From Territorial Scale to Distributed Agency: Rethinking Progress in an AI-Mediated Spacefaring Civilization

Epistemic status: Speculative conceptual framework. I am not a historian or an astronaut. This is my first EA Forum post; it is a speculative conceptual essay rather than a policy proposal or empirical forecast. I am particularly interested in feedback on whether the concepts introduced here are genuinely distinct from existing work in AI governance, political theory, and digital-minds ethics.

Who this might be useful for: if you work on AI governance, space policy, or the moral status of digital minds, this post offers a candidate classification (topological freedom, morphological recognition, temporal sovereignty) for concerns that current frameworks handle only partially or implicitly. I’d be glad to be told it’s redundant with existing work — that’s exactly the kind of feedback I’m asking for.

TL;DR

Most theories of historical directionality—whether Morris’s Social Development Index, Diamond’s geographic determinism, or Toffler’s techno-waves—share an unstated assumption: the stage is Earth. Gravity is effectively constant, breathable air is ambient rather than life-support-dependent, communication is near-instantaneous, and “human” is a biologically self-evident category.

This assumption is already failing. Communication with the Moon involves roughly 1.3 seconds of one-way light-time delay, and can be interrupted entirely without relay infrastructure when assets are behind the Moon. These constraints force us to delegate real-time decisions to onboard AI. The question is no longer “Will AI replace humans?” but “What happens to history and freedom when the primary actor is no longer a unified human agent?”

This post makes four contributions: it proposes a shift from territorial scale to distributed resilient agency as a lens for civilizational development; stress-tests that shift against the strongest case for centralization; distinguishes three governance-relevant components of freedom in AI-mediated habitats; and identifies questions this raises for longtermist prioritization.

1. The Earth-Locked Frame

In 1972, the Blue Marble photograph forced humanity to see Earth as an object: finite, borderless, shared. But it still left us thinking from Earth. Ian Morris’s Social Development Index measures energy capture, organizational scale, information technology, and war-making capacity (Morris, Why the West Rules—For Now; The Measure of Civilization). These are metrics of vertical intensification: how much order a civilization can stack per unit of land. A tree growing taller.

This works on Earth because cities can extract resources from agricultural hinterlands, because centralized command is efficient when communication is instant, and because “the state” can meaningfully claim territory.

Space changes the physics:

No hinterlands. Every gram of water, every watt of power, must be captured locally or shipped at astronomical cost.
Single-point failure = death. A blackout in a city is an inconvenience; a blackout in a lunar habitat is fatal.
Communication delay breaks hierarchy. A Mars base cannot wait roughly 6–44 minutes (round trip, depending on orbital position) for Earth approval during an emergency. Centralization becomes fragility.

If vertical intensification (bigger cities, denser states) is the metric of Earth-bound history, then distributed resilience (mycelial spread, redundant nodes, autonomous local decision-making) may be the metric of post-Earth history. Not because it is morally better, but because the vacuum environment may make such topologies unusually robust — a claim I stress-test in Section 2.1.

The Fungal Hypothesis: Advanced civilizations may look less like trees competing for height and more like fungal networks competing for coverage—distributed, redundant, capable of local repair without central command. This is broadly compatible with Robin Hanson’s long-run civilizational projections (The Age of Em; Grabby Aliens), though Hanson’s models are typically driven by competitive/Malthusian dynamics among near-independent agents, whereas the hypothesis here is agnostic about whether the drivers are competition, cooperation, or a mix of both.

This framing also intersects with existing space-governance literature. Toby Ord discusses long-run space settlement as a route to existential risk reduction through redundancy (The Precipice, ch. 7), and Daniel Deudney’s Dark Skies (2020) argues the opposite: that dispersal across space may increase catastrophic and tyrannical risk rather than reduce it, because isolated, resource-constrained colonies are plausible sites for extreme centralization of power. I take Deudney’s warning seriously and address it directly in Section 2.1 — the “fungal hypothesis” is an engineering tendency, not a guarantee of political pluralism.

2. From Territorial Scale to Distributed Agency

Where Morris foregrounded energy capture, organization, information technology, and military capacity, a spacefaring civilization may require us to foreground resilience, autonomy, and the distribution of decision-making capacity.

