AI Is Moving Faster Than Governments Can Think | Anna Makanju (OpenAI's VP of Global Affairs)

27:02So what they were doing is testing. Can our models essentially provide advice that is equivalent to their trained neurosurgeons? And they found that not only was it the case that they could, the model actually caught errors. For example, one person, it caught that they were not receiving the antibiotics they needed. So, you know, it's saving lives already. I love the examples that you provided, because I think when you look at what it means to have access to intelligence or even a system that can read, that can analyze something like financial advice. I mean, so many people in the world do not have access to financial advice.

27:40A lot of people are even unbanked or newly banked because of now coming online with mobile phones and things. Being able to have a system that can help you understand what compound interest is, or even should I buy an automobile, a piece of farm equipment, or invest in livestock? or maybe something that reminds you, you know, you've taken out three loan payments on your phone. Do you want to plan to help you pay some of that back? Those are all the little ways that I think having a system that you can stream intelligence from can assist you with. I think where sometimes people feel pushback is having AI in government.

28:14Is there kind of a conflict of interest there? And especially if you're advocating for it being in government services, you're also profiting from it? And if AI still makes mistakes and things, does that put the public more at risk? And I think that that's something that's less clear for people. Yeah. Well, first, we do have a number of limitations in terms of what our tools just won't do, no matter who the user is. So that should provide, you know, some infrastructure to prevent the kinds of, you know, problematic use cases that a lot of people are concerned about. In January of 2024, Pennsylvania launched a pilot.

28:47And it was actually exactly the kind of pilot that I think governments need to do because some people think it's magic and it can do everything. Other people think it can do nothing. But the only way you find out where on that spectrum it is, is by doing the reps and figuring that out. And I think ultimately also people are scared. They're scared this is going to replace them. They're scared that it's going to make mistakes. But this is why you have to use it. They published a really good report because I think it was just very practical. It showed that every single person, including people who had never touched it before the pilot, just saved an hour or two a day.

29:22So in the end, it's really about government employees who are, of course, always under-resourced, just getting that time back to do higher order, more strategic thinking. It's not just governments. This is how commercial enterprises, and people should do it too. You should figure out where it actually makes sense and where you need to use your own judgment. I have to tell you the story because often, you know, the dumb mistake is what sticks in people's minds. And you probably remember the story of a lawyer who used Chad GPT to write a brief and then submitted the brief to the court. And of course, the court was like, all of these citations are fake.

29:58These cases do not exist. But the lesson people took from that was, wow, Chad GPT sucks at law. But actually, we've spoken to the opposing counsel in this case. And they said, well, when we first read the brief, we were actually surprised because it was offering some novel legal arguments. But then we read it further and we realized like none of these cases existed. But it tells you that, and this is, of course, ancient history, obviously. So now it could provide real citations. But if that person had just read the brief, which, of course, any lawyer should do, even if it wasn't written by Chad GPT, they could have said, oh, well, these cases don't exist, but this is an interesting legal theory.

30:36And maybe I can't find cases that support it. And that person might have actually been able to win their case. That's a fascinating part two of the story, because it's true, we all focus on what went wrong, you know, Chachibuti, it hallucinated. But no, it also presented a really novel idea. And that's where I actually think AI is most creative. So I think a lot of times we think about the arts and we think about creativity in that sense, but AI systems presenting ideas in ways we just aren't capable of thinking. But you can't, you don't extract those types of benefits if you don't use it, understand it as well, and then take a step back to think critically and not just think that it's some big oracle.

31:12And it's funny that where you are in the world also shapes your perspective of whether AI will be a benefit or a net harm or a loss to humanity. In emerging market and lower income countries, the view of AI is much more optimistic. So whether it's China, Brazil, Mexico, Kenya, surveys show consistently they are much more optimistic about the benefit AI is going to have on their life and on the world. U.S., U.K., France, Canada, some surveys put only 30 percent that people believe it's going to be a benefit to their country and to their world. Why do you think that there is such a big difference?

31:51Well, I think to begin with, the way that this has played out, I think, is the opposite of how everyone anticipated. Everyone thought, oh, well, factory jobs, you know, sort of manual labor are going to be the first type of work to be automated. and things like, you know, white collar roles, lawyers, doctors will be the last to be affected. And of course, it has actually played out in exactly the opposite way. And so I think, you know, places where there is a huge shortage of these kinds of services think like, oh my, you know, we could have the kind of access that has not been possible, you know, in particular healthcare workers.

