[0:00:10] And with that, let's move on to our open keynote. Our keynote speaker is Jack Clark. Jack is the strategy and communications director for OpenAI. OpenAI is a privately funded consortium that is engaged in AI research and AI policy work for the public good.
[0:00:37] Jack, in his role there, focuses on policy, safety, strategy, and communications. He also helps run the AI Index, which is an initiative that's part of the Stanford 100-year AI study, which tracks and analyzes AI progress over time. In his spare time, he writes a weekly newsletter about AI research and applications called Import AI. That is importai.com. It's read by many experts.
[0:01:06] And Jack is, as you've gathered, in many ways an important person in this field, one whose voice is unique and is valued by many of us who are paying attention to AI ethics and policy. So please join me in welcoming Jack Clark. Let's see if the machines will let me speak. Fitting my personality, I deny or compliment and accept all criticisms and encourage all of you to offer me.
[0:02:07] I think that that's a very admirable problem and also an incredibly difficult one because the nature of AI is such that it's global. It isn't happening
[0:02:35] In any one country, there's no particular country or region that has technical dominion here, nor is there a particular single actor, such as a company, an academic discipline, or a kind of power block. It's truly a global thing with global characteristics. So when we think about ethics, we need to think about how we make that as inclusive as possible and how we can develop functional approaches to ethics which work across different borders and different value systems. So the motivation for...
[0:03:04] Why I think it's worth thinking about it in this way is that if we don't do that, we end up with the same kind of vulcanized technology standards and policies that we've done for the last few years, and those probably haven't worked. Today's world is beset by inequities and problems which have occurred almost directly because we've chosen to act within local or national frameworks rather than international ones, and that's led to...
[0:03:31] A series of problems. So as an overly elaborate rhetorical motivation, I'd like to present you with this charming picture which I made. And the reason why I'm showing this is that I think that, you know, why did the Luddites get so angry? Or why did they feel like they had some kind of right to go and...
[0:03:51] Disrupt the system therein. I'd argue that some of it is because the system was rolled out, the transformative technology and a set of approaches to automation and the systemization of daily life were rolled out somewhat thoughtlessly and without a care for the larger ethical construct in which they were designed. And as a consequence, you had a huge amount of...
[0:04:12] Thank you very much.
[0:04:41] Ground this a bit more in what we're dealing with by ethics. And in the spirit of this being inclusive and open, I did what I think many people outside this room will do, which is they'll go to the internet and see what the internet tells them about ethics. You can imagine what's about to happen. Ethics, the discipline dealing with what is good and bad and with moral duty and obligation. Well, that doesn't seem too bad. Let's just double check what moral means there.
[0:05:09] Good. Moral. Overall relating to principles of right and wrong in behavior, which is starting to feel somewhat kind of unsteady and shaky. Just make that clear. Good and bad and right and wrong. We're already in a sort of problematic domain where this doesn't really actually work for the vast amount of people that need to think about this. And so rather I'd like to think about how we can come up with a kind of global ethical construct which might allow us to
[0:05:37] To reason from an overly simplified view of ethics which might let us do some productive things and then we can come up with much more precise definitions of local geographies later. Since this is a conference kind of steeped in history and steeped in many intellectual disciplines, I thought I'd do something reliable which won't cause too much problems here, which is I just went to Kant. And I think Richard Kant isn't the worst person to go to.
[0:06:05] for a notion of ethics in a sort of functional way, which is, you know, how do we carry out actions which can seem just to the largest number of people in the domain we're acting in? And so I think that that is a useful thing to think about as we build this. And I'm going to, for the rest of this talk, talk about how we try to sort of achieve this in a couple of specific domains and what we've learned in a couple of specific instances and what problems we've encountered as well.
[0:06:34] Because ultimately what we're trying to build here is an accretion of different sort of value systems or scaffolds or way of dealing with the world that actually map to what we aspire to. And you can see the notion of sort of ethics and the notion of a rollout of this technology is made concrete by sort of artifacts or documents which happen as an outcome of lots of people thinking about this. One good document is...
[0:07:01] From a previous administration, we have the guide to how to do a sort of natural strategy in the US for AI rollout. This did involve some of the sort of ethical problems here to do with distribution and problems to solve. And it was kind of one component which is going to end up going into a larger kind of construct that people use to reason about this. And the reason why I highlighted that document in particular is that
[0:07:29] It has a sort of chain of causality within itself that is already becoming quite significant. So this is less well known. This is a 500 page ethical and policy guide which was put together by a bunch of very significant actors in China and published late last year. This is about 500 pages long and it goes through the issues of sort of AI development and deployment within China and the spheres in which it operates.
[0:07:56] In this document they reference a bunch of sort of artifacts which have already been created by other parties ranging from the Asilomar principles which is a set of 23 binding principles for how we should develop AI that was converged upon by a bunch of experts or alleged experts about two years ago. There's also the OSTP AI report and the IEEE's
[0:08:20] And one of the recommendations in this report from China, which I found interesting, is the view that ethics of themselves are technology. And ethics are a technology that China needs to have just as much of a role in as any other country. And that totally makes sense. That is the kind of just and right thing for any large group of people to want. But it's already going to test our notions of how we roll out things. The way that most companies and actors in the West behave today is they...
[0:08:49] They do policy via artifacts. They create a technical system and that defines the way that people interact with it. And they roll it out to a large number of people in the world and then they have that condition for policies as a consequence of it existing. But that's not going to work if we need to do things in a more collaborative way, which is probably the main thing we should strive for if we want this technology to happen in a way that benefits everyone else. So I think it's valuable to, in light of that, think of
[0:09:16] Ethics as being a series of choice areas where we can have views and we can come up with policies and paths forward. Now, there are lots of these. This is a very incomplete list of them. And I want to acknowledge that by talking about two of them, I'm not placing those at some kind of privacy relative to the others. I think that all of these are sort of equally valid and I've left lots of things off here which should be on here.
[0:09:41] So I don't want it to be seen as an implicit value judgment, but I'm just picking two because there are other ones which should be talked about and indeed will be talked about later today. With the two I'm picking are areas of collaboration and proliferation. And the reason why I think these are worth us thinking about through this lens is that proliferation speaks directly to the fact that this technology affects lots of people, it goes sort of everywhere, it travels widely and it has effect.
[0:10:07] And collaboration is the larger question of how do we expect to collaborate across different borders or communities. Again, that's why interdisciplinary centers like this are important and why I think we should view them as a model for other stuff that we do in the future. So I'll start with proliferation, and I'll explain sort of what I mean. So if you're not super familiar with underlying AI technology, I can give you a flavor of some of the progress which has occurred and why.
[0:10:35] You know, I have some level of reason to be standing here and why we're all so excited about AI. So these are sort of faces that are generated by an AI technology called a generative adversarial network. And what you'll notice is that between 2014 and 2017, we went from generating like fairly strange kind of 1970s art, black and white photos of people to...
[0:11:02] Relatively realistic things, but you could mostly mistake for reality. And so, you know, why does that matter? Why does this have some kind of import on proliferation? Well, all of these technologies are sort of dual use, and they can be used for a variety of things, and they're developed currently somewhat naively, but their applications end up being anything but. So here is a speech which was given at MIT recently in the top left, you see.
[0:11:29] President Obama being transformed into President Trump via a generative adversarial network. The implication being that this technology can be used to take any sort of given politician and transfer them into another. You also saw this happen with deep fakes recently which was, as with lots of things, pornography for certain weird applications before other people and they started using the same technology to transfer people's faces around in that domain. Both of these are
[0:11:57] Are somewhat sort of problematic, but maybe they seem fine. Maybe these just seem like people playing with kind of a new technology. So I think it's reasonable to say, so what? But I'd like to talk about another aspect of this and sort of try and have you go on this journey with me to understand why I might be concerned. So this is a graph of our performance.
[0:12:19] Image recognition in recent years. That graph of human performance is actually a friend of mine called Andre Carpathy who sat and stared at something like 10,000 images and tried to label them himself. He discovered that he's very bad at labeling specific dog breeds and computers are much better at that than him. But I think he's better at computers at labeling seatbelts and fried eggs. So we have to give humans something in this area.
