea was that this was going to be a gopher, broke push to try to resolve and deal with cancer, as we know,

it. Now, the reason why he brought it up in the state of the union is that the funding for that particular program is related to run out later this year, and he needs Congress to approve more funding for this. And so the idea is that with more research dollars going into this, that we have a better chance of making cancer a disease that people can live with. They're not necessarily talking about a cure here. The goalpost is that they want to reduce the cancer death rate by 50% over the next 25 years. And so this is not necessarily going to make it, you know, a way to get rid of cancer entirely, but we're talking about making it a more survival,

maybe a chronic illness. And so the other program he was talking about in the state of the union was this thing called ARPA H. This is a subsidiary or a part of the Moonshot program, but this is a research initiative modeled on DARPA, which was the Defense Department's Advanced Research Project Agency. And this was the program that led to the stealth bomber and led to the creation of the internet as we know it.

And the idea is that the government wants to find a very high risk but high reward project. So things that are kind of off the wall, things we don't know will work. And we're going to expect that a lot of these projects will fail. But if one does succeed, we expect immense benefits. The problem is, of course, something like this is going to be very expensive. Congress thus far has funded ARPA-H to about the tune of $2.5 billion last year. But the White House wants to actually triple that money. Yeah, because we're seeing some drugs that are hundreds of thousands of dollars a year for some cancer patients. That's right. We're seeing a lot of diminishing returns with cancer treatments that, first of all, a lot of these drugs are very expensive,

so the people who need them can't, we always get them. And while they are making improvements, we're paying a lot for these very marginal improvements. And so what we really want is a breakthrough that can actually move the needle. Well, let's hope that happens this time around. Let's talk now about electric vehicles. One of my favorite topics, if you're a listener to the show, you know that. You have a new story about how EVs are saving lives. And we're not talking this time about driving accidents. Tell me about what you're talking about. Right. You know, electric vehicles, they don't have an internal combustion engine inside them. They're not burning fuel, so they're not producing all these combustion byproducts, things like nitrogen oxides, carbon monoxide, all these chemicals

that can actually make it difficult to breathe and lead to other kinds of health problems. So it makes sense that you would see improvements in air quality, but what was surprising about this finding from this team of researchers at the University of Southern California was that it didn't take very many EVs to actually start seeing these effects. And so what they did was they looked at real world data in California. You know, California has been ahead of the curve in EV adoption. And what they were doing was they tracked EV penetration across different zip codes in California. And they measure that alongside emergency room visits due to asthma.

And what they found was that for every 20 EVs per 1,000 people, there was a 3.2% reduction in asthma-related ER visits. So it doesn't take many EVs, is what you're saying, for that difference to come. I mean, it kind of signals just how bad air pollution is. But at the same time, though, what they also found was that there was a big discrepancy in the areas that saw the benefits. You know, EVs generally are a little bit more expensive. The average car in the U.S., new car costs about $48,000. The average new EV costs $66,000.

And so wealthier areas, we're seeing the bigger declines in these asthma-related visits. But poorer areas are often the areas that have. worse air quality. And so the dividends of this would actually be better spent in some of the low income areas. And so it kind of signals to policymakers that we need a way to help ensure that those communities are also benefiting from this transition to cleaner vehicles. Well, one way we could hope that these new Biden tax credits for electric car purchases make them more accessible to more people. Yeah, that's something that Biden talked about during the state of the union address as well. Yeah, yeah. Speaking of carbon emissions, let's go to your next story, which is about carbon capture

specifically from smoke stacks. That seems like pretty logical, right? It does. You know, if we're worried about carbon dioxide, why not go straight to the source and capture it there? And that's something that scientists have been trying to do for a very long time, but the process that we have for doing that

is actually fairly expensive and energy intensive. You know, we do CO2 scrubbing inside, you know, it closed environments like submarines and spacecraft. But to do this on a power plant, you know, the big issue is the economics of it. It gets really expensive. So the conventional technique is using these chemicals called amines, and they require reheating the chemical in order to regenerate it.

And that process can actually eat up upward of 30% of power from the power plant. This is called a parasitic load, and that adds to the cost. And so the capture cost ends up being about $200 per ton. And electricity in the United States, you know, it's sold on competitive markets in most of the countries. And so if a power plant were to install this system, you know, it would raise their operating costs quite a bit, and they would be non-competitive. And so the big threshold or the big goalpost here is trying to make this a lot cheaper and a lot more energy efficient.

And a team of researchers at the Pacific Northwest National Laboratory said that they found a way to do that. They can actually now bring the cost roughly to about $40 per metric ton of CO2, by 200. By doing what? Well, what they found were these new types of CO2 binding liquids. So the problem with the conventional amines is that they also tend to absorb a lot of water and you have to use a lot of heat to get rid of that water. and regenerate it. But these CO2 binding liquids that they found don't absorb that much water.

They don't require anywhere near as much heat to reproduce and regenerate. And that means that the overall cost and the efficiency of the system goes down quite a bit, but they still captures about 90% or 97% of the CO2. And critically for policymakers, you know, this is below the social cost of carbon that's been established by the government. The government sets that, that cost at about $51 per ton. So if you can do that at about $40 per ton, then this becomes, you know, sort of a no-brainer if there is a carbon price that's ever imposed. Not just for this government, but for other poorer nations. Right. You know, coal power plants in the U.S. are already, declining, but in much of the world, you know, 80% of the world's energy still comes from

fossil fuels, and a lot of developing countries are still relying on fossil fuels, too. Poverty, so it'll get you there. Speaking of a water method, that goes right into my next door at my wheelhouse about surprising stuff about nature. And this one is about a new type of ice, ice that forms when you shake it really hard. This is really cool, so to speak. Yeah, it's something right out of a science fiction novel. I don't know if you've read Kurt Vonnegut's Cat's Cradle, but that was a big plot element in that story. But yeah, what they found was that if they chilled water, these team of researchers in the UK, they chilled water to minus 320 degrees Fahrenheit, and they shook it in this

container with steel balls. Cocktail shaker is kind of on the right track, but you'd have to be shaking it really vigorously. They were shaking this at about 20 times per seattle. second. And what they found was that it actually created a new form of ice. And what was special about it is that, you know, ice typically forms crystals. And when it forms that crystal structure, it's less dense than water. And that's why ice floats. But when they shook it with these steel balls, what they found was that it actually had roughly the same density as water. So this is ice that doesn't float, but is actually kind of neutrally buoyant. And also, it doesn't form a crystal structure. It doesn't have, it's more like glass than it is like a girl. That's cool.

That is really cool. It's a great way to end your segment. Thank you, Omer, for taking time to be with us today. My pleasure. Thanks for having me. Omer Airfant, science writer at Vox, based in Washington. We have to take a break, and when we come back, we'll be talking about chat, GPT, and the future of AI, and what's going on now with it. And we're taking your calls, our number 844-724-8255-8-4-4-Sight-a-4-Tock or tweet us at SciFri. Stay with us. We'll be right back. Science Friday, I'm my reflato. The past few years, it's been harder and harder to keep the squirrels from my prized tomato plants. It's been, well, nuts. So this year I'm giving up. I'll set dishes of nuts right under the tomatoes and build a ladder to make it easier for the squirrels to reach them. Wait a minute. That wasn't me. I didn't say that. That was actually an AI-generated version of me,

