Dan Williams is an Assistant Professor of Philosophy at the University of Sussex. He writes the Conspicuous Cognition newsletter, which brings together philosophical insights and scientific research to examine the forces shaping contemporary society and politics.
In this week’s conversation, Yascha Mounk and Dan Williams discuss whether AI is a genuinely transformative technology or just another tool, how the temptation to outsource thinking to AI may be eroding the cognitive skills of young people, and why the political battles we are actually having about artificial intelligence are not the important ones.
This podcast was recorded at a live event in partnership with Halcyon International School.
This transcript has been condensed and lightly edited for clarity.
Yascha Mounk: Welcome to this live episode of The Good Fight Podcast, coming to you from the Halcyon International School in London. Dan, I think one of the interesting things about artificial intelligence is just that it’s going to transform so many different aspects of our society. As I was preparing for the conversation today, I was wondering what I should ask you about. Should I ask you about existential risk? Should I talk about the socioeconomic impact of this? Should I talk about the thing you’ve thought a lot about, its impact on the public sphere and on misinformation? How do you think about what aspect of artificial intelligence to write about in your work, and how should we prioritize between them, or do we just need to be thinking about all of these impacts all of the time?
Dan Williams: I think we need to be thinking about all of the impacts all of the time. I think we are in the early stages of a huge technological revolution, the magnitude of which will be at least that of the Industrial Revolution, happening in a much shorter timeframe. I think it’s going to impact, as you say, not just the economy, not just the culture, not just our psychologies, not just our educational institutions, but the information environment, our social relations, and also all of the interactions between these different things.
So I think as a subject, it doesn’t really lend itself to hyper-specialized inquiry where you try to bracket off everything else. If you’re looking at the economics of AI, you are inevitably going to start asking questions about the impact of AI on our political institutions, on our political culture, on public opinion. Likewise, if you’re thinking about how AI is going to impact the information environment, the way in which we try to make sense of society, you need to be thinking about the economics of AI and the business models of these frontier AI companies.
I don’t think this is unique to this moment. If you go back to the early stages of the Industrial Revolution, precisely because it was this monumental technological revolution that changed every aspect of society in innumerable ways, it was this great era for polymaths, for people that looked not just at narrow, specialized fields, but huge questions straddling lots of different disciplines. I think that’s kind of how you have to approach AI today.
The flip side of that is precisely because these issues are so big and so complicated, and the technology is affecting so many different parts of society all at once, there is enormous uncertainty at almost every level. It’s very difficult, I think, to have any high degree of confidence about how this is going to unfold, which I think is one of the reasons why you see such spectacular levels of disagreement even about relatively basic things.
Mounk: The comparison to the Industrial Revolution is helpful and interesting. I’ve been thinking through that comparison in the context of a different set of work. But in 1800, should people have worried about the urban immiseration that might follow from the Industrial Revolution, at least in its initial stages, before helping the world get a lot richer, the kinds of things that Charles Dickens wrote about, David Copperfield, etc.? Or should they have worried about the possibility that it would revolutionize warfare and make possible the trenches of World War One? The answer is, well, we should have worried about both, and those two things end up being related in some really complicated but real ways.
So in a similar kind of way, you’re probably right that we shouldn’t choose between focusing on this or on that aspect of AI. We should be thinking about all of them at the same time. I do think, though, that there’s a kind of strange way in which some of the fights over AI feel to me like they’re not the most important ones.
The political issue that’s gotten the most traction is whether or not to build data centers. There’s a story to be told where perhaps you’re worried about existential risk, which is to say the possibility that AI might kill us all one day, and if there are fewer data centers, then it’s going to take longer to develop more capable AI models, and perhaps take longer to build robots that are actually capable of taking physical action. So opposing data centers might be a way of somehow lessening existential risk.
But I don’t think that’s what’s driving people. What’s driving people is: I don’t want this thing next to me in my community, which to me feels like it’s the one issue around AI that’s gotten any real political currency, because it can latch onto existing political fights, like whether or not to build housing. It’s just a way for NIMBYism, not-in-my-backyard-ism, to latch on to the issue of AI, and that’s why it’s the one AI issue that’s taken off. Do you share that sense, where the political battles we have about AI are not the important ones, or am I being insufficiently charitable here?