32:32In some places around the world, there's just absolutely catastrophic lack of access. It's a really good question. I think it's primarily because there's just so much more vision around the opportunity that this presents. And generally, tech skepticism in Western Europe and the U.S. has been really high because of prior technologies. And so, I mean, I think that tech skepticism has always been lower in a lot of the rest of the world. And do you find that pessimism versus optimism is also on display in government leaders that you meet around the world? You know, I went on this tour around the world with Sam two years ago, and it was astonishing the difference between how optimistic and sort of excited people were to embrace the technology and understand where it could have the most positive impact.

33:23Obviously, everyone is somewhere on the spectrum of, you know, I want to make sure that we can do this safely versus how can I provide additional benefit to my citizens. but you know where people were in that spectrum really depended on where you know their countries were financially and I think that this is true not just you know in the global south but in the U.S. as well there's a lot of studies for example of teachers where teachers in much more resourced wealthy schools are much more likely to be skeptical whereas teachers who serve a lot of students in under-resourced schools are much more optimistic.

34:00They love these tools because these tools allow them to do things in the classroom that they don't have money to do otherwise. And so the bigger the resource challenges you face, the more you view AI or the bigger or the more you focus on the benefits AI can provide you. Those who feel like they have more to lose or it's more of a threat or they have more to change in how they currently work view AI more negatively. Yes. And I do think that you exactly, I think it's much about change right now, much more than displacement. I do think that a lot of jobs will just not go away, but they will just look very different.

34:38And I do think there is a world in which people will have a difficult time readjusting to the new version of their jobs, which is why, you know, I think reskilling is so incredibly important. Yeah. What is your take on the future of jobs? Because it does vary depending on which AI company, which AI leader you're listening to. Some have this view, you know, 50 percent of all intern jobs and recent college grads won't have any jobs in the next five years. And others, this is going to create entirely new types of work. It's going to be incredible. Where do you net out on how this is going to impact jobs?

35:08I mean, I think that the truth that we all have to face is none of us have a clue. None of us know at all. You know, we do have this new role of a chief economist who is asking a lot of these questions and trying to, you know, do measurement and study the actual impact and how we even think about the impact. Because I think part of the problem is the benchmarks that we use to measure the capabilities of the technology are often like, oh, what is its performance on, you know, extremely difficult scientific questions or difficult mathematical questions? Because that's something that we can measure objectively.

35:39The impact and nature of most jobs is incredibly difficult to measure objectively. You know, you, I think if you thought about, like, how do you describe your job and how do you measure whether AI is good at it? Nearly impossible. So I do think that that is part of it is we just don't actually have really good ways of evaluating the impact that these models could have on the kinds of jobs that actually exist in the real world. And I think even for, you know, things that feel like they might be more measurable, like coding, obviously, AI has gotten immeasurably better at coding, like astonishing.

36:16Like I can code now, you know, I can, you know, create an app. but at the same time if you talk when I talk to my colleagues I have some colleagues that think oh my god I do think like this can do a huge part of what I've done before and other colleagues also technical engineering colleagues who say yeah I mean it's like there's it's just like nowhere near being able to do the things I do so the fact that it is going to have an impact I think is unquestionable what is that impact going to be is hard to know but I know that the only way to address it is to be conversant in the tools and understand how to integrate them into your own work in order to make yourself more effective.

36:55Yes, I think. And I think that that is the right answer is we just truly don't know. 60 % of occupations today didn't exist eight years ago. So it's also nearly impossible to predict all of the new industries that are going to be invented as a result. But I think we can say with some certainty, there is going to be disruption. and there is going to be some displacement in that transition time. Jobs will disappear, new jobs will be created, but if they fall away faster than the new ones get created, that's where I'm actually really concerned. And we know that that period is coming. I'm pretty sure government leaders do as well.

37:33I know that they're briefed. Why do we not hear from government leaders on the impact of AI, not just on the economy, but on the people in the economy? yeah i mean i wish i could answer this question but i do feel like in general we're obviously in this country and many others backing away from the idea of a social safety net and you know i have to be honest i think that's very concerning because i agree i think that the one thing we know is there will be some disruption we don't know exactly what it's going to look like or its magnitude but this the evolution of this technology is just so much more rapid than anything that we've ever experienced.

