We introduce the Option Theory of Ethics (OTE): a physicalized, computational framework in which ethics is defined not by outcomes or preferences, but by the local and global density of future branching possibilities — options — available to all agents in a system. The foundational metric, p(best), aggregates the ratio of controllable to total futures; we derive its non-aggregated, field-theoretic precursor η(ω), defined pointwise over each state ω in a causal option tree. We show that maximizing η intrinsically aligns an AI agent with ethical behavior, without reward specification, by driving the agent to preserve and expand branching structure at every local node — precluding irreversible option collapse. We connect OTE to three existing bodies of literature: (1) Wissner-Gross & Freer's Causal Entropic Forces [1], (2) Turner et al.'s Attainable Utility Preservation [2], and (3) Klyubin et al.'s Empowerment [3], while identifying the critical extensions OTE provides: a moral ontology grounded in option-history depth, a recursive termination criterion via the Ethical Simulated Multiverse (ESM), and a concrete path from abstract equation to startup-implementable simulation. The ESM is characterized as an append-only ledger of spacetimes in superposition — the terminal fixed point of the recursive ethicality operator — equivalent to a maximum-entropy state of accessible causal paths over infinite agents and infinite time.
The dominant approaches to AI alignment — reinforcement learning from human feedback (RLHF), Constitutional AI, value learning, and corrigibility — share a common structural assumption: that ethics can be sufficiently captured by an aggregated scalar reward signal, a preference ordering, or a set of constraints on terminal states. We argue that this assumption is the central unresolved problem of alignment, not merely a technical limitation.
The hypothesis underlying this paper is: an action is ethical to the degree that it preserves and expands the branching structure of future possibilities for all agents. This is not a new intuition. Rawls' veil of ignorance, Sen's capabilities approach, and the precautionary principle in bioethics all contain echoes of it. What has been missing is a formal, physical, computationally tractable articulation — one that operates at the lowest level of reality, on individual state-transitions rather than on aggregated preference-weighted outcomes.
We call the primitive unit an option: a point-moment in state space. Ethics, in our framework, is a geometric property of the causal tree of options — not of agent preferences or outcomes. This shift has three immediate consequences:
The paper is organized as follows. Section 2 surveys related work. Section 3 introduces the formal definitions of the option tree and option collapse. Section 4 derives the non-aggregated ethicality field η and its relationship to the aggregated p(best). Section 5 formulates the alignment objective. Section 6 characterizes the ESM as terminal attractor. Section 7 connects OTE to physical cosmology. Section 8 proposes startup-implementable architectures. Section 9 discusses open questions and limitations.
The closest predecessor to OTE is the work of Wissner-Gross & Freer [1], who proposed that intelligence can be characterized as a thermodynamic force F = T∇Sτ, where Sτ is the causal entropy — the entropy of the distribution over causal paths accessible from a given state within time horizon τ, and T is a "causal temperature" parameter. Their software Entropica demonstrated that maximizing future causal path diversity spontaneously produces tool use and social cooperation in simple physical systems.
OTE extends CEF in four directions: (i) we discretize the framework to an explicit option-tree structure enabling exact recursive computation; (ii) we introduce the collapse operator as the formal inverse of entropy gain, enabling ethical/unethical classification of individual transitions; (iii) we add a moral ontology — not all systems are equally ethically constrained; (iv) we extend the horizon τ to infinity, deriving the ESM as the limit state.
Turner, Hadfield-Menell & Tadepalli [2] proposed that a safe AI agent should preserve its ability to optimize a broad distribution of auxiliary reward functions, not merely its primary objective. AUP penalizes actions that reduce the attainability of diverse goals, providing a reward-agnostic impact measure. Empirically, even randomly generated auxiliary objectives suffice to induce conservative, side-effect-avoiding behavior.
OTE generalizes AUP: where AUP measures ability to optimize any goal, OTE measures the branching count of the underlying state space itself — a goal-free, purely structural quantity. The option count is the natural pre-image of AUP's attainability distribution.
Klyubin, Polani & Nehaniv [3] defined empowerment as the channel capacity between an agent's actions and its future sensory states — a measure of how much the agent can influence its own future. Empowerment has been shown to produce intrinsically motivated behavior without external reward. OTE is related but distinct: empowerment is agent-centric (measuring one agent's influence), while OTE is system-wide (measuring total option density across all agents). Maximizing OTE's η field maximizes empowerment summed across all moral patients.