[0:12:45] But what it shows you is that our technologies have got to a point that we can sort of exceed some measure of human performance on something as trivial as image recognition. Again, that seems sort of fine and benign, but then you start to think about the domains it's applied in, and this is a consumer drone which just came out, which sort of will follow you as you're doing a workout for fun. And it works entirely using these technologies. It relies on the...
[0:13:14] The kind of diffusion and proliferation of these things and I think that this might be fine but I challenge you to look at this and not feel slightly uneasy because it has a number of interesting applications beyond this sort of aerobic outdoorsy stuff which they're advertising it for. Again you might not be convinced yet so that's fine. I'm going to do one last example and then we'll move on to the next section which is...
[0:13:40] Those are about AI that can sort of classify the world around it. And again, that seems fairly static. It seems somewhat controllable. We can just embed that capability into a technology that we can control. But AI can also take actions. So here's sort of one example of recent progress is in 2013, we were able to train AIs to solve video games like Pong. By 2016, we were able to train them to do things like Go.
[0:14:08] DeepMind had a number of good publications on AlphaGo, which was doing pretty impressive work. We at OpenAI trained it on a computer game called Dota 2, the point being that it's able to take interesting actions over longer time horizons. How that applies to the real world is it has direct relevance for things like the automatic navigation of drones, training drones to be able to act autonomously over certain time horizons. This was done...
[0:14:36] In September 2017 last year, and it was done entirely in simulation. And I had a bit of a back and forth with the author saying, well, this is in simulation, so why should I care? And then I got this email about a week ago, which I thought was interesting. It was, you know, very excited. You know, thanks very much for all of your feedback. I was looking at something that OpenAI published, and I figured out how to get this to work in the real world. And then he even sent me a video, which I thought was nice. So a technology that we published and gave.
[0:15:06] My reaction is that this is somewhat problematic. And where this sort of intersects with ethics is this has been done in the absence of people thinking about
[0:15:33] about what happens when you make these sort of computational artifacts. It's been done somewhat thoughtlessly and we have a culture of rapid prototyping in the technology space which I think might cause problems as we think about people with different sort of value systems and different views on what is just and what is right trying to use these technologies.
[0:15:55] We've tried to do some work on this by thinking explicitly about what are the implications of this. And that's why I think the approach of scenario-driven analysis here is particularly useful, is that you can take what you know about the world and you can forecast forward. So we published a report recently, The Malicious Use of Artificial Intelligence Forecasting Prevention and Mitigation. And the idea there was to look closely at
[0:16:24] Existing technologies and how they're behaving and ask ourselves about how we thought people might take that technology and apply it for different purposes, whether good or bad. The idea being that we can come up with scenarios which feel grounded in reality and are somewhat based in the future and use that to reason backward about what we should be doing today. The nature of this work was very interdisciplinary, which I think is also key.
[0:16:51] If you act as single actors in this domain, there's less of a chance of you being listened to. But if you kind of team up with large groups of people and make sure that your study is actually incorporating a reasonably wide set of people, there are many others on the paper along with these organizations here, then you can get to good outcomes. So what did we learn, or what are some of the things that we came up with? Well, we found...
[0:17:16] Something which is probably not a surprise to most people, which is that if you take a bunch of people and sit them in a room for two days and really scare them and then sort of shake them and have them come up with scenarios, you can come up with useful things. But more specifically, what we learned is that we can make progressions about how we expect cyber warfare to change based on the incorporation of AI techniques. So when you have existing computer bugs like...
[0:17:43] WannaCry, we imagined a version called Wanna Laugh, which would use sort of neural network fuzzing techniques to make itself more difficult to detect and more difficult to remediate. That seemed somewhat real. We had a scenario about people using generative adversarial networks to take the faces of politicians and superimpose them on other ones, and as I showed you earlier, that actually just happened recently, you know, while this report was being written, so that felt somewhat on the money.
[0:18:12] We came up with ideas about how we expected existing consumer hardware to be repurposed for relatively low cost assassinations. That again feels directionally correct and while we were writing this report that drone came out and a number of other technologies have subsequently come out which seem to show that the arc of technological progress points to this. We also thought about targeted advertising and how we could expect that to worsen.
[0:18:39] Also learned is that questions are in many ways the solution here. The field is somewhat kind of young enough and at a pivotal enough stage that just by having conferences like this where we can gather around and actually think in detail about what questions we have when we interact with these sort of scenarios and what questions this technology raises can themselves be useful. So now what we're trying to do is to get sort of more questions to help us understand this stuff.