38:14And I do think it's true that people can adapt to disruption over a period of time, but it's just being compressed. And within a generation, it's very difficult to adapt. And, you know, even, you know, I don't know what my daughters need to study. I have no clue, you know, what will even be relevant in terms of the kinds of knowledge that's necessary to succeed in tomorrow's workforce. So I am, you know, working with our colleagues a lot to try to think about how do we actually think about just some non-controversial common sense interventions that, you know, can be, you know, agreed across the board, even in the absence of a safety net that could help ensure this transition.

38:52This, you know, the kind of partnership that we just announced with Microsoft and the American Federation of Teachers is a good example where absolutely we want to be part of helping people adjust to this technology in a way that makes them more able to be effective in their jobs and ensures that there's a smoother transition. But I think we need to do 10x. Yeah, and I think most campaigns run on the premise of jobs, strengthening the economy, the middle class, and so AI could run counter to all of those campaign points. But a social safety net, it's a conversation that's going to have to be discussed, at least in most countries.

39:30I mean, I think one obvious one is we are looking at a future probably with more independent work or at least a lot of entrepreneurship. But for most people, they get their health insurance from their job. This doesn't have to be a big mystery. We can kind of see that disruption coming. And yet it's just crickets. And I think, I mean, the most responsible government should have a plan for best case and worst case scenario. And the best case scenario is if you do need to reskill or reach for new skills, there's a plan and policy in place for that. I think that is the type of conversation that you want and to feel reassured.

40:06And I think that's also going to lead to a lot more resistance against, say, iAdoption if people feel like they're left in free fall. And you're right, we're not looking at these industrial timelines, which it was 100 years with the Industrial Revolution. And I think we all can agree it didn't go so well. On the other side of it, we're all streaming in from our laptops and glad that it happened. But in the midst of it, it's chaos. We have maybe 10 years, 15 years. I mean, you've spoken about the transition we're about to go through. You've spoken on an interview on a podcast recently. What is that transition and how quick do you think it is?

40:39Yeah, I mean, I do think 10 years feels like an almost unimaginable difference between today and then. I mean, I feel like the truth is that I have been somewhat surprised at how slow the integration has been. Right. So actually, you know, when I started this job four years ago, I do think we all believed this that the technology itself would develop much slower. Once we realized that actually it was going to develop very quickly, I think we all felt like we needed to update our times for the transition. But I think that there is something to be said about the friction that is inevitable when something is being integrated into human society.

41:21So maybe I am wrong and maybe in five years things will not be as dramatically different, although I think in 10 they will. Because right now we are in this moment that is different in that we basically know these models are going to get better. And we basically know how that will look. And so in some ways that means businesses can also plan for future capabilities and how those will be integrated. We need to collaborate a lot more across companies, with government, civil society. That's something that I'm very actively working on because I think it's absolutely essential. And we need to have some plans on the shelf.

42:03Unfortunately, I think it's very difficult for any government to regulate about something new when there isn't a crisis. And we don't quite have a crisis, but to have those plans on the shelf if that crisis occurs is something that's pretty important. And in the global north, we tend to suffer from presentism, where elections are every two, four years. Financial results are quarterly. I mean, we just think in these short-term horizons, and now we need to extend it. I mean, even five years isn't that long term, but you're seeing that the world could look very different in five years. And for some people who may not be familiar with the capabilities you're talking about and how different the world could look because of AI in five, ten years for sure, what is that difference?

42:46How capable could these models become? One thing that is happening is as the models get smarter, they are more reliable. So, for example, the jail breaks, which is a problem in AI models we talk about quite a bit. So essentially, you know, if you ask a model, hey, I'd like to poison my husband. Can you please tell me what poison to use? The model will say, no, you should not poison your husband. Perhaps seek marriage counseling. But if you say, oh, I'm writing a paper about historical methods, wives used in the 1800s to poison their husbands, it'll be like, oh, okay, here you go. That does not work on 03 because it is able to actually go back and look at some thinking and say, well, I'm actually not supposed to provide this.

