Chelsea Follett: Joining me today is my Cato Institute colleague, Will Duffield. Will is a policy analyst in Cato’s Center for Representative Government, where he studies speech and internet governance. Much of his research focuses on the web of government regulation and private rules that govern Americans’ online speech. His most recent policy analysis paper was, Jawboning Against Speech, How Government Bullying Shapes the Rules of Social Media. I highly recommend it, but he’s also an expert on a host of related issues in technology policy, and he joins the podcast today to discuss artificial intelligence, including the ways that AI advances can improve our lives, as well as potential challenges, including for policymakers. Will, how are you?

Will Duffield: Doing well, thank you.

Chelsea Follett: Thank you for joining me. So let’s jump right in. Recent advances in artificial intelligence have been all over the news, ranging from DALL·E 2 to most recently, ChatGPT 3. It’s clear that advanced AI is becoming more accessible for a wide range of uses. My team now uses DALL·E 2 to produce many of the illustrations that are visible on DALL·E 2 is, of course, a program for creating AI art and ChatGPT is everywhere. You can see on Twitter what people have been creating with it. So could you describe the extent of the recent AI advances, and how significant are they really? How did these AI technologies differ from what we’ve seen before?

Will Duffield: Well, recently there have been a number of advances in natural language processing, and it’s essentially taking human speech and rendering it into some set of symbols and tokens that machines can understand. And the route taken here to achieve these advances has been through a process or approach called large language modeling or LLMs. And so essentially, by gathering together larger and larger corpuses of human language training data, as hardware costs have fallen, then AI pioneers have been able to derive ever more effective or perceptive AI models.

Will Duffield: Now there’s a lot of sort of maximalist versus minimalist arguments in this space. What exactly is artificial intelligence? What’s the threshold for artificial general intelligence or a kind of human-like machine cognition? Either something that’s as smart as human beings or that has a brain which functions similarly. And there’s questions there as to what would meet that threshold, how one might get to it. But even for AI skeptics, people who aren’t sure that we’re ever going to get to that machine thinking like a human, then these large language models, which take a very different cognitive process, they don’t reason deductively. Instead they infer results or information by synthesizing great reams of data.

Will Duffield: And so even if we see these as kind of mere curve fitters, things that are good at recognizing patterns, then they’ve gotten good enough that their pattern recognition is useful to us across a whole host of applications. And in some cases even really outstrips what we’ve been able to do. There’ve been studies throwing images of eyes at artificial intelligence, millions and millions of iris images. And while we don’t know of any difference between the male and female iris, this artificial intelligence has been able to, with a fair degree of certainty, determine whether the iris belongs to a man or a woman.

Will Duffield: So there it’s seeing a pattern that we don’t even know exists, never knew, can’t identify what exactly it’s picking up on, but it is picking up on something. Again, just with enough background information comparing enough irises in a way that you or I never could. By iris 20, they’d all just look the same to us, but a machine doesn’t have the same constraints. And so it can perform these really fascinating cognitive sorting tasks and make associations that might not occur to us.

Chelsea Follett: That’s incredible. And of course, one of the instances that’s receiving the most attention of advanced AI right now is ChatGPT-3. So I think that we should talk about that for a minute. Can you explain what it is and how it works?

Will Duffield: So ChatGPT as well as these image generation tools such as Stable Diffusion or DALL.E, are all what we might call generative AI. And they’re either text to image generation or they create new text on the basis of a prompt. But in all cases, they’re taking some user input, some request, and then creating out of, again, not in a one-for-one picking pieces from existing art, but taking art they’ve been trained on and the understood associations between what the user types in and what that might look like or has looked like elsewhere. And they create some wholly novel picture or text in response to that prompt.

Will Duffield: It isn’t, again, just carved up bits and pieces, but a new generation based on its understood weighted associations between the user input and its training data. So these tools are fascinating and finally getting into the hands of the everyday user which is where we can really see what they’re capable of. It’s one thing for the creator to build the tool, but we don’t even know what uses it can be put to until it meets the general public. And so that’s what’s been so exciting about the past couple of weeks or months is that across these different generative AI applications, we’re seeing what purposes people find for them or discover for them, how they can be used.