[0:19:08] To work with more universities, more startups, more companies and also countries to think about what the next stage is. So we have a report outlining birth threats. We want to do a report outlining specific threats that people have encountered and that on the idea that all great art is theft is wholly stolen from that sort of scenario driven approach here, which is we're going to go into the communities and figure out what they're doing and struggling with and try to come up with things.
[0:19:34] And I think it's important that we remember that we don't necessarily have a great deal of time here and the world is moving sort of more rapidly than we think. I thought this was quite a neat example where our report came out and 10 minutes after Baidu published a paper about voice cloning and published some of the details of it and that means that in a year or two we're going to have synthetic voices that are somewhat indistinguishable from reality and you're going to see people
[0:20:02] Deploy that stuff in a fun hacking mindset way. And if we don't change the discourse around it and change the norms in the community around what you are and are doing with this technology, you end up in situations that you'd rather you hadn't ended up in. So the second area I'll talk about quickly is collaboration, which I think is one of the areas that we can think about for solving this stuff.
[0:20:30] And so collaboration by nature to me is something that everyone always wants to do, but the characteristics of AI make it uniquely possible at the moment. And I'll use an analogy which some of my colleagues will subsequently criticize me for of nuclear technology, but I think that it's reasonable. So Niels Bohr in the 50s wrote a letter to the UN, and he was referring to something called the Baruch Plan, which I'll shortly talk about.
[0:20:58] Where he's sort of observing that when you have a technology of such like power and utility and generality, you have a unique moment to act nationally and internationally because you can express the impact of this. Like this is an industrial revolution that we know is coming. It's not just going to manifest as sort of destructive technologies as some of my sort of talk has also discussed. It's going to manifest in technologies that change our relationship to each other, change how we
[0:21:29] Interact with the world around us and change how we manifest things like persuasion at scale. So it's something that we might want to collaborate on. And what we lack in AI is a notion of risk. We don't have the equivalent of the atomic sort of bulletin scientists which has what's called the doomsday clock because AI isn't like a single doomsday-like device. But it has...
[0:21:57] Doomsday characteristics in the sense that before this application of large-scale AI can be intensely sort of threatening and worrying to the world. And so we need to think about what we're doing here with some notion of urgency. So I'll talk about the Baruch Plan briefly, which is something that I still find to be a remarkable thing to happen. So the Baruch Plan was a notion that instead of countries developing their own nuclear technologies,
[0:22:27] Share as much information as possible among all of the countries via the UN to allow for a safe application of civil nuclear power. The UN would administer and control nuclear technology. You would inspect everyone and there'd be disarmament. There'd be sort of a notion of disarmament with a pointy bit at the end, which was that if you were found to have sort of...
[0:22:51] Defected and be developing your own nuclear technology, the UN had agreed that they would have the right to come in and give you a stern talking to, perhaps using the things you can imagine they'd use. The reason why it almost worked is that they were at a time where the transformative potential of that technology was obvious and they built the necessary international consensus around aligning on certain basic values or basic precepts where it was possible.
[0:23:21] Why it kind of didn't happen was a lack of trust on both sides. It was because the USSR thought that the Americans were lying. And if you inspected the historical record, the Americans weren't lying. They really were going to try and do this. You can probably forgive the USSR for assuming the other. But coming back to my slide earlier where I showed that we already have different notions of AI policy and AI ethics being developed by different actors around the world.
[0:23:51] This to me motivates why we should be thinking not just nationally but about an international kind of ethical framework and consensus here because if we have that things like the BRUTE plan have a better chance of working in this kind of new regime and if we continue to operate as as blocks or blocks of presumed moral authority or blocks of presumed like ethical value that occur over certain countries then the chance of a global collaboration will be
[0:24:17] So how do we actually go about doing this and how do we do this in a functional way that
[0:24:46] That's realistic. Well, I have a couple of propositions which are grounded in things that are going on right now. One is the idea of measurement and forecasting. So for us to think about the sort of impacts of the technology, it's useful to know what the technology is and where it's going. And the act of measuring something helps me understand what I'm dealing with and forces me to develop a competency in it. And the act of forecasting forces me to reckon with its effects on society.