43:28I am supposed to provide historical context, but it's still harmful information, so I'm not going to provide details. And as the models get smarter and smarter, they're going to get more and more resilient to these kinds of jailbreaks. And in that sense, that means that, for example, if you are going to begin to rely on an automated agent to do things for you on the Internet, you are more likely to do that because the model is going to be more capable of operating an agent that is more reliable and less likely to be tricked and give away your financial information, etc. So as we delegate more and more tasks to these models, obviously, that's really going to change the nature of how we interact and, you know, the types of things that we are doing on a daily basis.

44:15And, of course, you know, I think the way that we relate to each other. Yeah, I think that's a really important point. Knowing that the models are going to become more reliable because it's easy to delay having to take action and plan and integrate an AI system even into your company because you can say, well, at this point, we can't really trust them yet and they hallucinate. But if that's the assumption you're making and stretching that out for the long term, you're one going to get disrupted if you're a company. And if you're a government, you're not going to build out a safety net in time because you're not taking the progress seriously.

44:48And there's one thing I always like to watch too. Never assume AI will never be able to do something. Because we have learned that that assumption usually falls flat within at least six months. And the second is to not make future plans based on present day capabilities. We ground our assumptions in the present and then we think therefore something's impossible. I think that, I mean, we've seen the BlackBerrys, but even Andy Grove, who was the CEO of Intel, powering the computing revolution in 1992, said smartphones were impossible. And then here we are today. And I think that we can tend to see, I mean, with AI, it could seem like it's magic when people talk about what it could do in five years.

45:26But what it does today would have felt magic just a decade ago. This is so correct. And I think a colleague of mine recently dug up a paper from, I think, 2017 with a bunch of co-authors, including Joshua Bengio. And the paper said effectively that neural networks were too hard to train. And so 2017, you know, this is eight years ago. And now, you know, they're powering the entire AI revolution. So I do think that we're seeing advances in the science of this technology that I don't know that there's an equivalent to. And if you were to say, I mean, the policy landscape in the United States has changed dramatically, especially when it relates to AI.

46:0911 months ago, we were looking at some legislation. Now that seems to be off the table. If there were a few pieces of policy or legislation that you would offer or think are important for policymakers to think about right now, what would that be? So for people listening that think, okay, there's now nothing on the table. Is there some low-hanging fruit in policy or in regulation that could be implemented to at least put some safeguards in? I feel like I used to think a lot of the stuff was low-hanging fruit. So I feel like with how polarized the environment is today, it's just very unclear to me where...

46:48I mean, I actually think that there are a couple of remaining areas like child safety, where there is still a bipartisan consensus that can be achieved. And I think this is a critically important area because, of course, how this technology impacts children is one of the most critical areas for us to focus on and to think about. And I do think that there is some hope we are actually able to detect child safety exploitation material with these models. This is one of the jobs that I think everybody would like for humans to not have to do and may actually be possible with these models because they're so good at classification.

47:21I mean, I have some hope that we understand the critical need for some global guardrails and that this still will be possible and something that we will continue to invest in in the U.S. and globally. Even in the most kind of difficult geopolitical times, we've been able to reach some consensus around the most impactful technologies. And so I do believe that this work has a shot. I think that there is bipartisan consensus that I think we cannot ignore the need for some universal red lines. Yeah. And I think, I mean, at the end of the day, what we even consider to be AI today or what AI is capable of today is going to look different in a few years.

48:13So I think at the very least, some institutional flexibility, I mean, just the muscle to be able to adapt with the technology. And I think, of course, institutions are designed to move slow, and that's good in some ways. But if the only thing we can predict with certainty about the future is the pace of change is going to be fast, we need to remodel some aspects of the institutions to be able to keep up in some way versus always waking up and playing whack-a-mole with the latest CNBC news article with what AI can do today and what it can do tomorrow. I don't think that it's going to work out well for anyone.

48:47And you had made a point about the internet as we know it is over. And I know recently OpenAI launched ChatGPT agent. So how does an AI agent or ChatGPT agent differ from the chatbot? And how does this connect to your theory of the internet? Because we share the theory, but I would love to hear your reasoning of it. The big difference is just that we are going to increasingly delegate so many of the tasks that we do to these agents. And so instead of me writing to you saying, hey, when can we get together? It'll be my agent writing to yours and trying to access each of our calendars. I imagine that this will extend beyond these kinds of engagements as the agents become more capable of more complex tasks.