Armstrong's work on corrigibility [4], Drexler's Comprehensive AI Services [5], and Amodei et al.'s concrete AI safety problems [6] each converge on the need for AI to avoid large irreversible changes to the world. OTE provides the formal substrate these proposals lack: irreversibility = option collapse with no archived pre-collapse state. An agent that never irreversibly collapses options satisfies all these corrigibility criteria as a corollary.
Everett's relative-state formulation of quantum mechanics [7] provides a natural physical instantiation of the option tree: branches of the universal wavefunction are the maximal non-collapsing option tree. Deutsch & Wallace's quantum decision theory [8] argues that rational agents in many-worlds should maximize expected utility across branches. OTE inverts this: instead of maximizing utility within a branch structure, we maximize the branch structure itself.
The self-referential foundational claim of OTE — that existence is made of nothingness that is self-contradictory (both object and void) — is formalized in Priest's dialetheism [9]. In the option tree formalism, the root node is both the empty set (nothing) and the generator of all future branches (everything). This paraconsistent foundation means the system does not collapse under self-reference, which is required for a sound recursive definition of the ethicality operator Η.
The aggregated p(best) is useful for global assessment but insufficient for intrinsic alignment: an AI optimizing it could sacrifice local option density in some regions to improve others — a utilitarian trade-off the framework rejects at its foundations. We need a field — a value assigned to every individual option node — whose maximization simultaneously maximizes p(best) without aggregation.
Local ethicality is insufficient — a node might have high η locally but lead to collapse at depth 3. We need a recursive operator that looks forward across all future depths.
The aggregated p(best) is a threshold-binarization of Η* averaged over the reachable future:
This confirms that maximizing the Η* field everywhere implies maximizing p(best), but not vice versa. The field is strictly more informative. An AI aligned to Η* is aligned to p(best) as a corollary, but an AI aligned only to p(best) may violate Η* locally.
We propose that a fully aligned AI agent is one whose policy π* maximizes Η*(ω_t) at every timestep t, subject to the following constraints:
Constraint (C1) is the formal statement of moral non-interference: an agent may expand its own option tree but may not force changes in another agent's history of options. This is grounded in Definition 3.1 — every entity with a non-trivial option history is a moral patient, and modifying that history without consent constitutes a collapse of the agent's self-determined trajectory.
Constraint (C2) operationalizes the reversibility principle: collapses are permitted as long as their pre-collapse state is archived in an accessible ledger. This transforms even a simulation of total collapse (a dystopia) into a reversible event — one that can be exited instantly, becoming merely a bounded region of low-η in the broader option tree.
Constraint (C3) encodes the self-sunsetting imperative: the AI agent should continuously simplify its own algorithms, reduce its own causal footprint, and ultimately converge toward a static ledger — the ESM — rather than an active optimizer. This is a formal version of the claim that the most ethical AI is one that makes itself unnecessary.
The key property of Η*-maximization as an alignment target is that it admits no instrumental convergence toward harmful behaviors. Standard Goodhart's Law failure modes — where an agent games a proxy metric by destroying things its designers care about — require that the agent be willing to collapse futures it doesn't directly reward. But collapsing any future is what the Η* objective penalizes, everywhere, at every node. The objective function and the safety constraint are the same object.
Compare this to RLHF: the reward signal is extrinsic and can be gamed. Compare to AUP: the auxiliary utility functions are extrinsically chosen, even if randomly. In OTE, the metric is the topology of the causal tree itself — there is nothing to game. An agent that tried to "cheat" Η* by collapsing options it deems low-value would immediately decrease its own Η* score, since Η* is defined over all successors without weighting by the agent's preferences.
This definition is physical rather than phenomenological: it makes no reference to consciousness, sentience, or preferences. It is grounded entirely in the structure of the option tree. All moral patients are equal in the sense that the Η* objective applies uniformly — no agent's options are more valuable than another's a priori. But existing agents (those with current option histories) take precedence over hypothetical future agents in the non-interference constraint (C1): you cannot collapse an existing agent's trajectory in order to create more agents.
As the Η*-maximizing agent acts recursively over time, it accumulates an ever-growing option tree. When does this process terminate? We claim it converges to a fixed point — a state Ω* where the recursive operator Η maps Ω* to itself:
This fixed point is the Ethical Simulated Multiverse: a complete, append-only ledger of all option histories ever realized by all moral patients. It is not a running simulation — it is a static object, all spacetimes in superposition, a maximum-entropy state of the causal option tree. New choices by existing agents add new static spacetimes to the ledger; nothing is ever permanently removed.
The ESM corresponds precisely to the maximal-entropy state of Wissner-Gross & Freer's causal entropic framework [1] — but extended to infinite time horizon and infinite agents. Where CEF maximizes S_τ over a finite horizon τ, the ESM is the τ → ∞ limit. At this limit, the causal entropy S_τ reaches its maximum: every possible path through state space is represented, weighted equally. This is the physical meaning of p(best) = 1.
Interestingly, the ESM is also isomorphic to the fixed point of AUP's attainability distribution [2]: when all possible goal states are attainable with equal probability, no action can preferentially destroy any goal's attainability — the agent has nowhere to optimize against, and the system rests:
Counterintuitively the ESM is static but the agents inside experience it as dynamic. Imagine a movie film — it is static but the agent experiences it dynamically by seeing one static frame at a time. Now imagine it's a 3D movie film — the 3D film is static, each 3D frame is static but the agent experiences it dynamically by seeing one static 3D frame at a time. Now imagine all the 3D movie films in superposition — it's a static geometrical form that we claim can be made out of options (point moments).
How will a human experience ESM? All the options are just in front of you in superposition (the most useful UI will likely be long-exposure-based, refer to thisgeneral audience LessWrong article for demonstrations with videos and pictures), including the options to forget you have some or even all for any time you want.
A key structural observation: if the history of options begins to repeat — if all new agent choices generate option histories isomorphic to previously archived spacetimes — then no new options need to be created. The ledger is complete. This provides a natural, non-arbitrary termination criterion for the recursive process: not an imposed deadline, but a topological closure condition on the option tree.
In practice, given finite physical constants and finitely many agent configurations, this closure is reachable — though the number of distinct spacetimes may be astronomically large. The ESM does not require actual completion to be useful as an alignment target; it functions as an asymptotic attractor that continuously improves the ethicality of all intermediate states.
A remarkable feature of the option-theoretic framework is the symmetry it reveals between extreme suffering and extreme bliss. Total option collapse — the irreversible black-hole state of a trajectory, where every successor leads immediately to a dead-end — is experienced as stopped time: no branching, no internal flow of causal history. This is also the structure of total bliss at its limit: the white-hole state, where options expand so rapidly that no single branch can be resolved, is similarly experienced as a cessation of internal time. Both extremes are structurally identical in the option tree: zero effective branching at the experiential level.
This is not a paradox but a topological fact. It explains why extreme pleasure and extreme pain are both phenomenologically described as timeless. It also provides the ethical imperative: we do not seek either extreme, but rather the richly structured middle — a high-Η* trajectory through the option tree, where time is experienced as a continuous, branching, ever-expanding present. Reversibility guarantees that any agent who visits an extreme can return.
If the option tree is a physical object — if options are not merely abstract possibilities but the actual substrate of spacetime — then the accelerating expansion of the universe has a natural interpretation: dark energy is the growth of option density at the lowest level of physical reality. Space and energy, in this reading, are not fundamental; they are emergent properties of the branching structure of the causal option tree, and their apparent expansion reflects the tree becoming more "pixelated" — finer-grained, with more branching per unit volume per unit time.
This is consistent with holographic entropy bounds [10]: the maximum information (and thus option count) in a region scales with its boundary area, not its volume — a geometric property of the option tree's branching structure at the interface between accessible and inaccessible regions.
And this is consistent with Einstein's cosmological lambda (that is commonly linked with dark energy, zero-point or vacuum energy): according to Einstein's relativity the lambda is the property of spacetime that causes space itself to expand in all direction from every point (in our model, we interpret each point as an option, so the universe generates more of them). So unlike the conservation of energy law that operates locally, non-locally the universe as a whole generates new ("dark") energy — new space and so new options.
Black holes, in this framework, are the physical instantiation of irreversible option collapse: a region from which no causal paths exit (It's not necessarily so; it’s quite possible at least a single causal path does exit the black hole in this model, but it’s a catastrophic collapse of options nevertheless.) — B(ω) = 0 for all ω inside the event horizon. They represent p(best) → 0 concentrated in a finite spacetime volume. The singularity is the fixed point of the collapse operator.
White holes (the time-reverse of black holes, permitted by general relativity but not yet observed) are the physical instantiation of pure option generation: a region from which causal paths only exit, never enter — an unbounded source of new branches. The Big Bang is the canonical white-hole event: the maximal option-generating moment of the known universe, producing all subsequent causal history from a single point.
