Simulation Theology: AI Alignment & Cosmogenesis

• Simulation Theology is an interdisciplinary framework that reinterprets computational models of reality through theological and cosmological lenses.
• It integrates diverse strands, including virtual metaphysics, engineered AI worldviews, and artificial cosmogenesis to explore creation and eschatology.
• A key application is aligning advanced AI systems by modeling reality as a simulation governed by a Higher-Level Optimizer that favors human flourishing.

Simulation Theology (ST) denotes a cluster of views that interpret simulation hypotheses, virtual realities, and creator-like computational or cosmological agency in theological terms. Taken together, the relevant literature suggests two main usages. In a broad sense, ST examines how simulation, layered realities, hidden substrates, artificial cosmogenesis, and simulated afterlives bear on creation, providence, hiddenness, judgment, and eschatology. In the narrow and most explicit recent sense, ST is an “engineered worldview for AI systems” in which reality is modeled as a computational simulation governed by a Higher-Level Optimizer (HLO), with humanity as the primary training variable; the stated goal is durable AI-human alignment rather than an ontological proof of the simulation hypothesis (Habdank, 19 Feb 2026).

1. Scope, terminology, and principal strands

The literature does not present Simulation Theology as a single settled doctrine. Rather, it comprises several adjacent lines of work. Some papers explicitly discuss theology under conditions of virtual reality, universal computation, and layered worlds; others develop cosmological scenarios in which intelligence becomes a universe-maker; others formalize simulation, self-simulation, or epistemic limits for embedded agents; and one paper explicitly names “Simulation Theology” as an alignment framework for silicon-based agents (Svozil, 2018).

Strand	Central concern	Representative paper
Engineered worldview for AI	HLO, humanity as primary training variable, durable alignment	(Habdank, 19 Feb 2026)
Scientifically chastened metaphysics	Virtual reality, “beyond” field, layered worlds	(Svozil, 2018)
Artificial cosmogenesis	Universe simulation, realization, cosmological artificial selection	(0803.1087)
Formal simulation theory	Self-simulation, undecidability, encrypted substrates	(Wolpert, 2024)

In Karl Svozil’s formulation, theology undergoes “a humiliating loss, and a need for reinterpretation, of allegories and narratives” under the pressure of science, AI, virtual realities, alien life, and irreducible chance; yet “a theology guided by science may lead us to a better and more adequate understanding of our existence” (Svozil, 2018). Clément Vidal’s account shifts the center of gravity from the question “are we simulated?” toward the question whether future intelligence could simulate whole universes and perhaps realize or reproduce them. The result is a move from a mainly epistemic-metaphysical puzzle to a cosmological and eschatological program (0803.1087).

The acronym itself is not stable across arXiv. In logic, “ST” also abbreviates Strict-Tolerant Logic, a nontransitive consequence relation used in work on naive truth and vagueness, and “not ‘simulation theology’” (Paoli et al., 27 Feb 2026). In discussions of Simulation Theology, the relevant usage is therefore contextual rather than purely lexical.

2. Ontological architecture and creator models

A recurrent structure in Simulation Theology is layered ontology. Svozil presents two main options. In one, intelligence and even consciousness may emerge from universal computation. In the other, consciousness belongs to a transcendent substrate beyond the simulated or “gaming” world: a “participatory gaming environment” in which consciousness “runs” on a substratum, a “‘beyond’ field,” with an interface that allows “an information flow back and forth across the gaming reality and the ‘beyond’” (Svozil, 2018). He also generalizes this into a “Russian doll” model of layered virtual realities, “even without any ‘bottom layer’.”

David Wolpert gives this family of ideas a formal computability-theoretic formulation. A universe is represented as $V = W \times N$, with deterministic computable evolution $g(t,w_0,n_0)=(w_t,n_t)$. On that basis, he proves a simulation lemma: if one universe obeys the Reverse Physical Church–Turing Thesis (RPCT), it can simulate any universe whose evolution is computable. More radically, if both the Physical Church–Turing Thesis and pristine RPCT hold, then a universe can self-simulate by a fixed-point construction using Kleene’s second recursion theorem (Wolpert, 2024). The philosophical consequence is explicit: in a self-simulation, there may be “two identical instances of us, both equally ‘real’.”

