Actual Causation in Overdetermined Systems

44 concepts26 papers1 proposal

Imagine a committee of 11 members voting on a proposal. The proposal requires a simple majority of 6 votes to pass. When the votes are counted, the result is a unanimous 11–0 landslide.

Now, consider a standard counterfactual question: Did committee member Alice cause the proposal to pass?

Under the classic "but-for" definition of causality—which states that an event $A$ caused $B$ if, but for $A$, $B$ would not have occurred—the answer is a resounding no. If Alice had changed her vote to "no," the proposal still would have passed 10–1. In fact, applying this logic to every member yields a paradoxical conclusion: nobody caused the proposal to pass.

This puzzle is the core challenge of actual causation (or token causality) in systems characterized by redundant causation and overdetermination. When multiple independent factors are sufficient to produce the same outcome, standard counterfactual analysis breaks down, obscuring individual accountability and system mechanics.

The Anatomy of Redundancy: Overdetermination vs. Preemption

To understand why standard counterfactuals fail, we must distinguish between the two primary flavors of redundant causality: symmetric overdetermination and asymmetric preemption.

Symmetric Overdetermination

The 11–0 landslide is a classic example of symmetric overdetermination. Multiple causes operate simultaneously, and each is fully sufficient on its own to bring about the effect. As noted by Chockler et al. and Halpern et al., this structural redundancy creates a logical blind spot where individual contributions disappear because the outcome is already "guaranteed" by the remaining factors. Traditional counterfactual models struggle to assign responsibility here because no single agent acts as a strict bottleneck for the outcome.

Causal Preemption

Preemption introduces a temporal or structural asymmetry. Imagine Assassin A shoots and kills a target instantly. Meanwhile, Assassin B is waiting in reserve, prepared to shoot if Assassin A misses. If Assassin A hadn't fired, the target would still be dead because of Assassin B.

A naive "but-for" test concludes that Assassin A did not cause the death. Yet, intuitively, Assassin A is the actual cause, while Assassin B is merely a backup. As Künnemann et al. and Park argue, standard causal models fail here because they look only at the final outcome state, ignoring the physical control flow and the fact that the primary cause actively blocked or preempted the backup mechanism from executing.

Restoring Intuition via Contingencies and Sufficiency

To align formal logic with human intuition, researchers have developed frameworks that move beyond simple counterfactual dependence. The two most prominent approaches rely on contingencies within Structural Causal Models (SCMs) and the NESS (Necessary Element of a Sufficient Set) principle.

The Contingency Approach (Halpern-Pearl)

How do we make Alice's vote matter in the 11–0 landslide? We do it by changing the context. If we imagine a hypothetical scenario where 5 other members had voted "no," then Alice’s vote suddenly becomes the decisive tie-breaker.

This is the core mechanism behind the influential Halpern-Pearl (HP) definition of actual causality. To determine if an event is a cause, we are permitted to alter the background environment by holding certain variables fixed to "contingent" values. Halpern et al. refined this framework by restricting these contingencies to values that occurred in the actual world or represent "normal" defaults. By filtering out far-fetched scenarios, Halpern et al. successfully isolated actual causes without introducing arbitrary logical paradoxes.

Actual World (11-0 Landslide)
[Vote 1] [Vote 2] ... [Vote 11] ---> Motion Passes (No single "but-for" cause)

Contingent World (Freeze 5 votes to 'No')
[Vote 1-5 fixed to No] + [Vote 6-10] ---> Alice's Vote is now a Pivot!

The NESS and CNESS Frameworks

An alternative approach bypasses structural adjustments entirely by focusing on sufficiency. The NESS test states that a factor is a cause if it is a Necessary Element of a Sufficient Set of conditions. In the landslide example, any group of 6 "yes" votes constitutes a sufficient set to pass the motion. Alice's vote is a necessary element of any such set that includes her. Therefore, she is a cause.

To unify this regularity-based intuition with modern structural equations, Beckers introduced the Counterfactual NESS (CNESS) definition in The Counterfactual NESS Definition of Causation. Instead of evaluating the global outcome, CNESS analyzes specific causal subpaths. An event is an actual cause if there is no counterfactual scenario where the effect is produced through that exact same physical chain of execution. This prevents backup causes from washing out the responsibility of the primary actor, providing a robust solution to both early and late preemption paradoxes. This framework has been systematically extended by Beckers et al., Beckers, and Bochman to handle non-deterministic environments and complex conjunctive effects.

Why It Matters: Responsibility, Harm, and System Safety

Resolving these philosophical paradoxes is not just an exercise in formal logic; it is a practical necessity for designing safe, accountable, and ethical autonomous systems.

Graded Responsibility and Blame

In complex systems, binary causality ("did it cause it or not?") is often insufficient. Chockler et al. leveraged the contingency framework to quantify degrees of responsibility. An agent's responsibility is defined as $\frac{1}{1 + N}$, where $N$ is the minimum number of changes required to make that agent a decisive pivot. In a 6–5 vote, $N=0$, so responsibility is $1$. In an 11–0 landslide, $N=5$, reducing individual responsibility to $\frac{1}{6}$. This mathematical distinction allows systems to differentiate between a critical failure and a diffuse, highly redundant event.

Defining Harm and Moral Accountability

When an autonomous agent acts, how do we determine if it caused harm? If a self-driving car fails to brake, but a secondary system would have failed a millisecond later anyway, a naive counterfactual model might absolve the car of causing harm because the crash was "inevitable."

To fix this, Beckers et al. and Beckers et al. combined structural causal models with default utilities to formalize a qualitative definition of harm. By isolating the actual causal path, their framework ensures that an agent is held responsible for a negative outcome even if a backup cause stood ready to produce the same result. Building on this, Beckers integrated CNESS with epistemic conditions to formalize moral responsibility for AI systems, preventing algorithms from escaping accountability in overdetermined environments.

System Diagnostics and Risk Assessment

In engineering and software verification, redundant causality is ubiquitous:

  • Autonomous Driving: As demonstrated by Li et al., if a vehicle stops at an intersection where both a pedestrian is crossing and the traffic light is red, removing either factor leaves the vehicle's behavior unchanged. Advanced causal inference frameworks are required to disentangle these overlapping influences and correctly identify multiple risk objects.
  • Software Security: Coenen et al. extended actual causality to explain security hyperproperty violations. When multiple concurrent system inputs are capable of triggering an information leak, contingency-based frameworks allow automated verifiers to isolate the specific input sequence responsible for the exploit.
  • Temporal Planning: Even in early AI planning, as explored by Hogge, effective prevention of negative goals required addressing multiple independent causal chains, recognizing that terminating a single antecedent is insufficient when redundant pathways exist.

By moving from simple "but-for" questions to sophisticated models of structural sufficiency and path-based necessity, computer science has built the tools necessary to navigate a world where outcomes are rarely driven by a single, isolated cause.

Key Foundational Work

Causes and Explanations: A Structural-Model Approach (2000) — contributed 3 concept elements to this direction.
Prevention Techniques for a Temporal Planner (1988) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_overdetermination_60ffee6c,
  title = {causal overdetermination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-overdetermination/rdr_60ffee6c73074029b45b53dd786ba573},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_overdetermination_60ffee6c_v1,
  title = {causal overdetermination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-overdetermination/rdr_60ffee6c73074029b45b53dd786ba573/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake