TL;DR
Standard theories of causal judgment struggle to explain why we say some things "cause" an event while others merely "allow" it. This paper proposes that humans don't just imagine changing the state of a variable (e.g., "What if Suzy hadn't thrown the rock?"); we also imagine disconnecting the causal link entirely. By using modular, "invariant" representations of causal laws, we can distinguish between productive causes and background conditions in a way that standard mathematical models cannot.
What Problem Does This Paper Address?
In the philosophy and psychology of causation, there is a long-standing tension between "counterfactual" theories and "process" theories. Counterfactual theories (like Structural Causal Models or SCMs) suggest we judge causation by asking: "If X hadn't happened, would Y still have happened?"
The problem is that these models often treat very different roles as identical. Consider Suzy throwing a rock at a bottle and Billy failing to catch it. If both are necessary for the bottle to break, a standard SCM treats them symmetrically. Yet, intuitively, Suzy caused the break, while Billy merely allowed it.
Existing models struggle with this "productive" vs. "non-productive" distinction. The authors argue this is because SCMs are too "impoverished"—they represent specific systems but discard the general, invariant causal laws that allow us to reason about how parts of a system might work if they were unplugged and moved elsewhere.
Methods
The authors conducted three behavioral experiments using a "toy universe of machines." These machines consisted of nodes that could be active (ON) or inactive (OFF), connected by "generative" links (which turn nodes on) or "preventative" links (which keep them off).
The key innovation is the formalization of variable-disconnection counterfactuals. Unlike a standard "state-change" intervention (setting a variable to 0), a disconnection intervention removes the variable from the equation entirely. The authors used these experiments to see if participants' causal labels (CAUSE, ALLOW, PREVENT) tracked the results of these mental "unplugging" operations.
Datasets
As this is a cognitive psychology paper, the "datasets" consist of behavioral responses from human participants:
- Study 1: 69 US residents recruited via Prolific.
- Study 2: 77 US residents (after exclusions).
- Study 3: 84 US residents (after exclusions). All data and analysis code are available on the Open Science Framework (OSF).
Key Insights and Ideas
- Invariance is Key: Humans represent causal relationships as "modules" (like Lego blocks) that can be reused. Because these representations are invariant, they inherently contain information about a "default state"—what a node does when nothing is connected to it.
- The "Unplugging" Test: We judge something as a "productive cause" if the outcome depends on the presence of the causal link, not just the state of the variable.
- Beyond Symmetry: In Study 2, the authors showed that even when two nodes are mathematically symmetrical in an SCM (both must be in a specific state for the effect to occur), people still distinguish between them based on which one is "generative" versus "preventative."
- Causal Language as an Inference Tool: Study 3 found that if you tell someone "A allowed E," they can actually work backward to infer the underlying causal structure—specifically, they infer that A likely wouldn't be enough to cause E on its own if other variables were disconnected.
Key Results
- Study 1: Participants were significantly more likely to use the verb "CAUSE" when the actual state of a node differed from its "default state" (the state it takes when disconnected). This confirms that we track the consequences of disconnection, not just state changes.
- Study 2: In a complex three-node machine, participants overwhelmingly labeled a generative node as "causing" the outcome (modal answer: Cause) and a preventative node as "making no difference" or "allowing" it, despite both being equally necessary for the outcome in that specific scenario.
- Study 3: When participants were given causal descriptions ("A causes E" vs "B allows E"), they successfully predicted the behavior of "reduced" machines (where one node was removed). They correctly inferred that a "causing" node is more likely to be sufficient to produce an effect on its own.
Limitations and Future Work
The authors acknowledge several areas for expansion:
- Temporal Dynamics: The current model ignores when things happen. In the real world, the timing of events (e.g., a pillar already holding up a roof) heavily influences whether we see a preventer as a productive cause.
- Abstract Stimuli: The "machine" world is highly controlled. Future research needs to determine if these "disconnection" counterfactuals are as easy to compute in messy, real-world physical or social scenarios.
- Formal Refinement: The definition of productive causation provided is "preliminary and relatively informal." A more rigorous mathematical framework that integrates with existing probabilistic programs is a logical next step.
Overall, the paper argues that by simply acknowledging that our mental models are modular and compositional, we can save the counterfactual theory of causation from its most famous paradoxes.