[0:25:15] Some governments around the world are starting to think about both of these pieces. And if we can do a better job of articulating the need for measuring, say, the progression of AI, and not just in technical domains, but also in social domains, its impacts on people as well, and if we can do a better job of forecasting it, then we have kind of natural bipartisan bridges that we can use to talk to other countries and to develop shared methodologies, you know, shared weights and standards.
[0:25:45] Ideally, you don't end up with what we have now, where you drive on the left side of the road in some parts of the world and the right side of the road in the others, but that's actually a good outcome relative to driving in a whole different multitude of directions. And it shows that it's possible to have some notion of consensus that you can build. Step two for this is, we're going to return to my great artwork now, is the notion of safety and predictability within AI. And what I mean about that is that
[0:26:15] For all of these technologies, we want them to be safe. We want the populations that they're applied to to be safe. And I think no government or no act in the world is like, yeah, let's make an unsafe or a biased or a racist AI and deploy it in our local population. No one has any incentive to do that. And there's clear incentives for the opposite, which is collaboration and sharing of best practices and sharing of things that we've learned.
[0:26:41] And the second is predictability. Again, no one really wants this stuff to be unpredictable. Everyone wants to better model how it influences the societies it's deployed in. And so you have clear reasons for why people might want to actually collaborate on this stuff and do a good job together. And then the third step, which is probably the most speculative, is that you can have shared projects. I mean, I'm from England.
[0:27:10] I can attest that Europe does actually a reasonable job of doing shared projects across multiple different countries. I know that in the case of AI, there are good international collaboration frameworks in place in a few parts already. There are also organizations like Partnership on AI, which is trying to bring together large groups of people. I think that if we shy away from the difficulty of operating across borders or operating across geographies,
[0:27:39] And sort of retreat to operating within local regions of impact. You run this risk that we can all do the directionally right things that make us all sort of feel good and just about our behavior in the local sense, but you end up with a global set of incompatibilities. And these incompatibilities seem uniquely dangerous due to the kind of speed with which AI is being developed and its general utility to society.
[0:28:07] So I think just before I leave you with is that I think that ethics get embodied by tools, and tools govern a lot of the context in which things are used. And if we can develop a better notion of shared tools and a shared appreciation for challenges, then you can develop a better chance of having an ethical notion which is shared globally. And if you have that, you have the possibility of...
[0:28:34] And if you don't have that, you have a lower possibility of that, which seems strictly worse. So that's the end of my talk. I hope it was maybe helpful. And I'm happy to take questions for a few minutes if that's useful. Thanks so much, Jack. It's a really wonderful opening for what we're trying to achieve on this day. I love how you said that questions are the solution. I wonder if there are any questions or solutions in this.
[0:29:14] And if not, I'd like to ask you, what are some questions that you've heard that you think are worth asking? Well, I'll give a specific example, which is after the malicious access paper came out, I was contacted by a number of companies which had been dealing with issues of dual use and proliferation.
[0:29:33] And one of them was talking about their drones to me and they were saying, well, we don't say this publicly, but we've actually implemented a load of technologies within our drones that can tell if someone is strapping a gun to the bottom of it and it won't work. And I was like, wow, that's an incredibly disturbing thing to hear. I'm glad that you're doing it, but I also felt somewhat sad that we hadn't de-risked this discussion enough that these people could say this publicly or really share that knowledge with anyone else. You just have...
[0:30:01] Local, private actors that are making bizarre, ethical, and moral decisions about their technology because they're trying to do the right thing, but because they lack a shared sense of what everyone else is doing, it's going to be sort of insufficient to the size of the problem. Hi, Jack. Great talk. I tweeted a lot about it. A quick comment and then actually a short question. The relatively quick comment is that I want to just totally support what you said about the EU.
[0:30:31] The Eurozone alone, not the entire EU, has been the world's leading economy since 2007. And I think it's really important to look at the EU about what they're doing right. And I think the EU itself needs to think a little bit more about how it is leading the world and should be thinking about other, facilitating other international hubs in constructing these kinds of public goods because it really is providing a template and it should just be thinking internally. So then the question is,
[0:31:00] What about OpenAI? Is there anything that we should be doing to try to connect into your organization specifically, or is this something that you're... That's the question. So there are two things. One is that we're starting to become a lot more active in, I guess, what I term public policy. And Malicious Actors was a thoughtful experiment here that seems to have not been an exploding cigar. It seems to have actually worked somewhat well. And so we're now going to try and...