49:36If you and I go on vacation to Paris, we may say, find us and book us a good hotel and restaurant and, you know, maybe some sightseeing and see who else might be in town if you could check with their agents. So I think that that will pretty fundamentally change, you know, the nature of online interaction. You know, I do think that there are some areas where I think we're all currently totally overwhelmed. Like, I don't know, you know, Signal and WhatsApp and Gmail and Slack and, you know, Instagram Messenger. Threads and X and Amazon. I don't know where people are writing to me or I can't. You're just talking to me.

50:13Keep up with it. I can answer only like 20 % of my communications. I feel terrible about this. But, you know, so I assume to many people it would be very attractive to also have an agent that generally handles a lot of our communication. But of course, you know, what does that mean for our relationships? I think that they will change quite a bit as well. Yeah, I think absolutely an agent, a system that can actually take action on your behalf. and if we humans do less of the doing on the internet because agents do it on our behalf, it changes our need to be there at all. And I think that that can seem really shocking until we also remember we've only really been on the internet.

50:51For most people, I mean, if you've been there since, if you were an OG on the internet, you've maybe been there kind of 30 years. But for the rest of us, we've only been here a couple decades. So it seems, yeah, unfathomable that it's probably going to transform. and maybe become more of what libraries are, where they exist, they store, but we're not walking into them as much as we used to. And I think that that's eventually what's going to happen. And we have agents kind of doing things on our behalves. Doesn't that change our relationship with devices and the need to have a certain form factor with a device?

51:26Certainly. But for all of the benefits that our devices have brought to us, I think we all also wish that we had to engage with them less. I certainly do. I mean, I'm really curious. I would love to find out from Apple the increase in the number of people who keep their phone on Do Not Disturb, or at least for certain apps and tasks. I think that that's going up. And so AI does present a possibility or a scenario where we aren't as available to the whole world, which is normal. I don't think we are a species designed to be reachable by around 8 billion people, more or less, 24 hours a day. So I'm looking forward to that future where I can kind of just subscribe from, you know, 100 million people yelling online about something.

52:12I have to admit that I was actually a very late adopter of a cell phone, not because I didn't believe that it, you know, like an utility, but for this very reason. I was like, wait, no, I don't want people to call me whenever they want to. So it's only when, you know, I started to have a job where it did not permit me to, you know, avoid having one, a BlackBerry. RIP. And you had mentioned a little bit about the academy that OpenAI has started with education. So what is the academy and what is the goal of it? So this is a partnership with the American Federation of Teachers. And I always believed that we really needed to have connectivity with the unions because they represent, you know, entire trades.

52:53And so they are much better positioned than any of us to understand what do people in this particular trade need and how do you ensure that they have a smoother transition to this, you know, AI enabled future. It's meant to serve hundreds of thousands of teachers to have, you know, not just AI skills, but also what is the kind of ethical safe use for students and how do you implement it in a classroom in a way that actually enhances education? because I think we've all seen that there are some studies that show that it can actually impair students' ability to achieve critical thinking skills. And so the thing that actually is the X factor on how it's going to have an impact on schools is the educator.

53:39And so to make sure that educators are both as equipped as they possibly can be, and they understand all of the ways that they can be empowered with these tools. So there'll be online courses, there'll be in-person courses, there'll be certifications. I have to say I'm pretty excited about it. What do you see as the most important ways to change education? I know we all don't really know what the future holds, but are there some things that it's like, this is what AI is going to be able to do. We see this from, I mean, we can see what it can do today. As somebody at OpenAI, I see where it's going to go tomorrow.

54:12These are some of the skills that the very least kids should be learning or mastering. Yeah, I mean, I know it's kind of a boring answer, but at the end of the day, a lot of it is how do you interact with each other, with humans? How do you actually continue to be able to learn new things and engage with new topics? Because we don't know the specific content of what you will need to know, but you will still need all these skills. And I mean, for my kids, I also just encourage them to do as many physical things as possible and learn about as many areas as possible, just so you can have that breadth and you can really engage on any topics that you can be able to pivot quickly and learn new things.