And we did it by feeding audio samples of me into a program called Descript. It's the first time I've heard this. So this is just as surprising to me as it is to you. But it's not quite perfect yet, as you can hear, I hope. I think I still have a job. Well, we'll see for now, at least. Even so, it's getting harder to tell what's human and what's not. There is a growing number of programs popping up where you can create deep fakes of audio and video

to make people appear to say things that they actually have not said, like ChatGPT. Maybe you've tried that out if you have patience enough to wait online to get in, or others like Lenza, and then there's stable diffusion that create images. And just this week, we saw Google unveil their new AI-driven chat bot called Bard. Wonder of Shakespeare is spinning on this. And Microsoft announced that they will be using ChatGPT within their search engine. Bing. And let me ask you, have you used chat GPT, maybe to write a paper or a work assignment? Are you worried about the ethical implications of this new technology? That's what we're going to be

talking about. What direction would you like to see AI apps go? You make the call, only if you make the call. Our number is 844-724-8255-844-Sy-Talk, or you could tweet us at SciFri. Let me introduce my guests, joining me to talk about the current state of what is called generative AI. Dr. Melanie Mitchell, professor at Santa Fe Institute, based in Santa Fe, New Mexico, author of the book Artificial Intelligence, A Guide for Thinking Humans, and Rommand Chowdhoutary, founder and CEO of Parity Consulting and the Responsible AI Fellow at the Berkman Klein Center at Harvard in Cambridge. Welcome both of you to Science.

Friday. Thank you for having you. You're welcome. Dr. Mitchell, let's start off. Let's talk, start with what seems like a basic question that's actually kind of difficult to answer. What is artificial intelligence? What does it mean for a machine or an algorithm, Dr. Mitchell, to have intelligence? Yeah, that is a difficult question to answer because everybody has their own definition. So AI is really getting machines to do things that we believe requires intelligence in some form. So back, back, you know, in the early days of AI, playing chess was an example of something that people thought really required very high-level general intelligence. And yet we were able to get computers to play chess without really using anything like human

intelligence. And now it's gone even further. We're able to get machines to produce language and images and other media in a way that looks very human-like. So, whatever seems to require intelligence at the time, that's what we call AI, is getting it in machines. Yeah, so I would imagine if you ask three different people, they would say three different things of what they think intelligence AI is. Yeah, exactly. There's many different definitions. Can you explain briefly, I hope, how chat GPT or similar chat programs work, almost seems like a bit of magic. You ask it to write a term paper and it spits one out? Yeah, so this is a chat GPT is a kind of what's called language model, which is a program that's learned from vast amounts of human generated language.

And the way it's trained is it's given text, like a sentence and it's asked to predict a missing word. And it's doing that over and over again for huge amounts of data that's been trained on from. websites to digital books to all of Wikipedia and so on. And then now you can give it a prompt, like, you know, write an essay on the causes of the American Civil War, and it will then predict, in some sense, what words should go next over and over again. And will generate something that sounds very, very human-like. You know, listening to that little AI voice at the top of me, it doesn't quite get the pacing or the intonation, right? It's amazing how it got my voice, I think, correctly, but there are lots of others that can do that, right? We just don't have the public

radio budget to pay for a better one. Well, I think also we might need more data from your audio clips. So if it had enough data, it probably could imitate you pretty well. Dr. Chattery, how have publicly available tools like chat chit, which generates convincing language, and then you have lensa convincing images and even programs that can create a fake voice. How have they changed their understanding of what AI is capable of? That's a great question. So first of all, the big shift from, I would say, traditional artificial intelligence to generative AI is that the type of content that is being created by these models doesn't exist anywhere on the internet, right? It's not spitting back something that you see,

like a search engine, it is actually creating this. So when we think about these generative AI models, what I think about as somebody who works in machine learning ethics and AI ethics is what are the kinds of harms and stereotypes that exist in society that these machines can pick up on. So as Melanie said, these models are simply reflecting the data that's being put into them. So all of these AI models are not any source of truth, but they're actually reflecting what people have put into them.

Melanie, are you surprised about how well this works? Yeah, I've been very surprised. You know, I never thought that we could get human-like language generated with a machine that is so, in some sense, unintelligent. But it just goes to show how powerful, huge amounts of data can be and using a very, very large, complicated neural network program to make statistical models of that language. That is amazingly powerful.

