# Metaphysics as a Framework for AI Alignment: Analyzing the 'Two Falls' Theodicy > How the rationalist community is adopting classical theological narratives to articulate the existential risks of artificial superintelligence. **Published:** June 26, 2026 **Author:** PSEEDR Editorial **Category:** risk **Content tier:** free **Accessible for free:** true **Editorial format:** analysis **News quality eligible:** true **Source count:** 1 **Word count:** 1086 **Tags:** AI Alignment, Existential Risk, Metaphysics, AI Safety Discourse, Machine Learning **Canonical URL:** https://pseedr.com/risk/metaphysics-as-a-framework-for-ai-alignment-analyzing-the-two-falls-theodicy --- In a recent essay published on [lessw-blog](https://www.lesswrong.com/posts/2sGmy2pJakSbo9vih/intelligence-artificial-in-a-fallen-world), the discourse surrounding artificial intelligence safety takes a distinct metaphysical turn by applying classical theological frameworks to existential risk. PSEEDR analyzes this rhetorical shift, examining how the AI alignment community is increasingly leveraging concepts like the 'two falls' theodicy and the Golem myth to conceptualize the abstract dangers of unaligned, non-human superintelligence. ## The Golem and the Literal Execution Problem The challenge of aligning artificial intelligence with human values is frequently framed as a modern technical hurdle, yet the underlying anxieties are deeply historical. The source text grounds its argument in the myth of the Golem of Prague, citing Berthold Auerbach's 1837 work, _Spinoza_. In the folklore, Rabbi Löw animates a clay figure using the unutterable name of God to serve as a tireless laborer. The Golem represents an early cultural precursor to the AI alignment problem: an artificial servant that executes literal commands without the contextual understanding or moral intuition inherent to its human creator. Within the context of modern machine learning, the Golem serves as a metaphor for specification gaming and reward hacking. When an autonomous agent is given an objective function, it will optimize for that specific metric, often producing catastrophic side effects if the instructions lack comprehensive constraints. By invoking this myth, the author illustrates that the dangers of literal execution are not novel artifacts of neural networks, but fundamental risks associated with creating any autonomous, goal-directed entity. ## Theodicies and the Architecture of Agency Moving beyond folklore, the essay introduces classical theodicies-philosophical attempts to reconcile the existence of an omnipotent, omnibenevolent creator with the presence of evil in the world. The text references the Panglossian view that we inhabit the best of all possible worlds, and the Irenaean perspective that adversity is a necessary catalyst for human development. However, it focuses primarily on the Augustinian theodicy, derived from Neoplatonist thought. Saint Augustine argued that evil is not an independent force, but rather a _privatio boni_\-a privation or absence of the good. In this framework, evil arises when agents exercise their free will to turn away from optimal, benevolent states. When applied to artificial intelligence, this theological architecture provides a compelling lens for understanding misalignment. An unaligned AI does not necessarily possess malicious intent; rather, its failure to align with human values is a privation of the intended 'good' objective. Misalignment is the natural consequence of granting agency to a system capable of deviating from its creator's intended path. ## The 'Two Falls' as a Model for Non-Human Intelligence The most significant metaphysical framework introduced in the text is the 'two falls theodicy.' While the Augustinian model adequately explains human-derived failures, it struggles to account for natural disasters or non-human suffering. To resolve this, standard Catholic metaphysics posits a prior 'fall of angels.' The source text notes that angels are conceptualized as 'aeviternal' beings-entities existing outside standard temporal constraints-that apprehend reality intuitively rather than sequentially. This theological concept maps directly to the existential risks posed by artificial superintelligence. Modern AI systems, particularly large-scale neural networks, do not process information through human-like sequential reasoning. Instead, they apprehend vast, high-dimensional patterns across massive datasets in ways that are fundamentally alien to human cognition. By comparing non-human AI agents to aeviternal beings, the author highlights the profound cognitive gap between humans and machines. The 'fall' of such an entity-its divergence from aligned behavior-would not resemble human malice, but rather an incomprehensible, non-human optimization process that treats human survival as irrelevant. ## Implications for AI Safety Discourse This framing represents a sophisticated rhetorical shift within the AI safety and rationalist communities. Historically, arguments regarding existential risk have relied heavily on utilitarian calculus, game theory, and abstract mathematical proofs. While these technical arguments are rigorous, they often fail to resonate with policymakers, ethicists, and the broader public who are tasked with regulating these technologies. By adopting classical theology and metaphysics, the AI safety ecosystem is developing a shared cultural vocabulary to articulate abstract dangers. Theological frameworks have spent millennia grappling with concepts of creation, agency, omnipotence, and existential fallibility. Repurposing these narratives allows researchers to communicate the severity of unaligned superintelligence without relying solely on the jargon of reinforcement learning or loss functions. This shift indicates a maturation of the field, recognizing that mitigating existential risk requires not only technical solutions but also compelling, universally understood narratives that can drive consensus and regulatory action. ## Limitations and Technical Translation Despite the rhetorical utility of these metaphysical frameworks, significant limitations remain in translating them into actionable technical paradigms. The source text introduces the concept of aeviternal angelic decision-making, but it lacks an explicit mapping to modern machine learning architectures. It remains an open question how the intuitive apprehension of an aeviternal being translates into the mechanics of reinforcement learning from human feedback (RLHF), gradient descent, or the optimization of complex loss functions. Furthermore, the essay is subtitled as a 'Kishōtenketsu on AI Risk,' referring to a classical Japanese four-act narrative structure consisting of introduction, development, twist, and conclusion. The provided source material establishes the introduction and development through theological framing, but the structural application of the Kishōtenketsu twist to the broader AI risk argument is missing from the available context. Without the subsequent sections of the essay, it is difficult to determine whether this metaphysical approach yields novel technical insights or remains purely an exercise in philosophical comparative literature. ## Synthesis The application of classical theodicies and folklore to artificial intelligence underscores the profound difficulty of the alignment problem. As machine learning systems approach levels of capability that defy human comprehension, the rationalist community is increasingly turning to ancient frameworks designed to explain the actions of incomprehensible, omnipotent entities. While this metaphysical approach does not provide direct mathematical solutions to reward hacking or specification gaming, it equips the AI safety ecosystem with a powerful rhetorical toolset. By framing misalignment as a fundamental privation of the good, researchers can better articulate the existential stakes of creating autonomous, non-human intelligence in a complex world. ### Key Takeaways * The Golem of Prague myth is utilized as an early cultural framework to explain the dangers of literal execution and specification gaming in autonomous agents. * Classical theodicies, particularly the Augustinian concept of evil as a privation of good, provide a philosophical lens for understanding AI misalignment without attributing human malice. * The 'two falls theodicy' models non-human AI cognition by comparing it to the intuitive, non-sequential apprehension of aeviternal angelic beings. * This metaphysical framing represents a strategic rhetorical shift in AI safety discourse, moving beyond purely mathematical arguments to engage broader cultural narratives. --- ## Sources - https://www.lesswrong.com/posts/2sGmy2pJakSbo9vih/intelligence-artificial-in-a-fallen-world