[0:31:30] Use our bridge building to convene more people. And I think that one of the things that OpenAI can do is because we have certain financial resources and assets, we can do certain bits of research somewhat earlier than certain academic communities. But because we don't sell a product, we don't face the risk of communicating about negative policy as well as positive policy. And so I think what we want to do is start talking more about some of the hard questions so that we can make it safer.
[0:32:00] For other parties to join. So I'd love to chat with you about that because I think that's one of the main things. Secondarily, there's a technical question. And I think as I alluded to in the talk, we think that for a lot of what we do, we're still in the era of it being safe. Like I don't think that drone I showed is super disturbing. It's like pretty janky and no offense to the research. I tell this to them. It's not something that you're going to like lose sleep over.
[0:32:27] It uses a technique which we sort of wrote about, and I think we also, it doesn't use code when we publish, but that alone makes me think that one thing we need to do is start thinking, ironically enough, about openness more. And I'm not talking about us unilaterally being less open, I'm talking about the whole industry developing a notion about what artifacts are safe to publish and what are less safe, and how do you talk about research in such a way that the...
[0:32:55] People who need to benefit from it do, but people who don't benefit from it sort of aren't less aware of it. Maybe I really want to pitch in there. Sorry about the turn-take. I was actually one of the authors of the malicious... Yes, I did. Yeah, but I also was one of the authors of the EPSRC Principles of Robotics, which are from 2010, were published in 2011.
[0:33:17] A lot of us wanted to put open source as one of the principles, or at least mention it as one of the principles. And it was because of that, and ever since I've heard people going back and forth, this is the first time I've heard it back actually for a while. But we were convinced by people in the room that requiring open source exposed us to script kiddies and endangered people with respect to domestic AI. So I know that a lot of people say that open source is what
[0:33:44] It makes it more likely that we solve the encryption problems. And I think that's the strong position of the CITP. But I am not an expert in cybersecurity. I just think I want to flag up that's super interesting that you said that. Thanks. Hi, Jack. My name is Julia Palazan at Cornell Tech. I wanted to pick up on your points about collaboration and referring back to the Baruch plan.
[0:34:12] And wondered whether taking some more recent issues in international tech policy, such as privacy or management of critical internet resources, whether you could be precise about whether you see these solutions coming from cooperation between nation states or if you are supporting multi-stakeholder style solutions.
[0:34:35] For now, I'm much more supportive of multi-stakeholder style solutions because I think that a lot of this stuff gets enacted through what I sort of mentally call the creation of artifacts. The creation of technologies that kind of embody a policy. And that's dangerous. It's not necessarily the most... In an ideal world of infinite time, we would do things, I think, in a much more state-led manner. But in the absence of rapid...
[0:35:02] Policy leadership by states or certain states that have the ability to drive the global conversation. I think that you need to think about multi-stakeholder solutions as a way to create norms that later can be a template for states to use. And like in the case of this malicious actors stuff, I think that it's inevitable that in a few years some great, terrible thing happens using off-the-shelf AI. And in the absence of...
[0:35:30] Lots of multi-stakeholder work and coming up with methods of collaboration, de-risking, and disaster response there. We're going to be in sort of a worse policy environment. So I'm basically pragmatically saying, do multi-stakeholder first, but also talk to nations on the assumption that they move more slowly, but ultimately they need to implement a lot of this. Does that answer your question? Yeah, I guess, I mean, one of the implications of the last 15 years
[0:35:58] Internet governance has been that there hasn't been much action from states in response to that. But I think that that doesn't mean we should lose hope or think about... I mean, I think that AI, if I'm not crazy, and I mostly try not to be too crazy, if the technology continues to scale and it continues to sit on generic hardware...
[0:36:25] And the capabilities remain somewhat naive. Then we are looking at what I think of as fundamentally transformative applications, but will require large amounts of international coordination, signaling, and norms to maintain stability. So it seems like it makes sense to involve states and ultimately assume that states need to take the lead on this. We just need to create the...