Now, one of the big concerns about these AI chatbots is how prone they are to making up facts and other programs that create such realistic people to convince you of the facts, facts which may not be really true. For example, this week when Google unveiled this new chatbot, barred, it claimed that the web telescope was the first to take a picture of an exoplanet,

And of course, we have lots of pictures of them. How big concern is this, Dr. Chattery? It's actually a fairly large concern. There have been issues of misinformation, fake news and fake media for quite some years now. Really, the pivot that ChatGPT and Bard and a lot of the more readily available models do is now anybody can make it very easily. So we do have to be concerned about the type of meeting and how convincing it is. I'll also add that it doesn't take a high level of sophistication.

to convince somebody of misinformation. You may remember the Nancy Pelosi video. Plenty of people still believe it's true. It is not a high-quality deep fake. So the change has not necessarily been in how good the quality is, but how easy it is to make it and get it out there. That is really interesting. Our number 844-724-8255. Let's go to the phones. Winton in New Orleans. Welcome to Science Friday. How you doing? Thank you for taking my call. You're welcome. Go ahead. Yeah, so I am an entertainment attorney.

I work in intellectual property law, and I work with arts and entertainment professionals every day with copyright infringement and trademark infringement. And one of the biggest issues and worries I have surrounding AI generated creative arts is, A, the ethical implications, but also the potential impact they could have, A, on intellectual property law, and B, on the argument community as a whole, as the copyright law. offices already come out, ruled that AI-generated creative works are ineligible for copyright protection. And then we've seen, especially with some of my clients, we've seen them trying to figure out how to bring copyright and trademark infringement suits because they've seen AI-generated creative

works that are eerily close to their work or have fit out their trademarked image within an image that was created, and I'm following really closely the class action lawsuit that was recently filed last month against Mid Journey Stability, AI, and Deviant Art for copyright infringement after the Mid Journey CEO admitted in a Forbes magazine that, yes, it was being trained on hundreds of millions of artists work, and they did not reach out to them for any kind of compensation or giving credit or any of the traditional things you would do. All right. Let me get a comment for my guest.

So, yeah, that seems like something that has not been dealt with Melanie, Roman. What do you say? What's to tackle that one? Yeah, that's actually something I'm, yeah, I can tackle it. It's something we've been thinking about quite a bit. And, you know, excellent and completely spot on question about there's going to be a lot of evolution and IP law and understanding intellectual property. The things I worry about, one of them has been mentioned, which is how do we understand the origin

and appropriately give people credit for the work that they've done. And maybe I'm less concerned about the big-name artists out there and more concerned with the small guy, right? The people who are trying to sell their artwork on Etsy or places on the internet, and now it's been scraped by this model that treats the entire world like it's testbed, and their art is being reproduced for free. And as accurately mentioned, there's a lawsuit by Getty Images because it actually spits back images that still have the Getty copyright on them. So I'm curious how intellectual property law evolves, but also on kind of a more existential way of thinking about things. What does this mean for the future of creativity?

So what does it mean when a famous celebrity can, you know, maybe licenses themselves to be used? and an AI can just continue to generate Taylor Swift songs every day until the end of time. How can a real human being compete with that? How can we introduce new and novel creativity? Because actually what this AI is going to do is continue to make very Taylor Swifty, Taylor Swift songs. It is not going to be new or different or fresh. It's just going to be a rehash of what we know. So I actually, yes, the lawsuits matter.

Yes, the IP matters because people should be compensated for their work. I also am concerned about the future of creativity in general. Wenton, I hope you got some satisfaction from that answer. Yeah, I think it was spot on because it really goes back to the root of intellectual property and even goes back to, you know, the founding fathers and why they found it so essential to include it in Article 1, Section 8, clause 8, to protect innovation and scientific progress in the country. Yep. Thanks for the call. A lot of people don't realize that the copyright law or the ideas of copyright are right there in the Constitution. They realized how important it was.