Williams: I completely agree with you. I would say, in some sense, it’s understandable. There is this huge data center build-out, and many people don’t fully understand what motivates it, and it’s led by tech companies that, to put it mildly, are not very popular in many cases. But I do agree with you that in some sense it’s not really grappling with just how significant the relevant changes are going to be.
We hope you’re enjoying the podcast! If you’re a paying subscriber, you can set up the premium feed on your favorite podcast app at writing.yaschamounk.com/listen. This will give you ad-free access to the full conversation, plus all full episodes and bonus episodes we have in the works! If you aren’t, you can set up the free, limited version of the feed—or, better still, support the podcast by becoming a subscriber today!
If you have any questions or issues setting up the full podcast feed on a third-party app, please email leonora.barclay@persuasion.community
I also think it’s a little bit of a shame, in a way, that there’s this polarized conversation where, on the one hand, you’ve got people who think AI is just another technology, a normal technology in some respects, and so you can treat it just as you would any other technology. Then there’s this whole conversation, which we might get to, about AI and existential risk: is it the case that superintelligent, rogue, misaligned artificial intelligence might kill us all?
What I would say is those are not the only two frames. I think what the second frame gets right is it does grapple with just how powerful this technology is. We are in the process of building thinking, understanding, agentic machines that are going to radically transform the world. We shouldn’t be completely dismissive of the worries about misaligned, rogue, superintelligent AI killing us all. I think, for various reasons that we might get into, ultimately the arguments for that scenario are not very persuasive. But even if we avoid that specific scenario and we have aligned superintelligent AI systems, that is, AI systems that are more cognitively capable than human beings, at first disembodied AI systems and then autonomous robots, a world with that technology is going to be radically, radically different from the world as it exists today. There is nothing normal about it. In no respect can you think of that as just another technology. I completely agree with you that the politics of AI at the moment haven’t caught up with that situation.
Mounk: For a little bit of context, there’s a paper that has been very widely discussed in AI circles called AI as Normal Technology by Arvind Narayanan and Sayash Kapoor, where they argue that in order to understand how AI is going to roll out, you have to look at previous technologies. It turns out that many of the predictions of the leaders of the AI field, these two authors argue, are really naive, because they don’t understand real-world obstacles to the adoption of technology. You might think that the moment AI is in theory capable of translating any text, publishers are going to use AI to translate those texts. But in practice there are 17 million obstacles to that happening, and practically no publisher has started doing that yet. The same could be true in lots of different fields.
I have two points about that. The first is that there’s a great debate over this paper that was conducted recently via music videos, the song and the animation of which was produced by the new Claude Opus 5.5 model. So you have a great “paper” explaining why the robots are not going to roll out in the real world because of all of these obstacles, and then you have a great response to it called “Thus Spoke Compute,” arguing the opposite. It seems to me that the existence of those two videos is, in a way, proof that AI is not a fully normal technology.
But I take seriously some of the main themes in this paper. The other point I want to make is it doesn’t actually seem to me, if you read the paper carefully, that it’s necessarily at odds with those worries about existential risk. The main point of the paper is about how long it takes for a technology to diffuse into the economy, for existing firms and companies to change how they do their business. All of that might be true at the same time as there’s a real risk of us losing control over superintelligent AI.
I was hesitating a little bit in this conversation about whether we should go from immediate and comparatively small to distant and large, or the other way around. But since there were really interesting remarks by students, for the benefit of the podcast audience, explaining, for example, that they were concerned AI might be eroding their cognitive skills, that the temptation of going to AI whenever a problem gets hard may make it harder for them to really learn and develop the cognitive skills they’re invested in, perhaps let’s start with that. Do you share that worry, and what evidence do we have that AI may be undermining cognitive skills in that kind of way in the younger generation?
Williams: I absolutely share the worry. Before I address it, I’ll just double-click on one thing, which is that I have enormous admiration for Arvind Narayanan and Sayash Kapoor, the authors of that AI as Normal Technology paper. I agree with you there’s a lot going on in that general framework that they are outlining and defending, and you don’t have to accept all of it as a package deal. Certainly one of the aspects of it that I think is correct is that to understand how AI is going to affect society, politics, institutions, and so on, you can’t just focus on the raw capabilities of the AI systems. You have to think about how organizations, institutions, culture, incentives, and so on adjust to the technology. I think that’s highly relevant in the context of AI and education.
The cliché thing to say, which I think is true, is that AI is the best technology ever invented for learning and also for avoiding learning. Which of those two outcomes plays out in society won’t be determined just by the technology itself. It will be determined by the degree to which our educational institutions can adapt to this radically new technology. I think it’s fair to say that our educational institutions, whether schools or universities, were not designed, and they didn’t evolve culturally, in a world in which anyone has access to an incredibly smart, understanding, conversational machine. I think it’s completely understandable that our institutions are trying to adjust and think about how they can adapt in light of this new technology.
I think that’s a huge challenge. One of the real significant society-wide challenges with AI is that it’s so good at so many different things that we outsource so much of our cognition, our thinking, our decision-making to the AI systems that those skills really atrophy. I’m not really that worried about this when it comes to adults, because in a liberal society people can make their own choices. But when it comes to children, teenagers going through school, this is potentially a huge crisis. I think we really need to take this very seriously at all levels of society, as an enormous policy challenge.
Just one final thing: I think part of the issue is there’s still a lot of denialism about the technology itself. It’s still the case that lots of academics, lots of intellectuals, I’m sure to some extent it’s true of teachers and people who work in schools as well, aren’t familiar with, don’t understand just how impressive the technology can be.
Mounk: I was an undergrad at Cambridge, I live in Cambridge, and I was just reminded of an old joke about Cambridge: how many Cambridge fellows does it take to change a lightbulb? The answer is: change? That, of course, is not just true in Cambridge but in academia in general. It’s often the academics who most claim to want to change and revolutionize the world who are most resistant to any form of change in their own university or their own workplace. I’ve been astounded by how many professors make excuses to pretend that the assignments that have been assigned for their course for 27 years still work in the age of AI. They say, “AI could never do the assignment for my course. It’s too specific.” It’s like, of course it can.
But let me ask a brief preliminary question. In your teaching, have you found that the quality of students has changed? Are you seeing the signs of a great reading crisis? There are a lot of articles about this, and they all go viral, with people saying, “I used to be able to assign a whole book, and now my students can’t read two and a half pages anymore.” I will spike a football, as we say in the United States, by saying that in my teaching I haven’t found that. I’m very privileged to teach great kids at Johns Hopkins, a very selective university. I don’t find that they’re less capable of doing the reading or less engaged with the material today than they were five or ten years ago. But that’s one kind of anecdotal piece of evidence. Please give us a second one.
Williams: I’ve only been a full-time assistant professor doing lots of lecturing since 2023, which is when generative AI started being rolled out, and many universities still assign take-home essays. I remember saying in 2023, this is absolutely insane, because at the time AI systems were already able to produce relatively high-quality student essays. Now I think it’s the case that frontier AI systems can produce higher-quality academic papers than many actual academics, and we’re still doing take-home papers.
Mounk: Nowadays, the last time that I assigned take-home exams, I said I would never again. When I got a bad paper, I was sort of like, “Nice, the student actually did it themselves.”
What do you think, how do you think we should redesign the university, and perhaps the school, since we’re in a school here? It strikes me that there are two competing things here. We obviously want to make sure that people are actually reading text. Disagree with me if you think I’m wrong, but we still want people to learn how to write, because even if you can outsource that to LLMs in the future, that to me seems like an absolutely crucial step in learning how to think and reason carefully. At the same time, we also now have available to us individualized one-on-one tutoring in a form that just wasn’t available to everybody in the past. Presumably we want to use that in some kind of way.
I speak pretty good French, but I don’t write French very much, and I realized I’d totally forgotten the rules about when to make a past participle agree with the subject. I recently asked Claude to re-explain the rules to me. The basic rule is quite simple: if it’s with the verb être, “to be,” then it starts to agree with the feminine speaker, for example; if it’s with the word avoir, then it doesn’t. But then there’s a bunch of specialized rules. It started with a simple test, a really simple rule, and then added the more complicated rules. It was just a super-efficient way to learn what I’d learned many years ago but had sort of forgotten.