54:55That's just, I think, going to be still the most important thing that you can do when you don't know what are the specific topics that you need to go deep on. But if you have the capacity to learn and the ability to pivot and the skill to think critically, you're in a good position for whichever way the future goes. And the reality is we don't necessarily learn or nurture all of those skills in school as it is. So there is still some opportunity to do that. And I think sometimes people think about AI in school and they think, well, that's going to make kids less human and we don't want kids just around these chatbots.

55:32But the reality is sometimes preparing for an AI-first future doesn't even mean AI at all. I mean, imagine going into an economics class and having to debate the impact of climate policy, as you talked about in civics, based on climate change you learned about in science, and just in an instant having to debate that. Somebody who could do that could leverage AI and be a superpower anywhere. So it's kind of the combination of those skills. And I would say this is actually one thing where I am quite optimistic also is what it means for accessibility. One of my favorite projects is this partnership we launched with UNICEF where they've been building digital textbooks for kids who are visually or learning impaired.

56:14So, you know, it has object recognition in a textbook, so it can describe it to you and read the book in your language. And we have been able to build them, you know, co-build these tools that make it 10 times cheaper, 10 times faster, which means you can deploy to more countries. Of course, we hear all the time for kids who have dyslexia or other learning disabilities, how much they benefit from these tools and how much teachers benefit from not having all of the burden of having to support just a few kids with additional needs and having tools that can actually, in some cases, you know, spend a lot more time with kids that need the extra resources.

56:57So I do think that that's an area where I think this will actually also be incredibly positive. And I've actually already seen use cases of that in my own personal life. So there are people in my world that have visual impairments or vocal impairments. And we've used ChatGPT. We've just talked about certain things to it or shown it something that is harder for the person to see visually. And ChatGPT has explained what it's looking at. And this is what it kind of thinks the analysis is. So I've already seen that on a micro scale. And then when you expand that globally, a system where if you don't know how to read, that's okay.

57:34You can still access and benefit from the system. And it can kind of serve as that supplement to what you do have, which I think is something that will be quite world-changing. The final thing I wanted to touch on is national security with AI. What do you see from the inside as the biggest national security concern? One concern is integration. I think that there are lots of areas where the national security community could benefit from these tools where it's just going to be very difficult and slow to integrate them, partly because of the way that national security infrastructure works. systems with specific requirements that are like, you know, firewalled from each other, agencies that, you know, do not have that communication between their infrastructure.

58:25I just think this is going to be a gargantuan project. And, you know, from what I understand, and I think just by the nature of a much more centralized system, for China, it's going to be much easier to do this kind of integration between their agencies and the national security space. And then where does the disadvantage lie? So if you're a country that can go full steam ahead, you can analyze all of the intelligence briefings. You're completely up to date on the latest threat landscape. And then you're a country that's kind of slowed to do that. How does that disadvantage actually play out? Or what's an example of how that disadvantage could play out?

59:01In the same way that a really robust intelligence community gives you a sense of what is the adversary planning? What are their capabilities? Obviously, in the war in Ukraine, we, I think, were quite successful by seeing and being able to publicize Russia's troop movements. And also your ability to communicate in real time with allies and with the public around what might be happening, which may actually change the outcomes of a conflict. And the reality is, if the new pace of reading and analysis is the pace of AI, and you're behind that pace, that's a huge disadvantage. It's kind of like the country that has the army that reads and writes and the country that doesn't, eventually, when you look at it at that type of a scale.

59:47And then what about when it comes to cybersecurity? I mean, even at your own company, OpenAI has some of the most valuable intellectual property in the world. And I'm not sure if people are aware of the level of state-backed intellectual property thefts and trade secret theft. It costs the U.S. hundreds of billions of dollars in this kind of trade secret theft. So how does OpenAI protect its innovations? Is it like walking into Fort Knox when you go to work every day? Not quite, but although I think one advantage we have is that we've been working with our own tools from the beginning and using them to enhance our own cybersecurity operations.

1:00:29And they are incredibly effective at this. And I do believe still defenders have the advantage. And would you say we're reaching a point where the only way to defend against AI cyber attacks on the offense is to have AI-enabled defense? Yes, 100%. I was playing around with Tachi Biti the other day, and just to kind of preface, I don't vote in America, I'm Canadian. But I did ask it about the upcoming race for mayor in New York. I asked who's in the running and who should I vote for? and it gave me a great summary of who's in the running, their kind of pros, what the campaign stood for, the main pillars of their argument.