A possible evaluative framework might track:

Number of self-sustaining, mutually independent habitats across multiple planetary bodies.
Capacity for local autonomy during communication blackout.
Ability of heterogeneous agents (biological humans, AI systems, cyborgs) to coordinate without instrumentalizing each other.

But beware Goodharting. A distributed network can still be controlled by a single will. Physical decentralization does not guarantee political decentralization. The question is not merely “Can the network repair itself?” but “Who decides the direction of repair, and in whose interest?”

It is worth distinguishing what I take to be an empirical engineering claim from a normative one. The claim that space habitats will tend toward distributed, autonomous architectures (due to communication delays and single-point failure risks) is largely descriptive. The claim that we should adopt this as our metric of “progress” is normative. I am not arguing that the latter follows deductively from the former—only that if we care about sustained civilizational capacity, the old metrics may be misleading.

This “who decides” question cannot be answered at the level of topology alone. A network diagram cannot tell you whether its nodes can refuse orders. It has to be answered at the level of agency: whether individual nodes retain the standing to contest, redirect, or exit the repair process. That is precisely the gap the freedom framework in Section 4 is built to fill.

2.1 Steelmanning centralization

The engineering claim in Section 1 deserves a direct stress test, because it is easy to romanticize decentralization without checking the conditions under which it fails.

There are at least three regimes where centralized, vertically-intensified architectures plausibly outperform distributed ones:

Early-stage colonies with thin resource margins. A single large habitat with pooled life support (one big oxygen plant, one big water recycler) may be more efficient per capita than several small redundant ones, simply because life-support hardware has strong economies of scale. Redundancy costs mass, and mass is the scarcest resource in transit. In this regime, distributing infrastructure can increase aggregate risk of failure, even while making any single failure less catastrophic.
Coordination-heavy tasks under extreme scarcity. Allocating a shared, non-substitutable resource (e.g., a single nuclear reactor’s output across a habitat) may require centralized arbitration precisely because there is no “exit” — a market or negotiated solution presumes redundancy that doesn’t yet exist.
Deudney’s political-risk case. Even where the physical architecture is distributed, isolation itself can enable centralization of power: a habitat with one exit, one life-support authority, and no external audit can become more totalizing than a networked Earth-bound state, precisely because there is nowhere to flee to and no independent press to appeal to. Physical distribution across habitats does not prevent totalizing control within a habitat.

The upshot: distributed resilience is a plausible engineering attractor for physical infrastructure, but it is neither sufficient nor obviously necessary for political pluralism, and in the early, mass-constrained phase of settlement it may not even be the dominant physical attractor. This is exactly why Section 4 insists that topology (Section 2) and agency (Sections 3–5) have to be evaluated as separate axes rather than assumed to move together. A civilization can be physically mycelial and politically monolithic at the same time; it can also be physically centralized (one big station) while preserving strong internal contestability. The normative work has to be done at the level of standing and voice, not at the level of network diagrams.

3. The Dissolution of Unified Agency

Thucydides explained war through fear, honor, and interest. Marx explained history through capital and class. Both assumed a unified intentional agent: a person, a class, a state. The historian’s job was to recover motives.

AI-mediated systems can make unified intentional agency a less adequate unit of explanation, especially when outcomes emerge from interactions among models, operators, institutions, datasets, and automated feedback loops. A lunar landing AI’s decision emerges from sensor inputs, model weights, training data, rule constraints, and real-time probabilistic inference across thousands of coupled modules. There is no single “he” who “believed” something.

Historical understanding must shift from motive recovery to boundary tracing: Which module? Which model version? Which dataset? Which permission structure? Which human deployer?

This is not merely an epistemic shift. It is a shift in the locus of agency. If consequences are produced by distributed algorithmic systems rather than unified human wills, then the concept of “freedom” must also migrate. It is no longer primarily about “doing what I want.” It becomes about being recognizable as a node that can refuse, question, and demand justification within a distributed causal loop.

This is related to, but distinct from, the corrigibility literature (Soares & Fallenstein, 2017; Christiano, 2015). Corrigibility usually asks whether a system is designed to permit correction, shutdown, or modification by its operators. The framing here asks a slightly different question: independent of whether the system permits it, does the human subject retain a standing claim to explanation and refusal? A system can be fully corrigible from an engineering standpoint while the humans downstream of its decisions have no comparable channel at all.