Clearly that form of one-on-one tutoring is something that can be hugely impactful. But how do we make sure that that doesn’t just turn into “go off and play around with Claude” and we pretend that’s education?
Williams: I think the first step is having assessments, in-class assessments where you can’t use AI for the assessment itself, whether it’s a written exam, an oral exam, or whatever it might be, so that you are rewarding the student for being able to produce the work. Of course, if they want to use AI, as you say, for the process of learning, I think it can be an extraordinary tool for the acceleration of learning. That’s absolutely fine.
At the same time, it just is the case that AI is going to be increasingly embedded into every facet of society. One of the things we need to be thinking about as teachers and educators is how to prepare students to use AI as well as possible. I suspect there needs to be some kind of divide, where you’ve got assessments for core skills where, in the assessment itself, they can’t use AI, and assessments where you actively encourage as much and as savvy a use of AI as possible, something like that. But that’s really the best I can offer. I think it’s a huge question with lots of uncertainty.
Mounk: That’s pretty much what I hit on in my own teaching. Ideally I think that has to happen at the level of the university over the course of a degree. It doesn’t have to happen in every course. But since universities haven’t really adapted to that yet, I have one important element of assessment, a pen-and-paper exam in class, that I’ve never really done before in my teaching career, but that’s the only thing that ensures we’re not using AI to do it. Then I have another element where we either have to write a really ambitious paper for which we’re allowed to use AI, or produce a cool kind of resource. It’s a class on democracy and the rise of populism, and I said, look, you can use every AI tool you want to make a cool video about this, or to make a really informative website, or an app, or some kind of resource. You’re welcome to, and in fact encouraged to, use AI. You’re learning how to use AI tools to communicate some interesting set of ideas. But that’s only part of the grade. The other part is you’re sitting there with pen and paper, and you actually have to make an argument.
But I guess the question goes beyond that, which is: what should learning look like for a 15-year-old? What should learning look like for a 17-year-old? Is it that you have a lot of time, or some time, spent with something like an AI model for basically one-on-one tutoring, where you get really instant feedback, but then at the same time that means you can have a lot more time for really in-depth discussion? Is it project work? The one thing that seems pointless, and perhaps always has been pointless, is what ninety-eight percent of my schooling was like: a teacher standing up front and talking at you for forty-five minutes while you’re half asleep.
Williams: That’s always been difficult to justify as a practice. Speaking personally as a lecturer, it’s great if you’re the lecturer. I’m not so sure it’s so great for the students.
Mounk: The only skill in my life that I’m truly proud of is that I was able to sleep in high school, and teachers would get annoyed with me and say, “Yascha, did you fall asleep?” I’d say no, and they’d say, “What did I just say?” I was always able to repeat the last two sentences to them, verbatim, and it drove them crazy. So if you want to annoy your teachers, this is a skill I recommend developing.
Okay, so let’s go one level up from that. At the beginning there was high hopes about social media. If you go back to all of the writing about social media in the 2000s, and even in the early 2010s, it was all: it’s going to allow the opposition and young people to challenge these dictatorships in the Middle East and other parts of the world; it’s going to take on the rich and the powerful in the United States. Look at the Green Revolution in Iran, look at Occupy Wall Street. There was a lot of celebration of it. Then, obviously, our politics took a bit of a turn, and people started blaming social media for that. What explains the rise of Donald Trump? What explains Brexit? What explains all of these different developments? Well, it’s got to be social media.
So I feel like people are entering the conversation about artificial intelligence with the battle scars of the last set of fights and wars, and they’re assuming that just as social media turned out to be very bad for our public sphere, so too AI is likely to lead to the rapid spread of misinformation and disinformation, and we’re going to get even more cynical about the world, and all the bad things are going to happen. You disagree. Why?
Williams: I think if you look at social media and its effect on politics, it’s had two distinctive characteristics. The first is that it’s been a profoundly democratizing technology, in the sense that its most distinctive characteristic is the complete obliteration of establishment gatekeepers. With the internet, and then with social media, anyone with an internet connection and strong opinions can get online and start posting, expressing their opinions, and getting involved in public debates.
You’re right that initially there was a lot of optimism about this. I think a lot of that optimism was, and remains in some way, warranted. This is a genuinely great feature of social media. It’s also had lots of really deleterious consequences, because establishment gatekeeping sounds bad, but it also means some minimal standards of quality control. Once you get rid of that gatekeeping altogether, you get a public sphere where you get the expression of views which have always been popular among the general public but were previously marginalized: highly conspiratorial views, extremely bigoted views, quack science, and so on. Now, as a consequence of the democratizing character of social media, that plays a much bigger role in our politics than it used to. I think that’s closely connected to populism.
The second is the role of engagement-maximizing algorithms. The way these social media platforms recommend content, and keep people scrolling and glued to their screens, is by recommending content that engages people, that grabs their attention, and that often means rewarding, amplifying, sensationalist, divisive, conflict-oriented, generally quite stupid content.
So it’s had these two characteristics: it’s a democratizing technology, for better and for worse, and I think it’s generally dumbed down the state of public discourse. Many people thought you could take that kind of model and just assume that many of the pathologies associated with it, as you mentioned, fake news and misinformation, would just generalize to generative AI and large language models, and it would just be those problems on steroids. I think nothing like that is true. In many ways you should think of large language models as the anti-social-media. Whereas social media has been democratizing, it’s shifted power towards the communication styles, the beliefs of ordinary people. Large language models generally give people accurate, evidence-based, expert-aligned information. Not perfect, there are still problems with hallucinations, and we can talk about bias, but compared to the realistic alternatives, generally much higher-quality, expert-aligned information. Whereas social media, because it has this engagement-maximizing character, tends to amplify this sensationalist, conflict-oriented style of communication, at least at the moment, Gemini, Claude, ChatGPT, even Grok, which is Elon Musk’s large language model, tends toward a pretty dry, respectful, accuracy-oriented style of communication.
At least initially, you might be quite optimistic about that, insofar as it looks like it will, to some degree, correct some of the pathologies associated with social media. In fact, I think what’s going to happen is it’s just going to create a whole new set of both beneficial consequences and dangers, just very different from those we’re used to with social media.
Mounk: I agree with a lot of what you just said. I think, in particular, the assumption that somehow this is a new technology and it’s going to play out in a very similar, parallel way to this older technology is very likely to prove wrong. I also agree with you that, certainly at the moment, if you ask the counterfactual question, what kind of information are you likely to come across on whatever AI chatbot you use, relative to what you’re likely to come across if you open X, or Instagram, or TikTok right now, the relative quality of information you get is going to be a lot better.
At the same time, I guess I’m a little worried that the takes about the nature of social media that people had in 2005 or 2010 were the result of a more optimistic moment, but they’re also the result of what social media looked like at that time. Social media at that time was a lot less toxic than it became fifteen or twenty years later, and it may be that, in a similar way, the way we think about the standard use case of AI today is quite different from what it’s actually going to turn into being ten or fifteen years from now.
Let me start with the following question, to pressure-test it a bit: are people mostly going to get their information from AI chatbots, or are they mostly going to get it still from social media? To what extent are these AI bots actually rivals to social media? Secondarily, what use are some of the most divisive, toxic accounts on social media, what some people call conflict entrepreneurs, going to make of generative AI tools on those forms of social media?
If you assign me two worlds, one in which everybody gets most of their political information from chatting with Claude or ChatGPT, and one where everybody’s spending time on social media, I would agree with you that the first world is likely better. But if you say, no, perhaps sometimes we go to Claude or ChatGPT to ask about things we’re really motivated to find out, like what’s the best restaurant around here, but most people aren’t that interested in politics and don’t seek out political information in this proactive way, they get political information in this pretty passive, ambient way, then they’re still likely to get that on social media in the future. Except that, in the future, perhaps people who want to push force-narratives have much better tools available to them to produce persuasive content, and the ability of social media to hook us may be even bigger. To name one obvious example, not only does Instagram or TikTok now have an even better algorithm to serve you the most addictive possible video out of a universe of existing videos, they can create, on the spot, a new video designed to keep you hooked, in keeping with what kind of video you keep clicking on.
So when you think about how AI might displace social media, I think you’re right. But when you think about how AI may combine with existing social media, the impact may turn out to be a lot less positive.
Williams: You’re completely right that this more pessimistic scenario might unfold. I think the thing to remember is these things are not preordained. I think we have agency, both as individuals, more broadly as a culture, and then also in terms of policy, to try to address this, to maximize the positive consequences of the technology.
Clearly what’s going to happen is that they’re not complete rivals, in the sense that they’re not totally separate technologies. It’s already the case that on social media, LLMs are, to some degree, becoming embedded in different ways. One way is on the platform X, Elon Musk’s X, formerly Twitter: “Grok, is this true?” as a response to what people are saying or posting has become, to some extent, more normalized.
Mounk: As you said, despite some concerns about the nature of Grok, by and large the answers you then see from Grok are pretty well sourced, and they’re often a way of checking things that are exaggerated, untrue, etc. Somebody says, “Is it actually true?” and Grok says, “No, that’s not really true, that’s taken out of context.”
Williams: Exactly. You can think about more innovative ways of automating fact-checks with AIs. I also think, when we’re talking about potential threats of AI, it’s really important not just to focus on the negative use cases. You’re right, in a sense it can significantly automate the creation of propaganda, but it can also be used as a way of critically evaluating the kind of communication you’re coming across online. As that becomes normalized, provided these tools do give broadly reliable information, as I think they currently do, not perfect but compared to realistic alternatives, then content creators, even the very biased ones, will learn that they’re in a world where anyone can very quickly fire up their chatbot and critically evaluate or fact-check what they’re saying. So it creates this whole new set of incentives.
I think it’s also a very interesting question how trust in these chatbots is going to develop over time. You might hope, and many people have hoped, that ordinary people would place more trust in experts. This is obviously a huge political issue over the past ten or fifteen years, a very polarized one, with expertise itself becoming enmeshed in these broader political conflicts and culture wars. But expertise, up until about five minutes ago, has always existed in human form. Now we have these strange machine intelligences that seem at least capable of giving access to expert-aligned information, and that, at least in principle, needn’t have the kinds of biases and tribal identities that human beings tend to cling to.
Moreover, they’re not just going to be giving people information in the context of politics. People are using chatbots for all kinds of things: for medical advice, for fashion advice, to navigate bureaucracies, to get tips on how to cope with their relationships, and so on. To the extent that they’re building up trust in this source of information across all of these other domains, I think it will be quite natural to assume you can also trust them when it comes to politics as well. Provided the current situation continues into the future, and that’s a big if, that could be quite a positive, optimistic story.
Mounk: Let me push back on that a little bit. I’m trying to think through how that is going to look, and whether people really are going to be willing to outsource judgment in some ways to these AI models and perceive them as in some way neutral. In certain contexts I think that’s right: “is this true, Grok” is something that really does happen on social media, and I think by and large it has positive effects. But you might have two reasons why that might not be the case.
One is that there are real fears, on both sides of the political spectrum, that AI bots are systematically going to be politically biased. In one of the first papers that really sounded the alarm about the impact of AI in various ways, a set of authors argued a number of things, including the idea that these are just “stochastic parrots” and not going to be very capable, an idea that has become a kind of meme in the discussion. They were also saying, look, the kind of texts these models are trained on are very Western, very male-dominated, perhaps politically aligned with capitalism, and so all these models are going to be discriminatory against minorities, against women, and so on. Interestingly, the best evidence we have so far, and you can expand on this, you know the literature as well as I do, perhaps better, seems to suggest that’s not the case.
For example, if you force virtually any AI model to make a choice in a pairwise comparison between two states of the world, you can say, imagine a world where one white person is killed, or imagine a world where two Asian people are killed, or whatever it is, or the other way around, or you say one each, there are different ways of setting up the comparison. The point is that it actually ends up favoring the non-white lives over the white lives. This is a very consistent finding across different kinds of models. You can also ask general political questions, what kind of policy response do you think is right to the debt crisis, or something like that, and it tends to give very consistently center-left answers. So you could have a world in which actually both sides of the political spectrum feel that the AIs are biased against them, even if that’s not true, which I think you often get in moments of censorship: everybody feels like my side is being censored but the other side isn’t, which I think is one of the problems with a regime of censorship.