President Biden mentioned in his State of the Union speech about trying to rein in Silicon Valley. But this would seem to be more than just involving Silicon Valley. Dr. Chattery involves everybody who's participating. It does. Absolutely. And actually, prior to my current role, I was the, director of machine learning ethics at Twitter. So I am fully aware of what's going on in Silicon Valley. I mean, the other thing that these technologies are introducing is a real shift in, and a seismic shift in the market that is Silicon Valley. It is really interesting to see these actors like stable diffusion and open AI, not the big giants, right? This is not coming out of Google. Google came second or third, right? So it's interesting to see that it's not these big behemoths that are coming out with the state of the art models.

And actually, it is these smaller startups. We also have others in the playing field, anthropic, deep mind, and they may be closely related or affiliated with these companies. But again, they're not Google proper. It's not Microsoft proper. It's not Amazon. So I think we're also seeing a seismic shift as it relates to the tech layoffs and the declining tech of revenues.

Yeah. Just to reflect on what you just said, all these big companies, though, started with small companies, as individuals. So this is basically reinventing the wheel. Let's go to the phones and go to Tim. Tim in, is it, Elmhurst, Illinois? Hi, Tim.

Hi, thank you for taking my call. Yes, go ahead. I have a question that I've asked a lot of people in the computer field and you have yet to get a satisfactory or comforting answer. I just have to wonder what is to prevent AI, what, if anything, is to prevent AI from developing to the point of self-awareness. and self-control possibly leading to a dystopian scenario like the matrix or the Terminator, something like that.

The singularity has it? Yes. Yes. Melanie, what do you say to that? Well, there's a lot to say to that, that we don't really have a good definition of what self-awareness is. You know, we have it, but we don't know exactly what causes it. It certainly has to do with having a body and interacting with.

the world is something that these AI systems don't do yet. You know, chat GPT generates language, but it doesn't have any body or any way to interact with the world and make things happen in the world and get feedback to its body. So I don't think there's a very strong chance that AI is going to become self-aware in any sense, at least not until we understand what that means better. And, you know, the kinds of dystopian things that we see in science fiction movies, you know, we're just quite far away from that right now. But I think the technology has a lot more sort of near-term dangers, the kinds of things

we were just talking about having to do with human creativity and copyright infringement and the misuse of these kinds of AIs by humans. So the dystopia will come from humans. It won't come from the AIs themselves. Ramon, do you agree with that? Oh, absolutely. And she's spot on to point to the people. So at the end of the day, it's people who make these technologies.

But you are correct that generative AI and chat GPT is, you know, one step on the road towards artificial general intelligence as these companies want to build it. So that is the goal of these companies. But again, human beings are investing in these technologies, building these technologies, and it wouldn't create its own sentience. And again, as Melanie pointed out, we're not even, we don't even know how to measure well what human intelligence is. So how would we even calibrate a bar? Just a tweet came in from Gordon from Lewiston Idaho. He says, could AI correct political misinformation in real time? Could that be a good idea?

Well, let's take a break because there's a lot to talk about, and I think I'm not going to get into that question while we only have a few seconds to go. So stay with us. We're going to be right back talking with Dr. Melanie Mitchell, Dr. Ruman Chowdhury, about AI. Give us a call our number 844-724-8255, or you can tweet us at SciFRI.

Stay with us. We'll be right back very shortly. This is Science Friday. I'm Ira Plato. In case you're just joining us, we're talking about the explosion of new AI programs available to the public in recent months with my guests, Dr. Melanie Mitchell, professor at the Santa Fe Institute, Dr. Ramon Chattery, founder and CEO of parity consulting

and responsible AI. And Dr. Mitchell is author of the book, Artificial Intelligence, a Guide for Thinking Humans, Our number 844-724-8255, and lots of people want to ring in. I hope they're not bots that are sending us questions today. somebody says, Kara, on Twitter writes, I use chat GPT to write custom bedtime stories for my four-year-old. That's kind of cool. Roberto from Chicago is concerned we're losing human connection if we depend too much on AI. And let me ask sort of a similar question to you, Dr. Mitchell. I want to talk about something that AI is not very good at, and that's common sense, right?