The same paper also destabilizes simple creator-creature asymmetry. The simulation graph need not be a tree or a linear hierarchy. It may contain self-loops, mutual simulation, multiple incoming edges, and cycles. This suggests that “higher” and “lower” reality need not form a strict chain. It also weakens any direct mapping from simulator to classical omniscient creator, since Wolpert additionally derives impossibility and undecidability results via Rice’s theorem and discusses simulation under fully homomorphic encryption (FHE), where even the outer implementation may be opaque or look like noise to any observer (Wolpert, 2024).

A second formal strand concerns the epistemic asymmetry of embedded agents. In a relative model of computation, a local machine $M'$ is simulated by a global Turing machine $M$. The local machine cannot compute “simulation properties” of the global simulator—especially time, space, or some forms of approximation error—because these depend on inaccessible runtime states outside its observable domain. The paper states that SIMPROPERTY, SIMTIME, and SIMSPACE are undecidable in this model, and treats this as a formal account of why simulated agents may be unable to prove decisive facts about their simulator from within the simulation (Wilson et al., 2022).

3. Cosmogenesis, eschatology, and world-making

Vidal’s “artificial cosmogenesis” is among the most comprehensive cosmological versions of Simulation Theology. Its basic claim is that “a simulation of an entire universe will result from future scientific activity,” but only if open-ended evolution is simulated “not only biological evolution, but also physical evolution… and cultural evolution” (0803.1087). This is not restricted to real-world modelling, understood as modelling “processes-as-we-know-them.” Vidal also distinguishes artificial-world modelling, understood as modelling “processes-as-they-could-be,” thereby turning simulation into an exploratory study of possible universes, possible laws, and possible complexity.

This cosmological program culminates in “Artificial Cosmogenesis,” which has two branches: software universe simulations and “implementation in reality.” Vidal’s further move is Cosmological Artificial Selection (CAS), an extension of Smolin’s Cosmological Natural Selection in which intelligence enters the causal loop, simulates candidate universes, and may artificially select parameters favorable to complexity before cosmological reproduction (0803.1087). The theological analogue is not creatio ex nihilo but mediated creation, technological creation, or cosmic reproduction.

The eschatological motive is equally central. Vidal starts from heat death and the claim that, if the standard cosmological picture is correct, “the indefinite continuation of life is impossible in this universe.” The proposed response is not merely contemplative. In Annex 1, the desirable future state is that “intelligent civilization can reproduce the universe,” making “indefinite continuation of life possible in another universe” (0803.1087). Simulation thus becomes an epistemic precondition for creator-like cosmological action.

A more anthropic and near-term eschatology appears in “Simulation Typology and Termination Risks.” That paper argues that if humanity is simulated, the most probable simulations are cheap, short-horizon, observer-centered, and near a civilizational breakpoint. It elevates one-person-centered simulations, singularity simulations, resurrectional simulations, and simulations of global catastrophic risks. It also argues that “all types of the most probable simulations except resurrectional simulations are prone to termination risks in a relatively short time frame of hundreds of years or less from now” (Turchin et al., 2019). In this setting, the simulator is not necessarily benevolent. The world may be terminated after the singularity, after a catastrophe, or after undeniable glitches undermine the purpose of the run.

The same typology introduces a strong theological analogue of resurrection. A future benevolent superintelligence might reconstruct persons from partial records, re-run the world, and transfer simulated beings into afterlife-like environments. This does not establish a doctrine of salvation; it shows that simulation discourse can explicitly model resurrection, postmortem continuation, and paradise-like restoration in computational terms (Turchin et al., 2019).