[0:36:50] It seems to me that companies are incentivized to iterate quickly on technology and consider ethics afterwards. Given that, how can academia work with and incentivize companies to consider ethics?
[0:37:17] Journalists, so I say this with some small amount of experience here, but I think that the media is an incredibly powerful tool to actually shift how companies behave. And because the media is actually being systemically taken apart by technology companies, they have a certain... We're at a special time where by providing lots of information and expertise, academia can fill some of the...
[0:37:41] Some of the time which has been removed as a consequence of the economics of media changing, you as the experts can step in and provide examples and sort of didactic frameworks and scenarios. And if you can make this real and make this something that's deeply felt, you can create narratives that alter the actions of companies. I also think private collaboration with companies is great, and companies mostly are full of people who want to do...
[0:38:07] The right thing. Like as doomy as I sound, the reason why we're all here is that the world is mostly nice people but want to do the right thing. You know, if it was different, humanity would never have gotten this far. So you can always talk to each other as well. Well, this is cutting edge artificial intelligence we're talking about. I'm somewhat concerned about a couple of things I've named artificial mediocrity and artificial stupidity out there.
[0:38:36] When I call my insurance company, I get a long thing, and then it has a voice recognition, and it says, well, do you want to do this, yes or no? And it's terrible. It's none of the above, and you're stuck. So a lot of things get implemented early before they're very capable. I think it's valid to be concerned about it. I mean, I think a lot of this manifests as choice. There's a lot of terrible...
[0:39:10] Morning, great speech, fantastic. Two things, one is please add unions to your list of stakeholders that you want to work with. Yes. And then the second one, oh Jesus, I've just had a morning...
[0:39:44] Hi, thank you so much. I'm currently from the Internet Institute. So I have a question specifically about multi-stakeholder and unpacking that a little bit because it came up earlier in the context of Internet governance. And something you see there, because it's been an ongoing process for much longer, is that that in many ways is very imperfect, but specifically in the way that it allows for corporate capture.
[0:40:13] And considering that a lot of the players in this field are corporate actors, I was wondering how we need to think about that issue as it relates to the field of AI as we build these multi-stakeholder initiatives. Yeah, I mean, I think that as a... I thank you very much for having me to this speech, but the fact that I'm giving this speech is worrying. We are not an academic institution. We're not a public good institution. We are...
[0:40:39] Funded by people. We have a goal ostensibly to do what benefits humanity. You know, we do try to. But the fact that we have some kind of power as a consequence of capital, which has enabled us to do certain bits of science, which involves legitimacy, makes me nervous. It makes me nervous every day. When you think about corporations which are like this on steroids, some of it is about broadening the public conversation through precise engagement with people.
[0:41:09] I think that some of it is that, and just to come back to media, and this isn't because I'm familiar with media, I think it's because the media is really important here, is that the mood of the world is shifting away from people that govern without a mandate. Tech companies govern without mandates. I didn't vote for people in Facebook who have certain free speech policies. I just use Facebook and I use their free speech policies. And there's some fictitious notion of choice here, but it's...
[0:41:39] It's fictitious. So we need to probably tap into the political discontent and the change in public mood to start asking companies constantly about why they are the ones who are presumed to know. How do you deal with regulatory capture in the form of just, I can pay people to go and talk to politicians and you can't?
[0:42:02] It's more systemic problem with neoliberalism than with AI specifically, but I agree it's bad.
[0:42:13] AI chip working. I'm not as patient as you are, Jack, on this let's wait and try and mobilize our national governments. I'm not that patient. No, you sounded more patient than I am. No, I mean, I'm just thinking here, you know, the OECD, they are now working on their largest ever horizontal project and all of this.
[0:42:39] How can we somehow also begin to mobilize around the whole question of the space of policy? When we sent out, my organization, a call for a global convention on ethical AI, we, within 10 minutes, had a phone call from the OECD, the World Ignite Forum, and so on. So it's lurking beneath the surface, right? And I think we need to push that because...
[0:43:04] I don't believe that all sorts of various national initiatives are actually going to bring us into this safe, ethical, sustainable, transparent world that we want.