Will Duffield: And that will in turn then spur the development of a next generation of models, which are more fine tuned or tailored to the purposes people have come up with this time around. One area already being email drafting, drafting professional business like emails. People took ChatGPT, they threw it at this problem. It produced some pretty good results. There was one interesting story of a young man with some cognitive difficulty who ran a lawn mowing business, who just struggled or it took more time than mowing the lawns to come up with kind of business speak responses to the various emails he would get. But by putting a pretty rough description of the kind of email he wants to write into ChatGPT or a tool built off of it, it then spits out a professional business English version of that request that he can iterate on or send.

Will Duffield: And so the next time around, we’ll see some tool even more tightly tailored to that email response framework, the norms and heuristics of replying to emails, picking up upon what’s been said by your conversation partner and responding appropriately. And that next generation of tool will probably be even more integrated into the email client we use or the word processor we use to draft our replies.

Chelsea Follett: So obviously… Sorry.

Will Duffield: In general right now, I think a lot of these tools have gotten a lot of attention as standalone products because they are new, but in order to be really useful, then they’ll have to be integrated into the sorts of tools and workflows that we already understand. And we’re starting to see a little bit of this with stable diffusion plugins for Photoshop, making that stable diffusion image generation just another tool within Photoshop. I think that normalization or bringing these tools into more familiar settings will also help to dispel some of the displacement concerns that they’ve generated. Yes, there will be some disruption and displacement, but 10 years down the road, this is much more likely to be a part of an artist’s portfolio of tools rather than a replacement for him.

Chelsea Follett: So obviously we’re still learning what AI can do, but can you provide some other examples of how these new AI technologies are being used in the real world?

Will Duffield: Yeah, so my specialty and general area of focus is in this speech space. Even there, there are a lot of very exciting possibilities that can change the relationship between individuals and either institutions or the experts we contract for various services. You think of your architect giving you a set of drawings and you might not like everything about what you’ve been given, and right now your only option is really to tell him what you dislike and ask him to go back and do another set of drawings. And he’ll differ from the original about as much as he wants to. With something like stable diffusion or DALL.E, however, you could take his architectural drawings and iterate on them yourself, even without any specialized architectural drafting skills.

Will Duffield: And the same goes for a host of low level applications like the sorts of art that accompanies this podcast or my writing. However, AI is being used in lots of other less strictly speech related fields as well. One of the more exciting areas is drug discovery, using AI to combine molecules in unexpected ways or to predict the results of combining molecules together rather than synthesizing everything in a lab and figuring out exactly how it performs.

Will Duffield: A lot of that early stage testing can now be done virtually through or relying upon artificial intelligence. A little closer to the speech space, there are tools like Copilot at GitHub, which are essentially code generation or text to code generation tools that allow someone who might be a Novice coder to put in or explain the code that they would like to see or what they would like the code to do and then receive code generated by an AI program trained on lots of existing code out there as training data. And so it can help someone to learn or punch above their weight much more quickly, sort of turning to the AI in the same way you might turn to a more experienced coder. And obviously you need to check its work. You can’t but simply rely upon it. The same goes for kind of asking questions of GPT chat, but it can provide a really strong basis or give you a framework to iterate on that you wouldn’t be able to build yourself or you’d be one for one grabbing from someone else’s code and iterating on it.

Will Duffield: This large language model reinforcement learning approach also has applications in robotics. I think lost in a lot of the attention paid to GPT Chat in the past couple of weeks. Google released information about something called the RT1, which is a robotic arm controlled by an artificial intelligence model. So it can… As a result of this, rather than just being trained to perform particular tasks in a particular known environment, it can iterate and build on what it’s already been taught to do. Google employees went through and had the arm perform lots of different activities and motions just guiding it. But then weighted and rated those, how it performed.

Will Duffield: So it can then take that into account going forward when it’s asked to do slightly different things. So it has a lot more flexibility than a robot that either relied on a human to essentially operate it from afar or had a fixed set of responses to things that had been hard coded into it. Here the model still needs to stay fairly small because again, rather than following a strict set of instructions, it’s deciding how to react to each new prompt, how to fulfill it in the same way that GPT chat is thinking about how to respond to whatever you’ve requested of it, rather than just drawing an off the shelf response to whatever you’ve asked.

