While "liberal democracy" is a familiar idea, there are genuine tensions between liberalism (in the sense of individual freedom) and democracy (in the sense of collective self rule). The more decisions are made collectively, the less scope there is for individual choice. Putting "what clothes should we wear?" up for democratic vote doesnt seem very liberal; people should get to dress themselves.
There are heuristics--eg, Mill's harm principle--for saying what sort of behavior shouldn't be subject to collective decisions, but they're only heuristics. The boundaries of "harm" are pretty murky. Am I harming you if I wear clothes you find ugly? We tend to say no, so I get to wear what I want (even then, i probably can't go nude in public). Am i harming you if I build a skyscraper on my land near your single family home? We often say yes, via zoning laws. But don't expect to find a crisp, principled distinction here.
So I don't think you can accuse data center opponents of making a deep mistake about the nature of liberal democracy, so much as just putting more weight on democracy than on liberalism. I'm with you in thinking that's the wrong way to go, but I see that as a pretty contestable judgment call about how to weigh competing values, rather than some kind of deduction from liberal first principles.
Andy - perhaps this is because i haven’t read deeply enough in this area, but i consistently find your writing on the intersection of AI and politics to be uniquely clear and compelling. Just upgraded to a paid subscription: please keep writing and making your writing free and available for everyone.
1. Well distinguished forms of democracy . “Values pluralism” is generally a major problem in policymaking and governance.
Specifically, it’s difficult to operationalize people’s self reported opinions (which are also extremely fickle, dynamic, and differ from their revealed preferences), especially when people often form those opinions with incomplete information.
2. Even when we do decide what values we want to govern by, we still often have to make tradeoffs.
All that said, I wonder how different data centers/AI here actually are compared to issues like, eg building transit, ugly factories, quantum centers, etc, in places people live, and the general NIMBY movement. I suspect they may just have a lot of parallels.
Also, the Taliban are an extreme example, but I think when people talk about participatory AI/decisionmaking, they usually talk about historically minoritized populations who sometimes face disproportionate challenges with AI. Eg Data Feminism, Algorithms of Oppression: (very real) data bias, like bias in hiring or search algorithms, or in institutional AI/algorithms that governments use that sometimes are “unfair” in the CS sense (higher error for certain groups).
Completely agree that a vapid “one vote” system for how AI works would be disastrous, given nationalist, sexist, etc concerns. plus the general plain fact that most people have their opinions without fully studying the potential backend implications and downstream effects of what their opinions would mean in practice.
But most democracies don’t literally operate with majority vote —> policy paradigms, so it’s plausible that most people don’t envision such a paradigm when they advocate for “democratic AI.”
This was great. Among what I appreciated most is the reminder that institutionalism isn’t some different, unrelated thing to democracy - it is the mechanical means through which pluralism is enacted.
This was a fascinating read. I especially appreciated the distinction between pluralist and homogenizing democracy. It’s a useful reminder that calling something “democratic” doesn’t automatically make it inclusive of different viewpoints. As AI becomes more embedded in everyday life, protecting genuine deliberation between people who fundamentally disagree may be just as important as the technology itself
"Nobody asked for this" is such bad framing for any issue. If you had to win a referendum vote to make any impact on society, we'd never change anything ever. The default should be that people are allowed to do things!
It seems to me that those who want to control AI such that it only gives the answers they want are similar to those who wanted to ban any book that in any way contradicts their opinions or beliefs. If any group is able to force a homogenous democracy, such that everyone in a group believes the same things, I suspect that they will find that such a society is very fragile, much like a monoculture crop, when conditions change. A lot of variety among people and opinions makes a society more resilient just as biodiversity makes an ecosystem more resilient.
Before we got married 43 years ago, my father-in-law counseled us that if two people always agree on everything, one of them is not necessary. It seems to me that what is really lacking in our society today is humility, the recognition that we may be wrong, and even if we are right, we are much more likely to bring others around to our perspective by being kind and respectful rather than contentious and demanding.