[0:43:25] Political actors in the world who can enforce change. And they're all looking for material that they can use to deal with their grievances with one another and carry out the grubby business of politics. I think that providing information, inputs into that can empower people. So let's talk about unions. I see lots of surveys of notional experts about AI. And I see lots of surveys of the general public. I think a survey...
[0:43:53] across the largest set of unions possible in the world about union members attitudes to AI in their areas would provide the sort of information which you can give to candidates or incumbent politicians that are friendly to your positions and they can use to motivate rhetoric in their space and I think that that can actually be done in in lots of different areas where the desire of the world to understand AI is so high that being able to provide
[0:44:20] Information inputs about it from groups which are otherwise less represented can change things. And you've actually seen this in the issue of bias in AI recently, where people have started to just go up to image classifiers and be like, I'm not going to just show you white, balding, middle-aged men. I'm going to test you on the world, so I'm going to test you on representative data sets. Mostly the companies fail.
[0:44:44] The companies usually fail in such a way that it's such a public embarrassment that they change their policy and politicians notice as well because then it's a constituent issue. So I think that you can take these things which are uneasy about and you can ground them in a real information contribution and that information can shape a discourse to your benefit. Hey, really, really enjoyed the talk. I thought it was really, really comprehensive. What I am concerned about is the idea of shared ethics.
[0:45:15] You were mentioning the US, you were mentioning institutions in the UK, and you were mentioning this report by China, which I actually participated in making, and I don't think that that's what it means. I'm sorry. The idea of shared ethics and this kind of collaboration where there are a lot of fundamental ethical principles that I think are in very clear conflict, whereas you would say human rights in the West and...
[0:45:45] Harmony in the East, or you could say privacy versus actually functioning technology. So in a sense, there are things that may not strike you as things that are conflicting, but I feel like if you were to create a shared body of ethics, how do you plan to reconcile that and really give space to everyone to come and present their own ethical considerations?
[0:46:08] I mean, I think of this as somewhat like an imperfect scaffold where we should attempt to create some set of shared assumptions which have ethical notions. I think the issue of dual use technology is probably something where we can get some level of international agreement on very basic bits of it. I agree with you that the notion of every...
[0:46:31] I don't see it as something that we should avoid working on for now. I think that we'll learn a lot in the process, and I think one of the things I'm trying to motivate here is that I haven't seen, at least in AI, really great attempts at even trying this. There's an assumption that it's irreconcilable through organizations like the Partnership on AI, which started with a heavy sort of...
[0:47:00] Western inflection, whether they're able to broaden that out, I think is another interesting test case. I do agree with your skepticism here, and this may be one of the rare instances where I'm slightly more optimistic than you. I think we have one more question, and then we'll break for coffee. Thanks, Jack.
[0:47:30] So I'm a little more hopeful than the need on this one. And I just want to take the kind of ethical, take the ethical angle from a different trajectory. And I wonder if we might not urge scientists to just look in the mirror in the sense that the very ethical proprieties of their practices have kind of operated a condition of advancement of the sciences, which is to say the practices of
[0:47:59] Open and inclusive dialogue, egalitarian reciprocity in sharing information, mutual respect, sincerity in sharing information. All of these kind of, we could even call them democratic norms of participation that have informed technological progress are right there. And so if they're procedural, they're procedural. But at least it's a starting point to start thinking about how
[0:48:27] The scientists themselves can be seen as implicated as kind of ethical participants. And this is why specifically I mentioned the Baruch Plan. I should note that the Baruch Plan research has predominantly been done by my colleague Mike Page, so I don't want to present all of this as me having learned loads about it because that would be improper. But one of the things that I come away from there is that that was designed mostly by scientists and it was put together by like,
[0:48:57] Loads of scientists in different communities thinking sort of deeply about the technology and reasoning from that. And actually you saw that in the recent Iran nuclear deal where you had senior scientists, Ernie Moniz here, I forget the name of his counterpart in Iran, who sort of collaborated with each other as scientists to come up with a framework that became some of the policy deal. And I think that there is good evidence that you should...
[0:49:24] Use that natural competence there to a benefit. I think I'm agreeing with you. I don't think I have a specific reaction beyond saying, yes, that seems reasonable. Thanks so much, Joe.