But the other point is that there are, I think, real questions about where this political bias comes from. My assumption was that it tends to take center-left positions, roughly what my own set of positions happens to be, because when you want really high-quality text on which to train these models, a lot of it ends up having a slight center-left bias, because that’s the nature of a lot of journalism. There’s a reason why you’d want to train an AI model on the archive of The New York Times, but The New York Times happens to have a slight center-left lean.
Interestingly, I spoke to a very senior Silicon Valley figure a week ago who is adamant that he believes that’s not the case, that actually it’s all post-training. It’s not the selection of texts on which you train the AIs; it’s that once you’ve done all of that and you have a kind of raw version of the model, you have the AI give you a bunch of outputs, and you have a bunch of humans who rate the output. The reason for that is to make sure it doesn’t go off the rails and do immoral things, or swear at the user, or use terms that would be inappropriate. The problem is that all the people who’ve been recruited to do this are graduate students in the United States, they’re the smart people you can pay a relatively low sum of money to do this, and they, of course, are all left-leaning. Isn’t all of that going to get just as politicized as the old set of experts?
Williams: I think there remains a lot of uncertainty about exactly why it is that the outputs of large language models are as they are. I think there is expert disagreement about which aspects of the training process, whether it’s pre-training on specific kinds of datasets, whether it’s specific kinds of post-training for helpfulness and honesty, can be biased in the ways that you suggest. I think there’s still a lot of uncertainty about where to assign responsibility for different factors.
I also think this issue of bias is quite complicated. John Burn-Murdoch in the Financial Times did an analysis, and what it seemed to show is that ChatGPT, Gemini, and Claude had something of a center-left bias, Grok had something of a center-right bias, but overall they’re producing expert, accuracy-oriented information that has a net depolarizing effect, whether it’s on the center-left or the center-right. I think it’s very difficult to know exactly how these things are going to evolve over time.
Where I agree is that this is going to become a huge issue. As it becomes clearer that large language models are having an effect on public opinion, albeit one that’s quite difficult to unambiguously detect in specific cases, then it’s going to become a big political issue exactly what kind of content these systems are producing. I have a suspicion that it’s much more difficult to make these models exhibit certain kinds of biases than people think. I think it’s strongly in the interests of these companies, at least at the moment, to train their models in such a way that they place a really strong, high value on honesty and accuracy. If you try, for example, for commercial reasons or for more propagandistic reasons, to bias their output and behavior in some domains without it spilling over into other domains, I think it’s going to be very, very hard, given the way these systems get trained.
To make that concrete, if you’re trying to create a system that’s incredibly useful and competent when it comes to natural scientific questions, medical advice, business advice, legal advice, but then carve out an exception when it comes to politics and make it systematically biased in that domain, I think given the way the technology is currently designed, that’s just technically very, very difficult. It’s precisely for that reason that I think otherwise very different companies are converging on systems that, broadly speaking, are pretty accurate.
I’ll also just say one final thing, which is that there’s a way in which a certain kind of bias might be regrettable but needn’t necessarily undermine trust among a diverse audience. In the United Kingdom, people are constantly aggrieved or angry with the BBC and the biases it allegedly exhibits. Of course, it’s a human institution dealing with very complex issues. I’m sure there are countless biases, and the left is often right to point out right-wing biases, and vice versa. Yet I think, in general, people across the ideological spectrum can often have quite a high degree of trust in the kind of narrow factual reporting the BBC does, especially when you compare it with fringe, hyperpartisan, conspiratorial content online. So I think when it comes to large language models, this question of the precise ideological valence they exhibit is interesting, and I think it will be politicized. But in terms of the immediate impact at the moment, just the fact that they do tend to provide accurate information is something really salient and worth remembering.
In the rest of this conversation, Yascha and Dan discuss the meaning of intelligence in the age of AI, existential risk, and the impact of the Hugging Face incident. This part of the conversation is reserved for paying subscribers…
Listen to this episode with a 7-day free trial
Subscribe to Yascha Mounk to listen to this post and get 7 days of free access to the full post archives.