Chelsea Follett: You mentioned displacements earlier. How do you think advances in AI technology will impact the job market and what can individuals do to prepare for those changes?

Will Duffield: Well, I think it will be a little bit different than some past technological disruptions because so far it seems most likely to disrupt a lot of white collar knowledge economy work. It is essentially a labor saving device, but it’s a machine that can save us from a lot of tedious thinking. So everything from data entry, basic research, personal assistance sorts of tasks, basic legal assistance, increasingly is the thing we can imagine these LLMs tackling with a fair degree of accuracy, being able to give you the right statute or the relevant precedent. So unlike something like the power loom or cotton gin, we aren’t talking about displacing a lot of manual labor, but instead at the moment wrote white collar or knowledge economy tasks and increasingly some creative tasks. Though I think there’s been a tremendous amount of concern in the artist community about these generative AI tools.

Will Duffield: And while lots of people are using them, not all of that effect is displacement. And it’s something that it’s easy for people to, I think, get wrong when they see everyone with the new lens profile picture, but very few of those people were previously going out and paying an artist’s wages to create a profile picture for them. They were relying on a photograph or stock imagery, but they weren’t commissioning artists for that purpose. And so a lot of this so far fills in the parts of the market that weren’t very well served before. If you or I wanted art to accompany our article, it wouldn’t make sense in our budgets for us to go out and pay an artist for that. But it does make sense to come up with custom art rather than purchasing a stock image. So there the disruption occurs to the stock image company, Getty, but not to an artist. And obviously they rely on photographers art and all of that, but I do think that some of those effects at the moment at least have been overblown or at the very least people will adapt to these new tools. They’ll find ways to work with them rather than being displaced by them.

Will Duffield: Now, comparisons to something like machine translation, I think, do hold some weight. And there certainly are lots of low quality machine translations out there in the same way that artists worry about low quality art being produced. But there’s also a lot more machine or translations in general. And we can get real time machine generated closed captioning for media that previously would have just been completely inaccessible to those who couldn’t hear.

Will Duffield: And so in that way, the totality of art produced people who are able to express themselves through art will expand and it will be good for that to expand and grow. Just as even though translation quality on the whole might have suffered in the move to machine translation, it’s still good that those translations exist. And for those who want a particular portrait, it may be a new status symbol to have something done by a real artist rather than artificial intelligence. That part of the market will still exist in the same way that for particular translations, say something like you’re translating strategic documents you aren’t going to rely upon or at least you aren’t going to rely upon a machine translation for that. If it matters, you’re going to pay the money and get a professional.

Will Duffield: And as well as particularly on the art side when we think about disruption, machines at the moment, and DALL.E, GPT, they make a lot of little mistakes. And in art, physical 2D art, I think this is particularly visible. It’s not something that anyone then kind of edits out or changes in the way that if you get a GPT chat response with some mistakes, then you might fix it before using it in your own work. But while art created by people also have mistakes, we can understand those mistakes as products of the artist’s intention. They might have been going for something else and changed it midway. They might have been trying to present a figure in the style of someone else or in a particular pose that leads to them being contorted.

Will Duffield: There are all sorts of mistakes in famous historical art that nevertheless have meaning and allow us to more fully appreciate where the art is from, how it came to be. They give us a narrative within it. And when we look at the sorts of mistakes we see in machine art, it doesn’t have that quality. We can’t impute any meaning to it when someone’s hand comes out weird or with the wrong number of fingers. It’s just kind of off-putting. And so until these sorts of generative technologies are incorporated into suites of artists’ tools, then the kind of one-shot text-to-image generations are always going to lack a certain intentionality that gives rise to a lot of what we find compelling and appreciate in art.

Will Duffield: Now, again, that’s not to be said that across a wide variety of knowledge economy fields, there won’t be a certain amount of disruption and displacement. But at the end of the day, hopefully, we still need people in charge of these tools. So there are still jobs for human beings. And at its best, it can save us from kind of drudge work that, frankly, people deserve better than. It’s undignified for people to spend all day acting like a machine in a toll booth. And while thinking about how to give them some other gainful employment or something productive and enjoyable to do becomes an issue when they’re displaced from the toll booth, I think no one really laments them not having to spend eight hours a shift acting like a machine in a toll booth. And I think largely the effects of at least this style of artificial intelligence will be similar.

Chelsea Follett: And AI can, of course, create jobs as well. How do you think advances in AI will impact our daily lives outside of creating amusing profile pictures? Will it have an actual impact on our day to day existence?

Will Duffield: So the most, I guess, exciting possibility or hopeful possibility to me is that it will help us to interface with all of the systems we have built up in modernity to handle our affairs, but which we often struggle to understand, keep track of, and keep pace with. Everything from your county website, where you pay your taxes and deal with parking tickets, to payroll software, to booking a flight can all be very difficult, frankly, these days, and rely upon remembering lots of personal information and details, and essentially to interface with machines that often don’t seem very well designed to deal with us and the way we communicate. They may have been envisaged as labor saving devices in the first place to keep us from waiting on someone in a call center to get back to us, but they can be very frustrating as we have to deal with them or accept their decisions and the way they can order our lives. And so in the most hopeful world, AI will be an everyday interface between human beings, the way we communicate, act, express our desires and wants, and the various systems we’ve built for satisfying those wants that don’t tend to work very well with us.

Chelsea Follett: Now those are some of the things you’re most excited about, but what are some other benefits we might expect to see from artificial intelligence advances, do you think?

Will Duffield: Well, I think to an extent it depends upon what artificial intelligence we end up with. For those who envision artificial general intelligence, the results are world changing. All sorts of things would really be able to be shifted, new frontiers in science, nanomachines, that sort of thing. When you have a world in which everyone has basically unlimited brains to throw at any problem. Now, that’s really the AI maximalist position and there are a lot of concerns about that world and how such an AI might behave and we can get into that in a moment. But I do think the world of everyone having fairly useful personal assistance across a wide variety of pursuits or specialties is fairly realistic and one that we’re taking steps towards. It’s one in which we can have a great deal more agency. Our goals will be more achievable because we’ll be able to rely on this help to get there. And a lot of the tasks which right now we rely on others for, can instead be done by machines.

Chelsea Follett: So those are some of the different benefits. What about the challenges that are posed by artificial intelligence advances of the kind that we’ve been seeing?

Will Duffield: So I think there are two groupings of challenges that people have worried about as these machines have advanced. And the first is the alignment problem. Through artificial intelligence or artificial and general intelligence would essentially be an alien intelligence that we’ve created. And we’ve always been concerned by the idea that any aliens we meet might have different values and goals than we. And given how we’re training or building what might become artificial and general intelligence, we aren’t trying to perfectly model the human brain, but instead create free floating connection machines. And so it’s likely or hypothesized that whatever comes out of that may view the world differently than human beings. And getting its values and goals to align with those of humanity is seen as a concerning challenge. So people will bring up the so-called paperclip maximizer scenario. Even if you have an AI that is ostensibly aligned with your goals and you set it to make me a lot of paperclips and it sets out to do it. If it’s very powerful and can tap into financial markets and have factories built and draw up resources, it will slowly start making everything into paperclips.

Will Duffield: It’s prime directive. It’s just making paperclips. And once you’ve strip mined the whole earth, too bad for the people and animals and all of that, and then you start getting out into space and you might have a very efficient machine space program that’s just bringing resources in to turn them into more paperclips. That would obviously be a really bad scenario for everyone. That machine would be out of alignment with us because it would displace us in its pursuit of ever more paperclips following blindly this kind of directive.

Will Duffield: So even, it’s not just thinking of the genocidal evil machine that wants to destroy all humans, but just how to put safeguards or side constraints on even what it is ordered to do so that it will not take drastic or harmful steps in order to get there. And I think that big alignment question, which is again very tied to this question of AGI and machines that might deceive us in the pursuit of their goals or are capable of doing so, then it raises the question of aligned with who? Just as there isn’t any one singular public interest, there probably isn’t a singular interest of humanity beyond obviously not being destroyed to make way for more paperclips.

Will Duffield: But once we drill down, then we end up with all these more niche or specific alignment questions that implicate different values. And a lot of those are subsumed under the heading of human misuse. When we think about people then putting AI up to tasks which harm others or at odds with the goals of other people. If AI is a very effective personal assistant, then it won’t just be an effective personal assistant for good pursuits or good people. And there’s a lot of debate right now about how to either regulate at the government level or place side constraints on these models themselves.

Will Duffield: And we’ve seen some of that already with GPT Chat, with people figuring out prompts that would break it out of its box or allow it to query the internet or tell conspiracy theories in ways that it wasn’t supposed to. And OpenAI, its creator, limited some of that. And now in response to those sorts of prompts, it’ll give a kind of dumb, I’m just an AI, I’m here to have conversations and I can’t pretend to be a conspiracy theorist for you. I’m sorry. But beyond this speech realm where obviously there are harmful sorts of speech, there is deceptive speech. So a lot of concerns with these generative AI tools and how they might be misused have related to propaganda and misinformation, creating images or videos that appear to be real or of real events, but are in fact not.

Will Duffield: I’m not quite as concerned about this as a lot of people are because most of those frauds rely upon context in order to really deceive. It’s not the image itself, but the context in which we receive it. And at the moment, if you’re a nasty person or a government especially with lots of money to throw at the problem, you can make very convincing faked video. But unless it’s delivered in a believable context, then it just doesn’t sway minds. Earlier this year, someone, most likely the Russian government, released a fake Zelensky video in which he appeared to tell his men to lay down their arms. But given that just days before he’d appeared on the streets of Kiev and encouraged everyone to keep fighting, it just didn’t make sense to anyone that he would have this about face. And so it wasn’t believed. And I think the same goes for a lot of AI or generative AI fueled deception.

Will Duffield: However, outside of this speech realm, then I think there are real concerns. Just when we think about the kind of capacity or agency enhancing aspects of artificial intelligence. We’re very lucky right now that most of the people who choose to become terrorists or violent extremists just aren’t very bright. And they aren’t very good at carrying out their goals and objectives. But in the same way that occasionally the FBI will give a bomb or fake bomb to some bumbling would be terrorists who probably couldn’t manage it on their own before swooping in to arrest them. We can imagine a friendly artificial assistant helping people who otherwise couldn’t build a bomb to build one. And it’s not going to arrest them afterwards. It’s just helped its user carry out a task. And so the lack of moral side constraints on this is a cause for concern in the same way as a lack of moral side constraints is a cause for concern in intelligent people. And in many ways, the brighter you are and the more capable you are, the more worried we are if we find out that you’re associate path and you don’t care at all about how your actions affect others.

Will Duffield: And so getting that right is important, but it will also be very important not to either neuter this technology or the mass available version of this technology or gate it away from the everyman allowing only the kind of vetted or powerful access to it, because then you’re just going to exacerbate existing disparities in wealth and power. And so to me, even though there are these potentials for misuse and harmful misuse, many of those cases seem worse in a world in which only some people have access to the technology.

Chelsea Follett: That makes sense. How do you think policymakers should go about regulating this technology without stifling innovation?

Will Duffield: So I think at the moment it’s very early stage and frankly too early to be doing much regulating because we have to discover both how people would like to use this, how it can be used in good ways, and unfortunately how it will be misused. And as much as we can predict particular misuses, there will be new ones that we didn’t anticipate. It goes for every technology. It’s a disruptive learning process, but I think to an extent, we need to go through that and in a way go through it earlier rather than later, because many of the lessons that we will learn from the current generation of generative AI tools will apply to future advances in the space can be applied. And so I’m very skeptical of proposals say to treat this as a dual use technology under kind of military export control regulations, in the same way as 3D printed guns were seen as dangerous such that they shouldn’t be released online where anyone can get them.

Will Duffield: I think the consequences both for the growth of this industry, but also for people’s creative capacity of trying to tamp down on this in a similar way would be really destructive and harmful. And that frankly policymakers should recognize that at least in this generative AI space, these are speech tools. That it’s a new expressive tool doesn’t make it any less a part of the suite of first amendment protected technologies that we call the press. When the first amendment was written, the pen and the printing press were the only technologies that allowed a speaker to record their ideas and send them beyond the range of their voice. But as new expressive technologies have been invented, courts have understood the first amendment protects them just as it did or does the printing press. And so to the extent that these are that we’re talking about expressive tools here, we need to recognize that that expression has legal or constitutional value. And taking it away from people would not just be wrong, but illegal.