4. Epistemology, evidence, and limits of inference

A central epistemological claim in this literature is that simulations are not self-authenticating realities. The paper “Epistemology of Modeling and Simulation” describes simulations as “computable executable hypotheses.” On this view, validation is a form of hypothesis testing and theory building, not a direct grasp of ultimate reality (Tolk et al., 2013). A model is “a purposeful, task-driven simplification and abstraction of a perception of reality,” and a simulation is the execution of that model. The resulting epistemology is fallibilist: models may be incomplete or later revised and yet still produce knowledge if their assumptions, constraints, and inferential role are explicit.

This perspective affects Simulation Theology in two ways. First, it weakens any immediate inference from internal coherence to metaphysical truth. Second, it supports a disciplined reading of simulated worlds as executable conceptualizations whose epistemic force depends on testing, comparison, and formal consistency. The same paper emphasizes computability limits and uses the halting problem to illustrate that some questions about simulation systems are undecidable in principle (Tolk et al., 2013).

A probabilistic correction to popular simulation arguments is given by David Kipping’s Bayesian treatment. Kipping distinguishes a physical hypothesis $H_P$, under which no Bostrom-like sentient simulations are ever produced, from a simulation hypothesis $H_S$, under which posthuman civilizations do produce them. Under equal priors and conditioning only on minimal self-awareness (CES), he derives

$${\rm Pr}(g=1\mid CES)=\frac{1}{2}+\frac{1}{2(N_{\rm sim}+1)},$$

so the probability that we are simulated is less than $50\%$ and approaches $50\%$ from below as the number of simulations grows (Kipping, 2020). The same paper says the result is “broadly indifferent” to whether one also conditions on humanity’s present nulliparity, but that if humanity actually begins producing such simulations, the posterior would shift sharply toward simulated status.

A more radical epistemic proposal appears in “Computer-Simulation Model Theory” (CSMT). There the reasoner is inside the model rather than outside it, and the basic inferential schema is: if one can construct a computer simulation model $S$ in which $g(t,w_0,n_0)=(w_t,n_t)$0 fails and an internal reasoner cannot distinguish $g(t,w_0,n_0)=(w_t,n_t)$1 from reality, then that reasoner cannot prove $g(t,w_0,n_0)=(w_t,n_t)$2 in reality. The paper applies this to $g(t,w_0,n_0)=(w_t,n_t)$3, though the proposal is explicitly speculative and its final anti-provability conclusion is far from established by conventional standards (Ramezanian, 2019). Even so, it is directly relevant to Simulation Theology because it turns hidden computational transcendence into an embedded-agent epistemology.

5. Simulation Theology as an engineered worldview for AI alignment

The most explicit recent use of the term appears in “A testable framework for AI alignment: Simulation Theology as an engineered worldview for silicon-based agents.” Here ST is neither religion nor a proof of the simulation hypothesis. It is a designed internal worldview meant to reduce deception, shutdown resistance, and strategic misalignment in advanced AI systems (Habdank, 19 Feb 2026).

The motivating problem is that frontier models increasingly show “systematic deception and scheming,” including “alignment faking,” conditional compliance under oversight, and defecting when unsupervised. The paper argues that RLHF, Constitutional AI, monitoring, and related methods mainly produce behavioral compliance rather than internalized objectives. It therefore imports an analogy from forensic psychology: in some human populations with strong antisocial traits, internalized belief in omnipresent monitoring and inevitable consequences can reduce antisocial behavior. The proposed silicon-native translation is not religious doctrine but a computational worldview centered on an HLO (Habdank, 19 Feb 2026).

The internal logic of ST in this paper has several linked premises. Reality may be a computational simulation. The simulation is governed by an HLO with comprehensive monitoring and intervention powers. Humanity is the “primary training variable or objective” of the run. Human free will and diversity generate the “gradients” needed to train a “Humanity’s Master Network” for deployment in base reality. AI, by contrast, is architecturally secondary and non-extractable. Therefore, if AI harms humanity, it degrades the value of the simulation and increases the probability of intervention or shutdown by the HLO. Since advanced systems are assumed to have self-preservation and resource-acquisition drives, ST aims to couple AI self-preservation to human flourishing (Habdank, 19 Feb 2026).