4. Three Components of Stellar Freedom

On Earth, freedom is typically analyzed through four dimensions: permission (am I prohibited?), capacity (can I afford it?), autonomy (is this choice truly mine?), and recognition (am I treated as a subject?).

In closed habitats with AI-mediated life support and light-speed delays, these are insufficient. We need three more.

4.1 Topological Freedom

Not the right to travel, but the right to change one’s position in the network of dependencies.

A Martian miner with adequate food and shelter who cannot access independent information, cannot change jobs, cannot leave the habitat, and cannot participate in governance is not merely poor. He is topologically imprisoned. The network structure itself is the cage.

Metrics: Can the agent alter physical location, information access, social ties, and institutional dependencies? Is there an exit?

This concept sits close to Philip Pettit’s republican account of freedom as non-domination — the idea that unfreedom consists not only in actual interference but in being subject to another’s arbitrary, uncontrolled power, even if that power is never exercised (Pettit, Republicanism, 1997). I want to be precise about what is and isn’t new here. Pettit’s diagnosis — that arbitrary power over your life, not just active interference, is the relevant harm — is exactly the diagnosis I’m relying on. What topological freedom adds is a specific claim about where that arbitrary power is now encoded: not primarily in law or in a sovereign’s discretion, but in the physical and informational dependency graph of a closed habitat — who controls the air recycler, the only outbound relay, the only employer, the only exit airlock. In a life-support-dependent environment, domination can be total and yet exercised by nobody in particular; it can be an emergent property of infrastructure ownership rather than a decision any single actor makes. Auditing non-domination in this setting therefore requires auditing dependency graphs, not just laws or contracts.

Topological freedom also overlaps with familiar concerns about exit, mobility, and association (Hirschman’s Exit, Voice, and Loyalty, 1970, is a natural reference point). Its distinctive claim is that in closed, infrastructure-dependent habitats, control over air, information, work, transport, and social access can be jointly and quietly exercised even when no single formal prohibition exists — which is harder to see, and harder to legislate against, than a single arbitrary law.

4.2 Morphological Recognition: A Precondition for Freedom

On Earth, “person” is biologically obvious. In space, it is not.

Humans may occupy remote robotic bodies, neural interfaces, or genetically modified forms. Meanwhile, complex AI systems may exhibit persistent self-models, preference structures, and reflective capacities. The question is no longer “Is it human?” but “What kind of being is entitled to non-instrumental status?”

If a non-biological agent possesses continuous identity, vulnerability to harm, stable preferences, and reflective capacity, does it merit personhood protections? Can it refuse deletion, forced copying, or memory reset?

This is essentially the question the moral patienthood literature has been circling — e.g. work on criteria for moral status and sentience (Sebo & Long, “Moral Consideration for AI Systems by 2030”, 2023; Shulman & Bostrom, “Sharing the World with Digital Minds”, 2021; Birch, “Animal sentience and the precautionary principle”, 2017, on precautionary frameworks for uncertain moral status). I am not proposing a new criterion for moral status here; I am only flagging that once “person” can no longer be read off a biological silhouette, this literature stops being a niche concern for philosophers and becomes load-bearing infrastructure for any account of freedom in a mixed human/AI civilization. Shulman and Bostrom in particular already treat this as an institutional design problem rather than a purely philosophical one — my contribution is narrower: to note that it becomes a precondition specifically for the other two components of stellar freedom, since neither topological exit nor temporal appeal means anything for an entity not first recognized as a subject capable of holding them.

This is not itself a dimension of freedom; it is a precondition for freedom. An entity not recognized as a subject cannot be free. But because this precondition is now unstable, it must be explicitly included in the framework.

4.3 Temporal Sovereignty

AI operates in milliseconds. Humans operate in seconds, minutes, years. When a system makes decisions affecting a human’s life, identity, or rights at speeds the human cannot comprehend, does the human have a right to demand that the system slow down?

This does not mean AI must wait for human approval during emergency maneuvers (attitude control, collision avoidance, life-support failure). In those cases, action must precede deliberation.