and you found that out firsthand when you asked AI about yourself and it said that you had died. That's right. Early on I asked a version of ChatGPT to write a biography of me. And it wrote a very good biography of me except for the very last line, which said that I had passed away in November 22, which was kind of alarming. And one of the things that I figured out happened, it was basing. that on somebody of the same name who had actually died and it didn't have the common sense to figure out that we were not the same person. Oh, yeah. Is that because he just didn't do enough homework? Well, it did, I don't know exactly why, but it hadn't like looked at our two, the information

we both had on the web and any human would look at it and say, oh yeah, these are totally two different people, but it hadn't done that. That's very interesting. Let's go to the phones to Denver. Hi. Welcome to Science Friday. Hi. I was wondering what the impacts of AI is going to be on religion and a little more specifically on the development of theology. And then also, what are the ethical and moral implications of AI-generated sermons? That's a very interesting question. Dr. Chattery, want to tackle that? I will do my best.

Interestingly, well, interestingly, though, the Vatican and Pope Francis have actually been quite involved in AI ethics. So there are actually Vatican principles on the use of AI ethically. I don't think I can speak to whether or not, you know, sermons being generated by chat, GPT. I imagine it can, and you could probably try to do it today. I do think the thing that's interesting is, you know, and kind of related to the previous question, how people want to deify or anthropomorphize artificial intelligence. So the question's really astute in that we try to make these things human,

and they're not. They're programs, their computer programs that run online. So I think the interesting part, as it relates to theology, is, you know, human beings need or desire to have, some sort of a higher creature or being that is maybe omniscient and omnipotent and what that does for us in our direction and, you know, maybe again, I'm getting a little philosophical, but it is a fascinating question. Well, I did have Rabbi Mark and Marco Island on to talk about it. Are you still there?

Tell us what's on your mind? Ironically, anecdotally, I had occasion recently to ponder, when chat, GBT came out fairly recently, what the ability of the artificial intelligence program is to synthesize the abstractions of how people of faith extrapolate core values and ethics out of their own scriptural faith and to apply to contemporary issues of the important national and international concern. So I asked chat, TBT, what does the Torah have to say about Russia's invasion of Ukraine? And it very scrupulously reported that Torah or Five Books of Moses is regarded as sacred by the Jewish people. the war in Ukraine is an invasion by Russia of Ukraine, and inasmuch as the Torah was written in the Bronze Age, it has nothing to say about the war in the Ukraine. I posted this on a professional chat group for other rabbis with the observation, the rye punchline, that's why they call it artificial intelligence. And many of my fellow sermon writers said, I think our jobs are safe.

Thank you for sharing that. Most assuredly. Congratulations on your program and every compliment. your guest. Thank you. Thank you. Melanie, Ramon, any comment on the rabbi's remarks? Yeah, I think it's interesting that it kind of refused to make the connection between the Torah and this current day war, and that may be a result of it being, the chat GPT, being programmed by OpenAI to be very careful in trying to make such connections, because there's kind of of a trade-off between it being sort of truthful and not insulting or not profane or not, you know, harmful to humans and it being able to generate interesting, good text. And so there you see something

that it's being extremely careful about. And I think some of that is already, is part of the guard rails that the company has put on it. Michelle writes a long Twitter note that I'm going to just summarize the last part where it says the issue is less about the AI, more about human gullibility and the incorrect bias towards perhaps granting greater default legitimacy to something generated by a computer. Turn the lens on us more than AI, Melanie. What do you think of that idea? Well, certainly humans have been anthropomorphizing computers. ever since they existed. And interestingly, the very first chatbot, you know,

you might be ChatGPT's great, great, great-grandparent was in the 1960s, a program called Eliza, a very, very simple chatbot that pretended to be a psychotherapist, much, much stupider, if you will, than Chat GPT. And yet people thought that it really understood them and had very human-like qualities. So this notion of us being gullible in some sense or being more prone to believe that something that's talking to us understands us, that's something that's very inherent, I think, in human nature. Do you think we might have a spy versus spy? I'm talking about one of our tweets that came in about could AI correct political misinformation in real time. Could AI monitor stuff and, you know, try to decipher what is real from, you know, what is false? Could that be something useful it could do?

I can take that. Yeah, go ahead. Sure. And short answer is yes. I think, you know, I don't want to spend too much time either waxing overly poetic about it or saying it's all bad. I do think that these are the kinds of things or we need to think about how this technology can be used intelligently. I do think that a machine learning model or AI model is able to encapsulate a lot of information can actually provide helpful directional guidance towards things, right?

So imagine it as like an automated Snopes. Snopes is still human beings. If you're unfamiliar with Snopes, it's a website where you can debunk common myths that you'll see online or maybe you hear an urban legends. And those are human beings and human beings can be fallible as well. I think things like this could be useful for something like that. And actually related, we are already seeing, in the AI and machine learning world, we'll call it like adversarial testing, right? So for every good guy, there's a bad guy, and then there's another good guy.

So chat GPT comes out. Students start plagiarizing exams. Now the OpenAI folks actually have a model to tell you if text is coming from chat GPT. So the short answer is absolutely, and these are the things that could be used for very widely. Very interesting. Could I add a little bit to that? Sure, sure. Yeah, so yeah, I think that's absolutely right.

But it turns out that, you know, social media sites have been trying to get AI systems to monitor posts for hate speech and misinformation for a long time. And it turns out to be very difficult because it's such an open-ended problem and it's very subtle. And it really requires a much deeper sense of understanding than these systems have. So I think it's a harder problem than people think to just apply these systems to figure out if somebody's, you know, posting toxic speech or political misinformation. Is there any regulation that could help that? There is actually some regulation on the book. So there's the Platform Accountability Transparency Act that's just been introduced.

There is nothing on the books quite yet, right? So when we're talking about toxicity or misinformation, a lot of that sometimes as Melancholy, has correctly said can be very subjective. And this is, you know, these are the debates going on in Congress literally right now, right? People having different perceptions of who should or shouldn't be allowed to say what online, how they should say it, et cetera. And Melanie is totally correct. Where I do think models have been successfully used is giving directional information.

Otherwise, you can't parse out every piece of information online with human beings to see if it's incorrect or correct. But yeah, it's, it's, so the short answer is there is nothing clear. this is the kind of legislation that's being tackled right now all over the world, not just in the U.S. and TBD. Let's go to the phones to Stefan, I think, in Kansas City. Hi, welcome to Science Friday. Hi. Yeah, my question was about sort of the race to zero, which is something as a, like in a freelance contract artist's world, is something that really stinks because you'll have a client who just says, hey, I can outsource your arts to out of the country or out of your market, price you, out and it hurts my prices. And this should, in theory, cut those clients out. So I won't have to worry about those clients who want something for nothing anyway.

So I'm a little hopeful in that way. I don't know what you guys think about that. So you're saying this is a positive thing. You're not fearful of AI because it eliminates the people who are not going to pay anything anyhow. Right. And it can't create large bodies of work with consistent styles anyway, so it can't create a 32-page children's book with authentic art style on every

page yet anyhow. And then leadership, even my day job as a graphic designer, they can use the tools to kind of give me a better first base. So I don't have to actually have three or four different rounds of how's this, how's this, how's this? I can get straight to something more like what they want. And that could save me weeks. All right. That's something positive. Thank you, Stefan. This is Science Friday from WNYC Studios. Talking about AI, chat GPT. There was a little bit of optimism there from a few of our callers. Let's talk about some of the major flaws that generate AI tools. We've been talking about those. How about what is GPT? What is chat GPT or any of these actually good at? What's their positives? Melanie, let me start with you. Well, they're very good at generating articulate, fluent grammatical language. And this is something that could be very valuable. You know, I mean, people who are, say, not native English speakers.