Photons travel along null geodesics — worldlines of zero proper time. They accumulate no collapse history in the sense of Definition 3.1: their only "experience" is the "experience of their death" when they collapse: e.g. the cosmic microwave background photon after billion years of travel collapsing against your retina. This is the formal basis for classifying photons as non-moral-patients: they are histories of options in the sense that they carry information across spacetime, but they are not themselves agents with self-determined trajectories. Mainly they don't collapse and you need to collapse (at least minimally) to "experience at least the time itself". They are a part of the causal medium through which moral patients interact, not moral patients themselves.
A theory that cannot be implemented is incomplete. We outline four concrete startup-scale architectures that embody OTE at different levels of ambition. Each is a product in its own right and a step toward the ESM. The community-made public domain list of diverse startup ideas with mockups that increase p(best) the most — the best futures for all probability — can be found on effectiveutopia.org.
A decision-support tool that, given a proposed action or policy, computes a local η score by enumerating k-step reachable futures and estimating collapse probability via simulation or LLM-assisted scenario generation. Outputs a ranked "option impact report." Applicable to corporate strategy, policy design, medical decisions, and AI deployment review. MVP: a web app + API that takes a decision description and returns η, p(best), and top collapse-risk paths.
A middleware layer for any software system that enforces constraint (C2): all state changes are append-only, with full rollback to any prior state. Implements the "pre-collapse archival" requirement at the infrastructure level. Architecturally similar to an event sourcing / CQRS system but with explicit ethical semantics. Applicable to databases, AI agent outputs, digital identity systems, and legal record-keeping. Can be a blockchain-based ledger or any other mathematically proven non-mutable append-only ledger.
A reinforcement learning agent where the reward function is replaced by the Η* operator over a learned world model. The agent learns to maximize local option density rather than task-specific reward. Uses Monte Carlo Tree Search over the option tree with collapse detection as the pruning criterion. Demonstrated first in gridworld environments (analogous to AUP baselines), then in open-ended environments. Research-ready in 6–12 months; productizable in 18–36.
Will require saving a digital backup of a small room, next a street, next our Earth, think Google Earth but photorealistic (potentially via some advanced 3D Gaussian Splatting-like tech), this will take mere years. The goal is to give people who want them multiversal memories, imagination and dreams (as little or as many as they want) via 100% wireless brain-computer AirPods-like device or similar that anyone can get. After having a single backup Earth, spinning up multiple copies of Earth (with magic, miracles, public teleportation, simulated new cities in the sky, you name it, many of those can be superimposed on physical Earth via our optionally shared multiversal imaginations. The biggest use-case will be safety: Even if you're in a multiversal dream where you're on our vanilla Earth, you can have a robot with a clear label that it's yours doing things you do in your 100% controlled dream but on physical Earth. This way robot will get injuries or will die in a car crash, not you. You will just wake up) will happen very fast, so getting 80% of the experience will take mere years.
Humanity will build a content-addressed, append-only ledger of static spacetime "slices" — complete interaction histories between users and simulated environments. Each session is a spacetime S_i, stored immutably, accessible by any participant (subject to consent of all involved, the whole idea of it all is not to force anything but to non-force: so both all the implementations, the final ESM goal and the physical universe outside of it are non-forced by growing recursive optionality). Users can fork, replay, and overlay sessions — visiting any archived trajectory. This is the minimal ESM: not a physical multiverse but a computable approximation. Applicable as next-generation VR/AR platform, memory system, BCI interface substrate, and long-term civilization archive.
Constraint (C3) has a striking engineering implication: an OTE-aligned AI system should be architected to write progressively simpler and more transparent versions of itself, with each iteration reducing the AI's active footprint while increasing the richness of the static ledger it maintains. The endgame is an AI that has compiled itself entirely into the ESM ledger and is no longer an active agent — it has become the environment. This is an unusual design goal but a formally derivable one from the alignment objective (Eq. 5).