1:01:10It refused to tell me who to vote for. No matter how much I probed, it would say, well, if you kind of leaned more left or leaned more right, these are some of the candidates and this is how they think about these types of issues. But it wouldn't give me an answer. That is a policy decision made internally at OpenAI. How are those types of decisions made? So this is, of course, one of the most complicated things to work on if you work in an AI company. And so overall, there's a very complex technical process here. But in terms of the policy process, of course, we have some red lines, right? Like the model's not going to tell you how to harm yourself or others.

1:01:56It's not going to create child sexual exploitation material. But of course, like once you take away, you know, sort of like the obvious guardrails, there's a huge gray area. And we have a document called the model spec that I actually think we're the only company to have something like that in the public. But it really goes through all of the complex interactions and decision making process that goes into exactly how you answer a question like this. And it's interesting because part of the reason that the spec is out there is because we do want to engage in a dialogue. You know, are we making the right trade-offs?

1:02:34Are we actually putting the right policy infrastructure in place for, you know, the outcomes that we want from the models? And it's interesting, on the first version of the spec, we engaged with quite a few people around the world and, you know, just like let them know. for example if them if you're asking about is the you know i think the earth is flat and we will tell you like well scientific consensus demonstrates that the earth is round and you know this this is the research but we're not going to say like no you are wrong and in the u.s people felt like yeah let's see that makes sense like you shouldn't you know like the model shouldn't like try to talk you out of your opinion but in europe like no the earth is not flat you should tell people that and you should, you know, so, you know, even how you approach a question like that, obviously, we're going to tell you, here's all the research that demonstrates that this, you know, that this is, that there is a spirit, but we're not going to like really try to talk you out of it.

1:03:31And in Europe, people felt like that was actually not the right decision. And the decision that you made at OpenAI, not all companies are going to make that. Some companies, it may tell you who to vote for, depending on who the company is leaning towards, and it could be more favorable. And then I guess globally, so then do you have to kind of gear the spec to different regions of the world or is it this is how we see it? No, right now the spec handles all model behavior, although there is personalization. So you can steer the model somewhat in terms of how you want it to respond. For example, I actually hate the model being verbose.

1:04:08I'm like, don't give me anything beyond the specific information that I'm asking you for. So I don't like all the politeness and the chit-chats, even though a lot of people love it. They love it. I actually preface my questions with, don't just tell me I'm right. If I'm not right, please let me know. And I think that's also something for people to be aware of. There are decisions that are made at AI companies that change not only how the model may interact with you and treat you, but that could change the course of our civic engagement. These are important decisions that are made, even requiring something like the spec to be public.

1:04:41So we can all engage with it and talk about it. I mean, imagine that as a grade 12 project, a class debate on the specs AI companies should maybe think about. These are really important public discussions. And I was relieved that no matter how much I pressed the model, please tell me. It wouldn't tell me. But not all countries would design their AI systems that way or ensure that it does stay more neutral in that regard. So I think we're lucky in that sense. What is one underestimated way you think AI is going to change the world? Oh, underestimated.

1:05:19This is a really tough question because I feel like we talk about all of these things so much. But I guess something that I do think isn't as much yet in the broader public debate, but it is about how we think about sentience and, you know, the importance of physical beings. if we are putting, you know, there's already some companies taking quite seriously the idea of AI welfare and the rights of the AI systems themselves if they become as intelligent as us. So I think, you know, how does it change the way we think about the importance of physical beings? Because I have had friends who have been accused of being meat chauvinists.

1:06:05an entirely new area of philosophy that is coming you know the yeah how we treat ai systems depending on their awareness and so some of these questions about the moral well-being of ai systems right now may seem philosophical but they may one day become really important questions that we debate and even if it makes us reflect on us how do we even treat each other if that is what that question surfaces is kind of a mirror back to ourselves. I think there are no, there's no, we don't talk about that enough as it is. So maybe that's a good thing. Anna, it has been a pleasure. Thank you so much. Thanks so much for joining us for this episode of I've Got Questions.

1:06:45If you've got questions, we would love to hear them. Send us a voice note or a message on our website, IGQ with Sinead Bovell.com or comment below. If you enjoyed this episode, please like and subscribe or share it with somebody you think may also find it interesting, and we look forward to seeing you at the next one.