Of course, chat GPT is all English right now, but there will be other languages as well. Or people who, you know, want to generate something fairly generic, like an email. or a text or something or a short document. That's all, it's going to be an incredibly useful tool for that kind of thing. It also can sort of generate short answers to questions. You know, this is what Microsoft and Google are basing their new search engine strategies on, that we can then ask the search engine a question, and instead of giving us a whole bunch of links to look through,

it actually generates a short, concise answer. The problem, of course, is, you know, you can't always trust it to be correct or truthful or, you know, to contain the right information. But it has the potential to be extremely useful and help people, you know, in their daily work. Let me see if I can get a few more questions in before we have to go. Mike in New York. Hi, welcome. Hi, I'm an artist in Brooklyn, and I, this is more of a comment, but I've been thinking about the

that so much of this technology is just based on the overall human kind of dump that we've done over the last, you know, X amount of years onto the Internet. Maybe there's an opportunity for some sort of UBI, some sort of give back to the population that has basically built this database. Hmm. You mean people, people like you? Yeah, people like me. And so many people.

People like you, Ira. Thank you for that. So how would you see this happening, Mike? what would you see going on? I mean, you know, it would be a big structural change to make that happen, but it just seems like if all of this technology is really being based on all of our collective, you know, input into the world, like there's got to be a way for it to come back. I'm not, I make drawings and paintings for a living, so I'm not an economist, but, you know, that's my thought.

Let me ask, thanks for that call. Ramon, Melanie, what do you think? Is there a way to give back here, Melanie? Go ahead. Go ahead, Ramon. Go ahead, Ramon. Go ahead. Okay.

All right. They're like, we'll throw the hard question of UBI. No, and interestingly, what was just said is something I've heard as well. Like, how can we ensure that people are compensated for what they're contributing to it? So there are new, there's actually a new model that looks into some of the image generation AI and actually can point at the images that were possibly used. use to train it. So that could be one way. I think the difficulty here is identifying exactly which images led to the output or what text led to the output, because again, it's not a search

engine. It's not directly spitting back something that's crawling from the internet. It's actually generating this from a wide range of things. So I think the first problem to tackle would be how do you how do you do attribution? And then the second is how should someone be rewarded? I mean, another thing is we could just take money out of this all together and say, this should be a publicly available product then. Because if it is built on the labor of the world, then it should be available to the world. I think the thing that people are having issue with is not the existence of it necessarily, but the commercialization of it.

You've gotten back to the driver of society. It's all about the money and people making a profit off of these. These were not written just because somebody, you know, or they're not being produced just because somebody feels good about them. They're there to make money, right? So let's end on that. I want to thank my guest, Dr. Melanie Mitchell, professor at the Santa Fe Institute, of course, based in Santa Fe, New Mexico, author of the book, Artificial Intelligence, A Guide for Thinking Humans, Dr. Oman Chattery, founder and CEO of Parity Consulting, and she's a responsible AI fellow at the Perkman Klein Center at Harvard in Cambridge. Thank you both for taking time to be with us today.

Very much. Thank you. You're welcome. Now, here's Emma Gomez with some of the folks who helped make this show happen. Thanks, Ira. Our radio producers are Kathleen Davis, Shoshana Bucksbaum, and Rasha Auredi. Melissa Mayors is our office manager. Ariel Zich is our director of audience.

And I'm digital producer Emma. Thanks for listening. And thank you, Emma. BJ Leatherman and composed our theme music, and we help this hour from our audio engineers, Lisa Gosselin and Kevin Wolfe. Of course, if you missed any part of the program, or we'd like to hear it again, ask your smart speaker to play Science Friday or listen to our podcasts. Have a great weekend.

I'm Ira Flato. Thank you.