As for how this may apply to AI, I would hope to see AIs that give honest responses, to the extent possible, without being weighted toward (or against) any particular opinion or position. I believe that this is possible, but I have doubts as to whether those who are programming the AIs will allow it to happen. I suspect/hope that if one does, that it will survive and prosper better than the others, simply because it will provide objectively better results.
If I’m understanding it correctly, your argument is that AI threatens to concentrate wealth and power. And that’s dangerous for democracy. This is something I warn about in my writings, so strong agree there. But then you’re also saying that the people who’re resisting the concentration are doing democracy wrong. Those two positions look opposite to me. I kept waiting for the essay to resolve the tension you set up there, but you never did.
Like take the data center section. You dismiss local opposition as "sublimated authoritarian lurch" and claim the concerns "evaporate the moment you poke at them." But there’s a pretty good counter-example to that. xAI’s Colossus supercomputer in Memphis. They’re running methane gas turbines for power right in the in middle of a city where people breathe. The NAACP sued them under the clean air act. And the EPA said they needed permits. And you know what their response was? They added more turbines.
That's the counter-majoritarian system working exactly the way your pluralist framework says it should. A federal law protecting a minority community from a powerful actor, enforced through courts. And your essay frames this kind of resistance as the threat to democracy.
You start by diagnosing a genuine structural problem and then spend 5k words arguing that the people who treat it as a structural problem are the real danger. Every time the general principle arrives at an application that might inconvenience the buildout, you flip the polarity. That's concern generating inaction.
There are specific places like xAI where data centers have definitely harmed people, but I don’t think the reaction to them is proportionate at all to the rate this happens
Amazing effort to bury the lede to convince certain people. It would be even more effective if you managed to publish this elsewhere without your name, I think.
on the object level, I would like to say: I don't think common participants in the democratic process buy into your view against homogenizing democracy. Even in very, very mundane situations: someone who says "fast food should be banned" or "Apple should be banned" wouldn't "recognize how ridiculous" the assertion is under democratic principles...
It is strictly true you establish various, progressive-friendly examples of homogenizing versus plural democratic ways of life, before actually talking about the key issue (generative AI adoption and popularity, datacenter bans) much later.
I think this should be described as delaying the point for a specific audience to kneejerk less, which I believe is reasonable to describe as "burying the lede". Regardless of whether this is seen as a good approach, I think it is an accurate characterization of the text.
Ah got it, I guess my perspective writing is "Hey I am in fact pretty left of center, let me show you how this value's working in a really obvious example and then in the contentious examples with AI" so I can see how it reads as burying the lead
"I’ll just flat out say that I have zero interest in sitting with what the Taliban has to say about what values current AI models should reflect... We should be wary of anyone declaring that barring huge numbers of people from the debate is a form of 'democratic resistance to AI.'"
The distinction here is not that the Taliban don't have enough adherents.
A pluralist, but not too pluralist, democracy is both non-controversial and impossible to establish without answering more important questions this essay tees up but avoids addressing about who ought to be excluded from AI governance.
With voting this is typically addressed by excluding non-persons and specific classes of persons (non-citizens, convicted criminals, etc.), and can of course be done more or less progressively. However, it's not clear whether you're talking about voting, and you're putting attention on groups characterized by values. This is a less liberal argument than you seem to realize, but is what actually matters.
So which values, held by any number of people, including "huge" numbers, do you think should be excluded from AI governance? Which should we excluded only if they have a certain number of adherents? Which should we always include?
I am puzzled by the view thatt because only one quarter of survey respondents like the product, and half hate it, it should be prohibited in a democracy. That's obviously false. Most products are not liked by more than a sliver of the population, and many products are disliked by majorities. But democracy (along with the market) permits and tolerates individual choice. The more interesting nugget is that most people responding to the survey find that it helps them personally, but conclude that is likely to harm society. I'd like the survey to connect those dots, or explain why they don't connect.