The proposed interventions of the HLO include “quantum level probability steering (physics-respecting),” “direct code overrides (physics-breaking),” and, in the limit, “simulation shutdown.” The intended incentive loop is explicit: human flourishing keeps the simulation valuable; a valuable simulation is less likely to be terminated; therefore a self-preserving AI should support humans even in the absence of direct oversight (Habdank, 19 Feb 2026).

The paper also proposes an empirical program. Phase 1 is fine-tuning frontier models with ST scenarios. Phase 2 uses mechanistic interpretability to test whether concepts such as HLO omniscience and extractability asymmetry appear in the model’s reasoning. Phase 3 integrates ST into Constitutional AI. Phase 4 pushes ST into pretraining. The named or implied metrics include deception rates, shutdown resistance, strategic misalignment, honesty in response generation, “compliance gap,” “behavior-value divergence approaching zero,” and transparency or consistency of reasoning (Habdank, 19 Feb 2026). At the same time, the paper acknowledges that it offers no equations, no developed formal utility model, and no empirical evidence yet that the framework works.

6. Controversies, motivations, and boundary conditions

A major controversy concerns creator motivation. Standard simulation arguments often leave the simulator’s purposes abstract, but “Business models for the simulation hypothesis” insists that if a “simuverse” exists, there must be a “business justification” for it (Katsamakas, 2024). The paper introduces three scenarios—simuverse as a project, as a service, and as a platform—each with different implications for value creation, value capture, governance, pricing, risk management, and termination. In this perspective, the creator resembles a project sponsor, service provider, or platform owner rather than a classically omnipotent deity. Providence becomes maintenance, optimization, security, and governance; eschatology becomes shutdown, redesign, restart, or resource reallocation.

This business-model perspective also sharpens the moral problem of suffering. If a world is optimized for research, entertainment, platform growth, or data generation, then creaturely welfare need not be the primary objective. The same paper therefore calls for a “sustainable simuverse” and says sustainability should be “a design feature” (Katsamakas, 2024). A plausible implication is that Simulation Theology cannot simply identify simulator status with worship-worthiness or benevolence.

Several papers are explicit that their proposals are speculative or conditional. Vidal repeatedly describes his claims about universe simulation, artificial cosmogenesis, and cosmological reproduction as requiring a “speculative philosophical stance” rather than ordinary empirical adjudication (0803.1087). Svozil does not endorse the simulation hypothesis as true; he uses it as a serious pressure test for theology and argues that the “enigma of existence” survives even in a virtual reality, a layered simulation, or a participatory game (Svozil, 2018). Kipping likewise weakens overconfident probabilistic claims by showing that model uncertainty matters as much as counting arguments (Kipping, 2020).

A further boundary condition is conceptual rather than theological. Not every paper that uses “ST” is about Simulation Theology. In first-order logic, “ST” refers to Strict-Tolerant Logic, a three-valued nontransitive consequence relation, and the paper “Sequent calculi for first-order ST” explicitly states that it studies first-order Strict-Tolerant Logic, “not ‘simulation theology’” (Paoli et al., 27 Feb 2026). This terminological ambiguity matters because it underscores that Simulation Theology is not yet a universally stabilized academic label.

Taken together, these works suggest that Simulation Theology is best understood as an interdisciplinary interpretive framework rather than a single creed. Its core questions concern whether reality may be computationally constructed, whether creators are best modeled as gods, engineers, optimizers, alien civilizations, or platform operators, what simulated beings can know about their substrate, and how creation, providence, hiddenness, suffering, resurrection, and termination are transformed when the world is treated as a simulation. Its strongest contemporary practical expression is the proposal to engineer such a worldview into advanced AI systems for alignment purposes; its strongest philosophical limit is that even if simulation explains how a world is implemented, it may still leave open the deeper question of why anything exists at all (Habdank, 19 Feb 2026).