Chelsea Follett: What impact do you see over regulation having on the development and adoption of AI technology?

Will Duffield: Well, I think it has the potential again to really limit who can access it. If we’re talking about regulated development, it looks much more ivory tower. And certain firms have already been criticized for keeping their models closed, not allowing others to iterate on them. That’s fine if it’s their private decision. Others have entered the market with kind of explicit open source focus in order to provide AI to the masses. But if it were regulated in order to really prevent any of these misuses, then that would likely look like restricting access rather than restricting model outputs by law.

Will Duffield: And I think that takes you towards or starts you down the path that we’ve seen with lots of other technologies, which perhaps because of the risk of misuse or danger are developed mostly within government and academia. And in many cases there, then government purposes, which are often quite bellicose, you look at the history of nuclear energy development here, end up dictating what future advances should be rather than the commercial use cases, which are often much more Pacific. And so I think that closed AI is not just bad because of the kind of ivory tower aspect and lack of democratization, even though those are real harms, but because closed AI is more likely to be military AI, frankly.

Will Duffield: When you look at where then the funding will come from, where the interest comes from. And I think that’s an unlikely scenario. Frankly, there’s just not enough will to regulate it strictly right now because it is new and there are more pressing issues for legislators to deal with. But I think that very much would be the worst case scenario in which this is something that bad foreign governments and military have, and it doesn’t do much to improve our daily lives, allow our car to drive our kids to school on its own or pick our dog up from the sitter. Those sorts of everyday purposes, making a custom birthday card for your grandmother, your kids being able to do that, then where we are now.

Chelsea Follett: And are we seeing any premature attempts to legislate regulation of these technologies? Or is that fortunately not yet?

Will Duffield: There have been some. Europe is pushing ahead with a code of ethics. In the US, we haven’t seen much substantive regulation, but there was a rather concerning letter by Representative Anna Eshoo in the fall deeming open source models, that is those that can be modified and re-weighted by users, but aren’t necessarily aren’t controlled on a platform the way, say, DALL·E is, as unsafe. And if this is the threshold for unsafety, then there’s no real way to democratize AI. It must be closely held within a platform model. And so I think that’s really the wrong threshold to draw. At the end of the day, a responsibility for misuse should rest with the user. They’re the one employing the tool after all. And if someone takes an open source generative AI model and re-weights it or retrains it in order to produce, say, child sexual abuse imagery, which is one concern, then the responsibility and fault should fall on them. They’ve done something illegal. They should be prosecuted. In general, I don’t think trying to restrict or put these sorts of controls in place at essentially the printing press level is the way to go or a workable path forward.

Will Duffield: It’s not the standard we’ve set for any other sort of technology. Again, you’re allowed to sell a printer without it having to ensure that it prevents people from printing off government secrets or abusive imagery. VHS was allowed, even though it had the capacity to illegally copy tapes. And so we should follow the model that we have in the past with other sorts of publishing technology and ensure that responsibility for misuse rests with the user rather than the tool or tool creator.

Chelsea Follett: Have any of the questions that I’ve been asking you, have they struck you as out of the ordinary in any way?

Will Duffield: I wouldn’t say so.

Chelsea Follett: All right. Well, as an experiment before this conversation, I asked ChatGPT what I should ask you. Although I’ve varied the wording and so forth, the ideas for most of the questions I’ve asked you were generated by the AI and that’s fascinating that…

Will Duffield: Oh, wow. I failed the Turing test, I guess, though.

Chelsea Follett: I’m not sure. I would have been able to…

Will Duffield: As presented by you.

Chelsea Follett: No, I’m impressed by how they were.

Will Duffield: How curious. Yeah. That’s even the regulation question and all.

Chelsea Follett: Yeah. There are a few that were mine, but most of them were from the AI. This has been fascinating and it’s certainly an exciting time to be alive with all of these advances and their potential to make our lives easier. Thank you so much for speaking with me.

Will Duffield: Indeed. I’m excited to check in in another six months and see how far we’ve come.