Governing the Ungovernable: What Three AI Communities Taught Me About Hyperobjects
(research diary · Fairfax, VA, USA / Sydney, Australia · October 2025)
I just returned from presenting at the Markets and Society 2025 conference from the Mercatus Center at George Mason University, and I’m still processing what happened. Not the presentation itself—that went fine. But the question that came afterward, from a regulatory scholar who’d spent twenty years working on tech policy: “If AI is truly what you’re describing, aren’t we just… screwed?”
He wasn’t being flippant. He was genuinely grappling with what it means to govern something that exceeds the conceptual boundaries of our governance systems. Something that, as Prof Jason Potts argues in his brilliant recent piece, might be best understood as a hyperobject: an entity so vast in spatiotemporal scale and complexity that it withdraws from human perception, making it fundamentally illegible to the institutional architectures we’ve built to manage things.
The paper I presented examined three experiments in decentralised AI governance—Prime Intellect, EleutherAI, and Hugging Face—through the lens of Elinor Ostrom’s design principles and Architect Christopher Alexander’s form-fit-context framework. But what started as an analysis of “alternative governance models” became something stranger: an investigation into what it means to govern not over AI, but within it.
And I’m not sure we’re screwed. But I am sure we’re asking the wrong questions.
The Control Theory Trap
Traditional AI governance operates on a simple assumption: AI is a discrete object with clear boundaries that can be regulated from the outside. You identify the thing, you write rules about the thing, you enforce compliance with the regulations. It’s the same logic we apply to pharmaceuticals, aircraft, or financial products.
The problem is that AI isn’t really a thing in that sense anymore. It’s an infrastructure. It’s a knowledge production system. It’s a network of models, datasets, compute resources, human practices, institutional arrangements, and feedback loops that spans continents and constantly reconfigures itself. It’s non-local, interobjective, and partially withdrawn from our ability to perceive it as a unified whole.
In other words: it’s a hyperobject.
When I started researching how communities were actually governing AI development (not how they should according to policy papers, but how they were in practice) I kept encountering the same pattern. Every attempt to govern AI from the outside, through traditional regulatory mechanisms, was running into the same wall. Not because regulators were incompetent or regulations were poorly designed, but because the fundamental ontology was wrong.
You can’t regulate a hyperobject with control theory. You need something else.
Three Experiments in Governance-Within
The cases I studied represent three radically different approaches to AI governance. Still, they share a common insight: if you can’t govern AI from outside, maybe you can govern from within by building institutional arrangements that are themselves entangled with the AI ecosystem they’re trying to shape.
Prime Intellect is building a decentralised protocol for collective AI ownership, leveraging blockchain infrastructure to coordinate global GPU contributions and democratise access to model training. Their governance isn’t imposed on the network—it is the network. Rules are encoded directly into smart contracts. Monitoring happens algorithmically through their verification system. Decisions emerge through token-weighted voting by stakeholders who are economically invested in outcomes.
This is governance through entanglement at its most literal: you can’t separate the governance structure from the technical infrastructure. The form is the fit; the context is the fit. It’s all one recursive loop.
EleutherAI took the opposite approach: radical informality. Born as a Discord server in response, amongst many things, to GPT-3’s closed release, they proved that a loose collective of volunteers could replicate state-of-the-art models through “science in the open.” No formal membership. No constitution. Just norms, transparency, and shared purpose.
What made this work wasn’t just the community’s passion; they recognised that openness itself was a governance mechanism. By conducting all research publicly, they created a distributed sensing system in which anyone could detect emerging patterns, identify errors, or raise concerns. The governance emerged from the practice, not from rules imposed on it.
Hugging Face represents a third model: layered governance that nests corporate structure within community participation. As the “GitHub for AI,” they manage a massive platform where hundreds of thousands of models get shared, used, and scrutinised. Their governance combines top-down policy (content guidelines, licenses) with bottom-up participation (community moderation, collaborative documentation).
What’s fascinating is how they’ve created recursive feedback loops between the platform, the community, and the content. Model cards become governance artifacts. Forum discussions shape policy. Community values get encoded into licensing standards. It’s not just platform governance about AI: it’s governance through the socio-technical assemblage of platform, models, users, and practices.
From Governance Of to Governance With/In
The conference presentation ended with what I thought was my conclusion: these three cases demonstrate that private, decentralised governance can complement, or even substitute for, state regulation in managing AI development. They show that polycentric arrangements—multiple decision-making centres coordinating without central control—can be remarkably effective.
That’s true, but it’s also incomplete.
