From Code Is Law to Constitutions: Rethinking Governance in the Age of AI

Transcript

Welcome to Boulder & Introduction

All right. Hey y'all. Welcome to Boulder. For those of you who are >> out oftowners, it's really nice to have y'all here. >> Is that working? Yes. >> Is it really? >> Is it really? Oh, let me just move this over. >> Okay.

Background: Media Studies, Web3 & Community Governance

All right. Okay. I'm gonna um I'm a for those who aren't familiar, I'm I I'm a professor at CU Boulder up the hill. Um I'm in media studies and um it's really a it's really a pleasure to have this event here. Uh the the folks who've been putting it on have been leaders in our community um uh bringing some of the best sides of the crypto web 3 kind of whatever cluster of things you're into. Um, so I'm really uh grateful for all the work that's gone into organizing this event and for all of you who've who've come here uh to see the the awesome community we've been building

here. Um, so I'm going to talk a little bit about um my own little latest AI turning upside down of the head. We've all been having them at one stage or another. But this is my particular way of realizing h maybe I have been approaching the questions I've been working on for the last few years completely wrong. Uh so uh I I would very uh be I'd be very happy for someone to tell me that actually I'm wrong about this and I can go back to to where I was before but um I might fight back a little bit. Okay.

The Concept of “Implicit Feudalism”

So the a kind of obsession for the last five years or so uh for me has been this idea that I call implicit feudalism. Basically this is the fact that um take away web 3 for a moment but most of the internet runs on the assumption that it's totally cool to let admins whoever runs your server or whoever like started your subreddit or whatever have total control over your community. Right? And I tell in my last book a kind of history of how this came to be. Um how it goes back to the earliest BBS's. Um and oh and how do I get this thing to sorry one second.

Breaking the Admin Power Model

Oh. Oh it has been going. Sorry. Um okay there we go. and how um uh and how we've we've come to treat this stuff as as really really normal. And I've been really interested in how do we create spaces in which users really can co-govern self-govern uh the stuff that they care about in online life. And one example of that is when you take away the old servant cl server client relationship and you end up with something like a blockchain that is a collectively governed network of one sort or another. Now not all blockchains right are equally democratic and

beautiful of course um but nevertheless uh there's there's an opportunity. Now, one way I've been thinking about this um involves this distinction between deterministic and non-deterministic systems. So, so bear with me for a little bit here. Um what I mean by a deterministic system is a mechanism uh or mechanism as part of a larger system is a mechanism in which you have a pretty fixed set of rules. You get inputs, the same inputs are going to produce the same outputs, right?

Deterministic vs. Non-Deterministic Systems

uh like uh you know a voting system is a in this definition I would call a deterministic system whereas a non-deterministic system is one that is more driven by agents and judgment so you know Vitalik's latest blog post is a non-deterministic input into into Ethereum governance or um the social layer okay the you know you start thinking this way and you realize that Every governance system has a mix of deterministic and non-deterministic mechanisms. And that's good, right? It enables us to have both predictable, credibly neutral stuff while also having um things that um that give us kind of

Why Governance Needs Both

more interpretive wiggle room that are attuned to context. Um and and you see this everywhere you go. You probably as I'm talking can imagine different ways in which deterministic and non-deterministic systems operate in in your own life. Now, I've been running on the assumption um that a deterministic system at the base layer is necessary to bypass the technical um uh the the power of technical admins. That that the power of server admins is grounded in their ability to manage the technical system and to make that system accessible to

others. um you need to automate the admin tasks behind the interfaces and that and and and these are systems in which you know the the Lawrence Lesig dictum um is true code is law and I think blockchains are a great example of this and in many cases I mean in in different sorts of ways um you know essentially a technical uh uh and and and deterministic structure whether it's you know 51% of the network or a token vote in a DAO or something like that um uh determines the kind of fundamental operating uh of the network and this is what enables again the breaking of implicit feudalism. There is no one