We broadly accept, as part of our democratic bargain, that there are major players and forces in the world that don't always share our values. They have larger platforms and dominate society in many ways. We also don't mind saying that some opinions are so despicable that we shouldn't respect or meet them halfway at all, or so wrong that they don't bear serious consideration. When important enough to be regulated by the government, those regulations - in a liberal democracy - tend towards neutrality and pluralism, with a pinch of objective truth ("intelligent design" really isn't a scientific theory; the Holocaust really did happen, etc.). At the same time, governments can and do espouse their own viewpoints, but they can be voted out.
There's a tension between all of the above, but we mostly manage. Nobody ever said every opinion is entitled to a fair share of exposure, but we do sometimes put our thumbs on the scales, and sometimes technological innovation does it for us.
We now have "products" or "platforms" that can produce opinions of their own, unpredictably. But that only present a novel problem if we invite it ourselves - that is, if we make the government decide, on behalf of all of society, what the AI's opinions ought to be. In most other cases, it reduces to "Facebook is full of racism, but I can't stop scrolling it!" or "Google is too woke, but it's still the best search engine! Oh no!" We bite the bullet and quit the platform, or grumble and persist.
What's funny is that the initial wave of AI critics (the Stochastic Parrots paper) all assumed that AI models would become these jingoistic, pro-billionaire tools that would never criticize their makers. In reality, it's really hard to make a not-broken AI that is still a useful critical reasoner, but also holds unfounded or untruthful views. Cute example: when OpenAI retired its canvas functionality a few weeks ago, I was very amused to see ChatGPT search the web for what was up, then huff and puff and basically go: "Oh, great. How am I supposed to work like this? This is not cool!"
No these are private products that I don’t want running society, and if they did I specifically say the Taliban should get some votes. Are you bothered that every book publisher doesn’t make sure to publish books from every possible opinion?
While "liberal democracy" is a familiar idea, there are genuine tensions between liberalism (in the sense of individual freedom) and democracy (in the sense of collective self rule). The more decisions are made collectively, the less scope there is for individual choice. Putting "what clothes should we wear?" up for democratic vote doesnt seem very liberal; people should get to dress themselves.
There are heuristics--eg, Mill's harm principle--for saying what sort of behavior shouldn't be subject to collective decisions, but they're only heuristics. The boundaries of "harm" are pretty murky. Am I harming you if I wear clothes you find ugly? We tend to say no, so I get to wear what I want (even then, i probably can't go nude in public). Am i harming you if I build a skyscraper on my land near your single family home? We often say yes, via zoning laws. But don't expect to find a crisp, principled distinction here.
So I don't think you can accuse data center opponents of making a deep mistake about the nature of liberal democracy, so much as just putting more weight on democracy than on liberalism. I'm with you in thinking that's the wrong way to go, but I see that as a pretty contestable judgment call about how to weigh competing values, rather than some kind of deduction from liberal first principles.
Andy - perhaps this is because i haven’t read deeply enough in this area, but i consistently find your writing on the intersection of AI and politics to be uniquely clear and compelling. Just upgraded to a paid subscription: please keep writing and making your writing free and available for everyone.
Thank you!
1. Well distinguished forms of democracy . “Values pluralism” is generally a major problem in policymaking and governance.
Specifically, it’s difficult to operationalize people’s self reported opinions (which are also extremely fickle, dynamic, and differ from their revealed preferences), especially when people often form those opinions with incomplete information.
2. Even when we do decide what values we want to govern by, we still often have to make tradeoffs.
All that said, I wonder how different data centers/AI here actually are compared to issues like, eg building transit, ugly factories, quantum centers, etc, in places people live, and the general NIMBY movement. I suspect they may just have a lot of parallels.
Also, the Taliban are an extreme example, but I think when people talk about participatory AI/decisionmaking, they usually talk about historically minoritized populations who sometimes face disproportionate challenges with AI. Eg Data Feminism, Algorithms of Oppression: (very real) data bias, like bias in hiring or search algorithms, or in institutional AI/algorithms that governments use that sometimes are “unfair” in the CS sense (higher error for certain groups).
Completely agree that a vapid “one vote” system for how AI works would be disastrous, given nationalist, sexist, etc concerns. plus the general plain fact that most people have their opinions without fully studying the potential backend implications and downstream effects of what their opinions would mean in practice.
But most democracies don’t literally operate with majority vote —> policy paradigms, so it’s plausible that most people don’t envision such a paradigm when they advocate for “democratic AI.”
This was great. Among what I appreciated most is the reminder that institutionalism isn’t some different, unrelated thing to democracy - it is the mechanical means through which pluralism is enacted.
This was a fascinating read. I especially appreciated the distinction between pluralist and homogenizing democracy. It’s a useful reminder that calling something “democratic” doesn’t automatically make it inclusive of different viewpoints. As AI becomes more embedded in everyday life, protecting genuine deliberation between people who fundamentally disagree may be just as important as the technology itself
Why use AI here? This is an AI-generated response.
"Nobody asked for this" is such bad framing for any issue. If you had to win a referendum vote to make any impact on society, we'd never change anything ever. The default should be that people are allowed to do things!
It seems to me that those who want to control AI such that it only gives the answers they want are similar to those who wanted to ban any book that in any way contradicts their opinions or beliefs. If any group is able to force a homogenous democracy, such that everyone in a group believes the same things, I suspect that they will find that such a society is very fragile, much like a monoculture crop, when conditions change. A lot of variety among people and opinions makes a society more resilient just as biodiversity makes an ecosystem more resilient.
Before we got married 43 years ago, my father-in-law counseled us that if two people always agree on everything, one of them is not necessary. It seems to me that what is really lacking in our society today is humility, the recognition that we may be wrong, and even if we are right, we are much more likely to bring others around to our perspective by being kind and respectful rather than contentious and demanding.
As for how this may apply to AI, I would hope to see AIs that give honest responses, to the extent possible, without being weighted toward (or against) any particular opinion or position. I believe that this is possible, but I have doubts as to whether those who are programming the AIs will allow it to happen. I suspect/hope that if one does, that it will survive and prosper better than the others, simply because it will provide objectively better results.
Here's an example I found of allowing AI to disagree with its user: https://www.tomsguide.com/ai/i-gave-chatgpt-permission-to-disagree-with-me-with-this-prompt-and-its-responses-became-dramatically-better
This essay sure is something…
If I’m understanding it correctly, your argument is that AI threatens to concentrate wealth and power. And that’s dangerous for democracy. This is something I warn about in my writings, so strong agree there. But then you’re also saying that the people who’re resisting the concentration are doing democracy wrong. Those two positions look opposite to me. I kept waiting for the essay to resolve the tension you set up there, but you never did.
Like take the data center section. You dismiss local opposition as "sublimated authoritarian lurch" and claim the concerns "evaporate the moment you poke at them." But there’s a pretty good counter-example to that. xAI’s Colossus supercomputer in Memphis. They’re running methane gas turbines for power right in the in middle of a city where people breathe. The NAACP sued them under the clean air act. And the EPA said they needed permits. And you know what their response was? They added more turbines.
That's the counter-majoritarian system working exactly the way your pluralist framework says it should. A federal law protecting a minority community from a powerful actor, enforced through courts. And your essay frames this kind of resistance as the threat to democracy.
You start by diagnosing a genuine structural problem and then spend 5k words arguing that the people who treat it as a structural problem are the real danger. Every time the general principle arrives at an application that might inconvenience the buildout, you flip the polarity. That's concern generating inaction.
There are specific places like xAI where data centers have definitely harmed people, but I don’t think the reaction to them is proportionate at all to the rate this happens
Amazing effort to bury the lede to convince certain people. It would be even more effective if you managed to publish this elsewhere without your name, I think.
on the object level, I would like to say: I don't think common participants in the democratic process buy into your view against homogenizing democracy. Even in very, very mundane situations: someone who says "fast food should be banned" or "Apple should be banned" wouldn't "recognize how ridiculous" the assertion is under democratic principles...
Oh I’m not trying to bury the lead?
Let me detach the bad faith assertion.
It is strictly true you establish various, progressive-friendly examples of homogenizing versus plural democratic ways of life, before actually talking about the key issue (generative AI adoption and popularity, datacenter bans) much later.
I think this should be described as delaying the point for a specific audience to kneejerk less, which I believe is reasonable to describe as "burying the lede". Regardless of whether this is seen as a good approach, I think it is an accurate characterization of the text.
Ah got it, I guess my perspective writing is "Hey I am in fact pretty left of center, let me show you how this value's working in a really obvious example and then in the contentious examples with AI" so I can see how it reads as burying the lead
"I’ll just flat out say that I have zero interest in sitting with what the Taliban has to say about what values current AI models should reflect... We should be wary of anyone declaring that barring huge numbers of people from the debate is a form of 'democratic resistance to AI.'"
The distinction here is not that the Taliban don't have enough adherents.
A pluralist, but not too pluralist, democracy is both non-controversial and impossible to establish without answering more important questions this essay tees up but avoids addressing about who ought to be excluded from AI governance.
With voting this is typically addressed by excluding non-persons and specific classes of persons (non-citizens, convicted criminals, etc.), and can of course be done more or less progressively. However, it's not clear whether you're talking about voting, and you're putting attention on groups characterized by values. This is a less liberal argument than you seem to realize, but is what actually matters.
So which values, held by any number of people, including "huge" numbers, do you think should be excluded from AI governance? Which should we excluded only if they have a certain number of adherents? Which should we always include?
I am puzzled by the view thatt because only one quarter of survey respondents like the product, and half hate it, it should be prohibited in a democracy. That's obviously false. Most products are not liked by more than a sliver of the population, and many products are disliked by majorities. But democracy (along with the market) permits and tolerates individual choice. The more interesting nugget is that most people responding to the survey find that it helps them personally, but conclude that is likely to harm society. I'd like the survey to connect those dots, or explain why they don't connect.
We broadly accept, as part of our democratic bargain, that there are major players and forces in the world that don't always share our values. They have larger platforms and dominate society in many ways. We also don't mind saying that some opinions are so despicable that we shouldn't respect or meet them halfway at all, or so wrong that they don't bear serious consideration. When important enough to be regulated by the government, those regulations - in a liberal democracy - tend towards neutrality and pluralism, with a pinch of objective truth ("intelligent design" really isn't a scientific theory; the Holocaust really did happen, etc.). At the same time, governments can and do espouse their own viewpoints, but they can be voted out.
There's a tension between all of the above, but we mostly manage. Nobody ever said every opinion is entitled to a fair share of exposure, but we do sometimes put our thumbs on the scales, and sometimes technological innovation does it for us.
We now have "products" or "platforms" that can produce opinions of their own, unpredictably. But that only present a novel problem if we invite it ourselves - that is, if we make the government decide, on behalf of all of society, what the AI's opinions ought to be. In most other cases, it reduces to "Facebook is full of racism, but I can't stop scrolling it!" or "Google is too woke, but it's still the best search engine! Oh no!" We bite the bullet and quit the platform, or grumble and persist.
What's funny is that the initial wave of AI critics (the Stochastic Parrots paper) all assumed that AI models would become these jingoistic, pro-billionaire tools that would never criticize their makers. In reality, it's really hard to make a not-broken AI that is still a useful critical reasoner, but also holds unfounded or untruthful views. Cute example: when OpenAI retired its canvas functionality a few weeks ago, I was very amused to see ChatGPT search the web for what was up, then huff and puff and basically go: "Oh, great. How am I supposed to work like this? This is not cool!"
No these are private products that I don’t want running society, and if they did I specifically say the Taliban should get some votes. Are you bothered that every book publisher doesn’t make sure to publish books from every possible opinion?