The deeper insight is that we’re witnessing a fundamental shift in what governance means when the thing being governed is a hyperobject. We’re moving from governance of AI (external regulation of a discrete object) to governance with/in AI (institutional arrangements entangled with the computational infrastructures they’re shaping).
This has profound implications:
First, it means we can’t wait for comprehensive regulatory frameworks before acting. By the time we’ve mapped the territory, it has changed. The governance experiments happening right now in communities like these three aren’t just alternatives to regulation—they’re proto-forms of what regulation might need to become.
Second, it suggests that the most effective governance will be recursive and adaptive rather than fixed and comprehensive. Not rules that anticipate all scenarios, but systems that can sense, respond, and evolve as AI reveals new dimensions of itself. Infrastructure that governs and is governed simultaneously.
Third, it implies that transparency takes on new significance—not just as an ethical principle, but as an epistemological necessity. If we can’t fully perceive the hyperobject, we need maximum visibility into the parts we can see, with mechanisms for sharing and synthesising those partial views. EleutherAI’s “science in the open” isn’t just ideology; it’s a survival strategy for governance under hyperobject conditions.
Fourth, it points toward a role for decentralisation that goes beyond ideological preferences for markets or communities over states. Polycentric governance may be epistemologically necessary for hyperobjects precisely because no single vantage point can grasp the whole. You need multiple centres of perception and decision-making, each with partial but different views, coordinating through both competition and cooperation.
So, Are We Screwed?
Back to the regulatory scholar’s question. If AI is truly a hyperobject, if it genuinely exceeds the conceptual boundaries of our governance systems, then aren’t we just screwed?
I gave him the academic answer at the conference: that these cases demonstrate alternative pathways, that polycentric governance can work, and that iterative adaptation matters more than comprehensive planning.
But here’s the more honest answer: I don’t know.
What I do know is that treating AI as governable through traditional control mechanisms (as if it were pharmaceuticals or aircraft) is definitely not working. The mismatch between our governance forms and the context we’re trying to govern is producing increasingly dangerous misfits.
These three communities aren’t solving the governance problem. But they’re doing something potentially more valuable: they’re conducting live experiments in what governance might need to become when the thing to be governed withdraws from whole perception. They’re building proto-institutions that can sense, respond, and adapt rather than control and contain.
Prime Intellect shows us governance as algorithmic infrastructure. EleutherAI shows us governance as distributed sensing through radical transparency. Hugging Face shows us governance as recursive feedback between platforms, communities, and artifacts.
None of these is the answer. But collectively, they’re sketching the contours of a question we’re only beginning to formulate properly: not “How do we govern AI?” but “How do we build institutions that can govern with a hyperobject—that can operate within, learn from, and adapt to an entity that will always exceed our ability to perceive it fully?”
The scholars at Mercatus get this. The Markets and Society conference isn’t just about AI policy—it’s about how institutions, markets, and social orders emerge and adapt in the face of radical uncertainty. That’s why this work resonates there. We’re not just studying AI governance. We’re studying institutional evolution under hyperobject conditions.
And maybe that’s actually the hopeful part. Because humans have always lived with hyperobjects—climate, “capitalism”, radiation, the internet. We’ve built institutions that function despite not fully grasping what they’re embedded in. These institutions aren’t perfect. They create their own problems. But they’re proof that governance under partial perception is possible.
The question isn’t whether we’re screwed. The question is whether we can build institutions nimble enough, transparent enough, and reflexive enough to govern with rather than govern over. To learn as we go. To sense and respond rather than plan and control.
I think these three cases suggest we can. But it requires abandoning the fantasy of comprehensive oversight and embracing something messier: governance as continuous experimentation, adaptation, and sense-making within the hyperobject itself.
A Note on Method: The full research paper “AI Governance in Motion: Aligning Form, Fit, and Context Amid Decentralization” is available as a preprint here. It’s much more systematic and rigorous than this Substack post, with detailed case analysis, theoretical frameworks, and proper academic argumentation. This piece is meant to pull out the ideas that keep me up at night and share them in a more conversational mode.
Coming Soon: That promised return to post-text knowledge infrastructures. Because if we’re governing hyperobjects, we need knowledge artifacts that can be hyperobjects too. Stay tuned.
Your Turn: Are you working on AI governance in communities, companies, or government? What misalignments are you seeing between governance forms and the contexts they’re trying to fit? What experiments are you running? I’m genuinely curious what others are experiencing—reach out at luishernando.lozanoparedes@uts.edu.au or find me on the usual platforms as Luis Lozano Paredes.