Abstract
Identifying the "actual cause" of an event is a fundamental challenge in AI and philosophy. Traditional definitions often fail in scenarios involving preemption or overdetermination. This paper proposes that the difficulty arises from treating causation as a monolithic concept. By formalizing distinct criteria—Strong Sufficiency, Consequentiality, Sustenance, and Contributory Cause—within functional causal models, the author provides a robust framework that resolves long-standing paradoxes and aligns more closely with human intuition.
Introduction: The Problem of Actual Cause
In causal reasoning, determining the "actual cause" in specific scenarios is notoriously difficult. Consider two arsonists starting a fire, or a prisoner executed by lethal injection while a backup firing squad stands ready. These scenarios can be represented by isomorphic (structurally identical) causal models, yet they lead to different intuitive conclusions about what "caused" the outcome. Previous definitions, such as Judea Pearl’s "causal beam" or the Halpern-Pearl "ASC" definition, often produce counter-intuitive results in "preemption" cases where one cause precedes another.
Methodology: A Taxonomy of Causal Criteria
The paper builds on Pearl’s Functional Causal Models but shifts the focus from finding a single "true" definition to providing a taxonomy of causal criteria. This approach allows for a more nuanced interpretation of events based on the specific context and the observer's focus.
- Strong Sufficiency: Captures the physical process where an event sets in motion a chain guaranteed to bring about the effect regardless of outside interference.
- Consequentiality: Represents the "but-for" condition—would the outcome have changed if the event hadn't happened?
- Sustenance: A hybrid of strong sufficiency and potential consequentiality.
- Contributory Cause: An event that is a necessary part of a larger sustaining cause, even if not sufficient or consequential on its own (e.g., a single vote in a landslide).
Key Insights: Subjectivity and Isomorphism
The most profound insight is that the "correct" cause depends on the questioner's psychological focus. Is the goal to identify the physical process (Strong Sufficiency) or the critical decision point (Consequentiality)? The fact that identical structural models yield different answers suggests that causality requires information outside the model—specifically, the criteria chosen by the observer. Furthermore, "Strong Sufficiency" is identified as the only transitive relation, aligning with our intuition about physical chains of events.
Key Results and Applications
The proposed framework successfully navigates classic philosophical thought experiments that have historically challenged causal models:
- The Train Switch: The model correctly identifies flipping a switch as "strongly sufficient" but not "consequential," matching the intuition that it only changed the path, not the arrival.
- Voting: It explains why a single vote in a landslide is not "consequential" but a block of votes can be a "strong cause."
- Explanations: By utilizing multiple definitions, the model generates more precise explanations. An explanation is treated as a hypothesized event serving as a specific type of cause for an observation.
Limitations and Future Work
While the paper identifies that psychological factors drive the choice of causal criteria, it does not formalize how humans make these choices. Additionally, the computational complexity of checking multiple variable instantiations for "Strong Sufficiency" may be high in large networks. Future research could integrate these discrete causal types with probabilistic models to handle "degrees of responsibility" more rigorously.
Summary updated