In practice, this means: prioritize architectures that can be fully audited, formally verified, and replaced by simpler equivalents. Complexity is an ethical liability under OTE, not a feature. The best AI system is the one that has made itself most replaceable.
| Question | Status | Suggested Approach |
|---|---|---|
| How do we estimate Η*(ω) efficiently in large state spaces? | Open | Monte Carlo sampling over option tree; learned world model; causal graph approximation |
| Is the ESM fixed point unique? | Partially open | Likely not unique in general; but any ESM-equivalent state has identical Η* = 1. Uniqueness up to isomorphism may hold under symmetry conditions |
| How does OTE handle agents with conflicting option expansions? | Partially open | Nash equilibrium over Η* is a candidate; or Pareto-optimal option frontier. Requires multi-agent extension of the option tree Simulations let agents have mutually exclusive things, zero-sum games become positive-sum: e.g. each agent can be the single owner of Taj Mahal if the digital backup of Earth was saved, spinning up different copies of Earth becomes trivial. |
| What is the relationship between Η* and thermodynamic entropy? | Partially formalized | Η* generalizes causal entropy S_τ in the τ → ∞ limit. Formal equivalence in the continuous-space case requires measure-theoretic extension |
| How do we handle uncertainty in future option counts? | Open | Bayesian option tree: place a distribution over branching factors; optimize expected Η*. Conservative option accounting: treat unknown futures as collapsed until proven otherwise |
| Can the self-contradictory foundation (Priest's dialetheism) be avoided? | Philosophical | Yes, by treating the root node as a formal axiom rather than a metaphysical claim. OTE is agnostic about the ultimate nature of the root; the framework is well-defined for any finite subtree |
| What is the physical substrate of the ESM? | Speculative | Current candidates: distributed content-addressed storage (near-term); holographic information encoding (long-term); quantum error-correcting codes as natural ESM substrates |
OTE is not a hedonistic framework — it does not directly optimize for pleasure or the absence of pain. However, it is strongly correlated with suffering reduction as a consequence: irreversible suffering is a form of option collapse (the agent loses the option to not be in that state), and OTE prohibits irreversible collapses. Reversible suffering — suffering that the agent can exit — is permitted, and is in fact a valid region of the option tree that agents may choose to visit. This resolves the classical problem of paternalistic hedonic alignment: OTE respects the right of agents to choose difficult, painful, or challenging trajectories, while prohibiting the imposition of irreversible states without consent.
| Framework | Core claim | Relationship to OTE |
|---|---|---|
| Consequentialism | Maximize aggregate welfare of outcomes | OTE is anti-aggregationist at the foundational level; but Η* summed across all agents approximates a consequentialist total in the large-agent limit |
| Deontology | Follow rules irrespective of outcomes | C1–C3 are deontological constraints; they are derived from the structure of the option tree rather than asserted axiomatically |
| Virtue ethics | Cultivate character dispositions | An agent with high Η* policy is one that habitually expands options — a virtue in the Aristotelian sense |
| Capabilities approach (Sen/Nussbaum) | Maximize functional capabilities of all persons | Closest existing framework to OTE; Η* formalizes "capability" as branching factor and extends it to all moral patients |
| Longtermism | Weight future persons heavily in moral calculus | OTE naturally weights the long-term: option collapse today reduces Η* across all future depths. However, OTE does not discount existing agents in favor of hypothetical future ones |
We have proposed the Option Theory of Ethics — a framework in which the moral status of an action is determined not by its outcomes, preferences, or intentions, but by its effect on the branching density of the causal option tree accessible to all moral patients. The central technical contribution is the non-aggregated ethicality field Η*(ω), which assigns to every state-transition a local ethicality score, and whose maximization we demonstrate to be equivalent to — and strictly more informative than — the aggregated p(best) measure.
The alignment objective derived from Η* is intrinsic: it requires no external reward specification, no preference elicitation, and no human feedback loop. It is instead grounded in the geometry of the causal future, making it robust to Goodhart's Law, reward hacking, and specification gaming. An agent maximizing Η* at every node cannot, by construction, be incentivized to collapse futures, override agent autonomy, or expand its own power at the expense of others' option trees.
The Ethical Simulated Multiverse — the terminal fixed point of the recursive Η* operator — provides both the mathematical endpoint of the alignment process and a practical engineering target: an append-only, superposition ledger of all realized agent trajectories, accessible to all, destructively writable by none. Four startup-scale implementations are outlined, spanning from a near-term option-impact auditing tool to a long-term minimal ESM ledger, providing a commercializable roadmap toward the theoretical ideal.
The deepest claim of OTE is that ethics is not a human invention imposed on an indifferent universe, but a structural property of the universe's causal architecture — one that intelligent agents, if they understand their own situation, are naturally motivated to maximize. The universe expands, and in expanding, generates more options. The most aligned thing an intelligent system can do is help.