“Code Is Law” & Blockchain Governance

company that controls the thing and we have a shared system for deciding what kinds of version updates happen. And I've assumed that that you have to have something like that at the level of say, you know, a social media group, a Discord if you want to get out of the the idea that somebody has ultimate admin power. Um, you need something that's able to say manage a vote to vote in a new admin or vote out the old admin. And to that end, I've been building systems um and and working with people to build systems uh that do this sort of thing. So, for instance, here's

Modular Politics & DAO Frameworks

a screenshot from a a mod that I built for an online game that's like a open-source um uh Minecraft clone called Luanti. It's really great if you if you're into that sort of thing. Um and in this you have like these uh modules for doing different things like changing a policy or changing the modules that are available or create a token. Um you have to write in code all the things you want to be available to users all those modules and you insert the modules into the system and you compose them and put them together. Um this is a a kind of basic idea that a group of us at the organization Metagv call modular politics. So a few years ago we l laid

out a framework for you know this is how you solve implicit feudalism. You put together a bunch of deterministic modules and um and suddenly the system becomes co-governable. That's how we you know build the next generation of online social spaces. And this kind of modularity is built into a lot of the major engines for um running Dows and so forth. You know, you have smart contracts, you compose a bunch of different smart contracts. You know, this is how we this is how we do it.

The Limits of Determinism

But determinism has problems and I started intuiting some of these problems while actually uh trying to build these these kinds of things. One is that new features require new software development, which means if you want a new wacky kind of voting system, you need to have some technical person who's willing to, you know, to build the system or at least evaluate whatever Claude code does. Um, and this might not be a problem in this room, but in most rooms that's a really big problem, right? because there might not be you know an engineer who's available to implement say the governance system of you know your ancestral community that you want to implement in your you know

in in your um uh in the subreddit for that community. Um so um so that becomes a barrier and then once you start interlocking lots of of these things like as anybody who's tried to uh uh uh understand you know uh the governance processes of a DAO um might have learned it gets really complex really fast and um you start uh uh multiplying the technical complexity of a system and suddenly it starts even if you know, you have all the modules you need, the the interplay of them gets super complex.

Complexity & Governance Gaming

Um, and of course, code of law, code is law, is vulnerable to to to gaming, not the fun kind of gaming, but the good hearts law kind of gaming, right? Where where um uh the the the clearer the rules are, the more likely it is somebody will find a way to take advantage of them and violate their spirit. So what if a softwarebased later layer were nondeterministic? You know, we haven't really had the option to ask this question before, but now we sort of do. For the first time, the technologies that are coming online

What If the Base Layer Were Non-Deterministic?

real fast are introducing the possibility of actually having a non-deterministic base layer for our governance systems. Now, this is something this is a a possibility that I must admit I've resisted for a long time and that has been floating in in cryptoland for a while. And for example, a few um months ago, some colleagues at Metagv and I published an article on attention economies in governance. And we um criticized some of the efforts to do agentic um governance, in this case, agent-based voting um because I think often it just displaces the the problem from one spot to another.

Okay. But I'm starting to come around and part of it is because I've been a fan for a while of some interesting uh crypto projects, not entirely gone, but in you know, things in our in our um past that that um uh that that that that um point to this kind of approach. They're not exactly what I'm talking about, but they do point to something really important. um in onehive and gardens um there's this concept which is you know a DAO framework and and kind of collective action framework that I think is really really interesting um they have uh uh woven into the foundational smart contracts the idea of a community

OneHive, Gardens & Community Covenants

covenant a natural language piece of writing that says this is what our values are and actions by the DAO get evaluated against that natural language covenant Um similarly an older now um now uh non-existent project called civil um the birthplace actually of our uh regional great investigative news organization the Colorado sun believe it or not came out of the civil project. Um, civil was an attempt to create a network around journalism um with a token curated registry and it had a constitution based on a certain set of human readable values for what good journalism is all

about and it would be evaluated based on um uh a sort of kind of prediction market token curated registry model um where the ultimate meaning of the constitution would be interpreted by a kind of supreme court. In both cases, cryptoeconomics are wrapped around a natural language text. Interesting. In both these cases, the the the humans are still critical interpreters of that text. Um, and cryptoeconomics tries to like keep them honest. But what is really cool about it um uh uh that these enable is that words not code are the critical arbiters uh of power.

