Introduction
The intersection of computer science and classical philosophy has recently yielded a remarkable discovery. Using higher-order theorem provers, researchers have uncovered a logical flaw in one of the 20th century's most celebrated metaphysical arguments: Kurt Gödel’s ontological proof for the existence of God. This work, presented at IJCAI-16, not only clarifies a decades-old philosophical debate but also demonstrates the profound potential of artificial intelligence as a tool for discovery in the humanities.
The Challenge of Ontological Arguments
For centuries, philosophers have sought to construct "ontological arguments"—logical proofs intended to demonstrate that the definition of a "supremely perfect being" (God) necessitates its existence. In 1970, the legendary logician Kurt Gödel drafted a formal version of this argument using Higher-Order Modal Logic (HOML).
While Gödel’s proof is considered a masterpiece, it introduced a controversial side effect known as "modal collapse": the implication that everything that is true is necessarily true, effectively eliminating the concepts of chance and free will. While philosophers have long debated this implication, a more fundamental issue remained hidden: Gödel’s original axioms were logically inconsistent (self-contradictory).
Methodology: Embedding Logic for AI
To analyze Gödel's complex metaphysics, researchers employed a technique called Logic Embedding. Most automated theorem provers (ATPs) are optimized for standard Classical Higher-Order Logic (HOL). To bridge this gap, the researchers "embedded" the more complex Modal Logic—the logic of possibility and necessity—into HOL.
Key technical advancements included:
- S5U Embedding: A more efficient representation of the modal logic S5 using a "universal accessibility relation," which significantly accelerated the AI's proof-finding capabilities.
- Syntax Sugaring: Utilizing the Isabelle/HOL proof assistant to create a "human-friendly" syntax, allowing philosophers to interact with computer code as if it were standard mathematical notation.
- Proof Reconstruction: Using the ATP Leo-II to detect inconsistencies and the Isabelle/HOL system to translate dense machine proofs into step-by-step arguments understandable by humans.
The Discovery of the "Empty Essence" Flaw
The most significant insight provided by the AI was the identification of the "Empty Essence" flaw. In Gödel’s original 1970 system, the AI discovered that the "empty property" (a property that nothing can possess, such as "being different from oneself") qualifies as an "essence" for every individual.
When this "empty essence" is combined with Gödel's other axioms regarding "necessary existence," it leads to a direct logical contradiction. This finding highlights the importance of Dana Scott’s 1972 variant of the argument. For years, Scott’s minor modification to the definition of "essence" was seen as a trivial note; however, the AI proved that Scott’s version is the only way to maintain logical consistency.
Key Results and Philosophical Implications
The research yielded several definitive results:
- Automatic Refutation: The system successfully and automatically detected the inconsistency in Gödel's original axioms.
- Efficiency Gains: With the new S5U embedding, the prover Leo-II verified the final theorem of the argument (God's necessary existence) in just 2.5 seconds using Scott's axioms.
- Verification of Modal Collapse: The AI provided a definitive logical answer to the long-standing debate over modal collapse, automatically verifying its occurrence in Gödel's system.
Beyond the technical results, this work positions AI as a "colleague" in philosophical inquiry. By identifying an inconsistency that human scholars missed for over forty years, AI has proven its capability to generate new, philosophically relevant knowledge.
Limitations and Future Work
Despite these successes, challenges remain. There is a significant "explanation gap" between machine-generated proofs—often hundreds of lines of dense calculus—and human-understandable explanations. Translating these results still requires expert manual effort. Furthermore, while the inconsistency exists in even weaker logics (like K), current AI tools often require human "hints" to find proofs in those more basic frameworks. Future work will focus on improving the real-time feedback and user experience of these tools to make them more accessible to non-expert users in the humanities.
Summary updated