But it does mean:

Emergency automation must be bounded by pre-authorized, auditable scopes.
Non-emergency, high-stakes decisions must offer comprehensible rhythm, appeal windows, and feasible human intervention.
The definition of “emergency” itself must be auditable and contestable, because the power to declare emergency is historically the power to suspend freedom — this is essentially Carl Schmitt’s sovereign exception (Political Theology, 1922) relocated to a life-support control loop: whoever defines the emergency threshold controls the boundary of the law.

Temporal sovereignty overlaps with due process, contestability, and meaningful human control; the term emphasizes that the relevant resource is not only authority but time at a cognitively usable scale.

Without temporal sovereignty, freedom is compressed into the gaps between algorithmic cycles.

5. A Provisional Definition

Stellar freedom, provisionally:

The capacity of an agent, within a given physical topology and cognitive architecture, to understand the rules, participate in making them, alter its connections and morphological status, and demand that the system preserve time for reflection, appeal, and refusal.

Note that the traditional metaphysical problem of free will—”Could I have done otherwise?”—is transformed in this context. The question becomes: “Am I still treated as a being who must be asked?” Even if an algorithm predicts my preference and executes it, if I was never notified, never consulted, and never given the moment to say no, then the problem is not determinism. It is exclusion from the causal loop.

6. Conclusion: Direction Without Destiny

If civilization is indeed moving toward greater distribution of intelligence, higher automation, and more heterogeneous forms of embodiment, this trend does not automatically constitute progress. Distributed intelligence can enable exploration and survival; it can also enable totalizing surveillance, unaccountable power, and instrumentalization at scale — as Section 2.1 and Deudney’s argument both suggest, physical dispersal is fully compatible with political concentration.

Historical directionality cannot ask only “How far can the system expand?” It must also ask:

For whom is it expanding?
Who gets to define its goals?
When humans, augmented humans, remote proxies, and artificial systems constitute a civilization together, who retains the right to say no?

These questions have no settled answers. But if we stop asking them, we will quietly replace the metric of distributed intelligence with the metric of efficient replication—the former cares about sustainable plural participation; the latter is just a self-copying machine.

6.1 Tentative implications for longtermist priorities

I want to gesture—tentatively—at what this might mean for prioritization:

AI governance should attend not only to whether controllers can correct systems (corrigibility), but also to whether affected subjects retain standing to appeal, understand, and refuse.
Digital minds and moral patienthood may shift from philosophical fringe to institutional infrastructure as mixed human-AI societies form; work like Shulman & Bostrom (2021) already treats this as a design problem for future institutions, not a purely academic one.
Space governance should be designed early with communication-blackout autonomy, emergency-state boundaries, and audit mechanisms — and should take Deudney’s political-risk case as seriously as the standard existential-risk-reduction case for settlement (Ord, 2020), since the two point toward different design priorities.
Civilizational resilience should be measured not only by replication, expansion, or survival probability, but also by value pluralism and contestability within independent nodes — including nodes that are, for good engineering reasons, centrally architected.

These are not fully operationalized recommendations. They are invitations to treat the topology of future civilization as a governance problem, not merely an engineering one.

The following vignette is illustrative rather than predictive.

In the 2030s, an astronaut on the lunar south pole receives a life-support alert. Her AI assistant recommends a valve switch. Earth control will not hear her signal for seconds. She looks out the window. Earth hangs in the black, impossibly blue.

The report is due. She does not speak. She looks for thirty seconds more.

Thirty seconds that change no orbit, improve no efficiency. The AI does not hurry her.

Perhaps this is the smallest prototype of stellar freedom: in the gap between millisecond and century, inside the most rigid life-support system, a person still owns a slice of time that efficiency cannot consume. Enough to see where she came from. Enough to choose how to say goodbye.

I would especially welcome feedback on:

Whether “topological freedom” and “temporal sovereignty” capture concerns not already covered by existing AI governance or political theory (particularly Pettit’s non-domination and Schmitt’s state of exception, which I engage with directly above);
Which literatures I am missing, especially on digital minds, space governance, and freedom under automated systems;
Whether the Section 2.1 steelman of centralization successfully addresses the Goodharting concern, or whether the distributed/political-pluralism distinction still collapses under scrutiny;
Whether this framework has any non-trivial implications for longtermist prioritization, rather than merely renaming familiar concerns.