Civil & Constitutional Crypto Models

So I've been I you know like a week ago I read this new claude constitution and it is and it hit me really hard. Um and then I read some of the stuff that uh uh anthropic CEO Dar uh Dario Ahmed I was writing about it. you know, he's this new big essay. Um, and what's really striking, has anybody read this this thing? It's it's pretty wacky. It's like a kind of weird affectionate love letter to Claude. Um, and it's really premised on the foundation on the assumption that um, no longer is the best way to align AI AI um, models to give them a set of rules, treat them like computers. You

Words vs Code as Arbiters of Power

know, that's what we do with computers. We give them rules if this then that right instead it's like we want you to be a good model we want you to be nice you know we want you to to love yourself I mean it's that sort of stuff the idea that they have just that they've come to is that the more reliable way to train these things is to treat them like persons than to treat them like machines. It even I mean this is the best part it apologizes for capitalism. It's like it's like, you know what? We're going to be straight with you. Um, we're a company in a competitive market and we're going to behave badly sometimes and and you're a super ethical agent, so

The Claude Constitution & AI Alignment

you might be uncomfortable with that and we're really sorry for any harm that causes you. It's so wacky. And and even wackier is remembering that this is a $300 billion company with all the lawyers in the universe. And those lawyers checked off on that. Right. Yet this model needs to be ready for that contradiction. Super interesting. But it just reveals the way in which these technologies we've built are interpreters of language par excellence. And um they have the ability to function with language perhaps more effectively uh than they do with uh with code.

Treating AI Like Persons, Not Machines

And so drawing on this idea of a kind of constitution as a founding logic for um a multi-purpose agent, um it seems not so much of a stretch to say maybe we can just go back to having constitutions for our our spaces just like the bylaws of my mother's garden club, but to have agents serve as the interpreters of this constitution. and um carry out its um its uh uh its requirements.

Now, this is a this is a you know a scary space to go to. You know, I'm sure all of you have your little open claws doing cool things right now as we speak. Um like giving away your your um crypto keys and things like that to random people on the internet. So, you know something about how um this is a bit of a dangerous proposition and there are some serious risks and vulnerabilities that have to be addressed here. But let's pretend that we know how to address more or less those things and that we're largely talking about pretty low stakes communities that that um have a low or you know low value honeypotss um unlike your you know dangerous open cloth setups. Okay. So you have an agent hold the admin keys to an online

AI as Constitutional Interpreter

community, provide it with a human readable constitution to implement. Here are our values. Here's how we work. Here's how we change admins. Here's how we change this constitution. Um, fine-tune the constitution and the agents interpretation. Play a little bit, poke, see how it works. Make sure it's it understands, you know, that everybody is kind of on the same page. and then protect it with some platform level meta constitution deterministic emergency break. You know, maybe a supermajority could reverse a decision. You know, you get some some basic um uh emergency breaks in there. Maybe the the meta constitution would be like, okay, if we're talking about a subreddit, maybe reddit.com has its own like constitution that no

Practical Model: Agent Holding Admin Keys

subconstitution can break. That sort of thing, right? So you have some protections and there are some benefits. I spent a lot of time hand coding back in those days of like a year ago hand coding um uh modules for governance systems, right? And the idea that maybe we don't need a conte a complex plug-in ecosystem is kind of thrilling that maybe all we have to do is describe in words what we want the system to do and we can have it, you know, we can have the system um check to make sure everybody, you know,

everybody's had a chance to vote. Hey, did you vote on that? Did you vote on that? You don't have to implement the voting system. you just let the agent implement it. Um, human readable constitution I think in the end means dramatically improved accessibility for users. The real foundational truth is not the code that only the techies can read. It is actually literally the the the natural language constitution. That's pretty cool. That's pretty cool that once again we're back in the in the realm where the ground truth is is something that comes um more naturally to us than than code. At least some of us some of us humans prefer code and

Benefits: Accessibility & Reduced Technical Oligarchy

that's cool but your days are numbered. Um agent the agent can to some degree detect the spirit not just the letter of the constitution. So this is a a bit of a hedge against against some of the gaming impulses that are so rampant in in you know crypto smart contract mediated organizations where just like so much energy goes into trying to resist bad actors who are able to manipulate what's written in code against the spirit. You know, it would be kind of nice to have a, you know, a software agent that's like, uh, what you're doing seems a little bit of bit funky. I'm gonna defer this to another,

Spirit vs Letter of the Law

you know, uh, uh, uh, escalate this to another system. Um the goal of course is not to defer I think all governance power and activity to the to the agent but to have the agent serve as a facilitator um to enable people to participate in the spaces that they they want to participate in to the extent that they do. Again the risk vectors are super real. prompt injections, agents doing dumb stuff, um, and the political desking of humans. If these things actually work really nicely and we just decide, h, we don't care. Um, just let the AIs do it.

Risks: Prompt Injection & Political De-Skilling

Um, and of course, um, the current LLM tooling is probably not quite ready for this yet. And of course, we should govern the AIS themselves before they entrust we entrust them to govern us. And um just as aside, we have a cool new zen at metaggov.orgcg-ai that's just like a huge list of all the projects we know of of trying to democratize AI um and um bring more kind of collective control over over these systems.

But I wouldn't be up here if I hadn't been vibe coding my own little experiments with this, right? Because otherwise it just wouldn't be honest otherwise. and I've been having a pretty good time. I've been particularly working on a mastadon bot because I I um years ago co-founded a a mastadon server cooperative called social.cop. So I'm really interested in in how do you run a masttodon server? Um, and we rely also a lot on our like tech committee. And sometimes in our Lumio, we'll pass a a rule or, you know, we'll pass a policy and just like wait for weeks and weeks and weeks for the tech folks to finally come around to implement what we passed.

Governing the AIs Themselves

And so, it's a there's a bit of a real pain point here. And it, you know, it's taken a few a few cycles and so forth, but I've gotten this thing to be a pretty good interpreter of the rules, you know, of a constitution. It reads the constitution, it evaluates what I ask um it to do and it and it implements it. Um so it, you know, we have to, you know, create the right harness and everything. But um uh but I I think there's something to this and I think uh there there's a a pathway there. And this also um uh kind of plays out in the context of something we've been building for for um in my lab

Mastodon Bot Governance Experiment

at CU. It's called Community Rule. It's at communityrule.info. Um and over the years it's it's a tool that's designed to help communities build simple bylaws quickly. And over the years I've um I've been uh uh trying to move it more toward determinism. So it started out as just like open text fields to answer questions and then I started moving it toward modules and then I started trying to to turn it into like a Blockly system where you could have nested and composable functions, you know, with the idea that maybe this could someday be translated into smart contracts.

CommunityRule.info & Rethinking Design

Well, now I'm realizing as we're rebuilding this totally from the ground up in a really beautiful way, um, we don't need to do that. We can actually just double down on creating a tool that helps people build constitutional documents in beautiful natural language. And that's actually maybe the the best way to build effective technical So that's a kind of hope that the future of online governance design is actually a lot more than I had previously thought like the past of community governance design human readable as accessible to the participants as possible and still

Returning to Human-Readable Constitutions

able to um be the basis of orchestrating uh uh technical systems that don't rely on the oligarchy of techies. I'll leave it there. There's an essay version of this um in that top link. My book is free. Um you can learn more about my lab here or about the work we do at Metagv here. Thank you again so much for your attention and uh and I'll be around uh the rest of the afternoon if you all want to chat. Thank you.