Causal machine learning promises to elevate artificial intelligence from mere pattern recognition to true reasoning. By modeling the world through Directed Acyclic Graphs (DAGs) and structural equations, models can compute counterfactuals ("what if we had acted differently?") and isolate invariant mechanisms from spurious correlations. However, this mathematical machinery relies on a foundational assumption: the qualitative causal graph provided to the system must be correct. When this structural assumption is violated—a phenomenon known as causal graph misspecification—the guarantees of causal inference quickly degrade.
Why Graph Errors Propagate: The Mechanics of Misspecification
To understand why causal models are so sensitive to graph structure, we must look at how causal effects are identified. In the framework formalized by Judea Pearl in The mathematics of causal inference, a graph is not just a visual aid; it is a strict mathematical statement about conditional independencies ($d$-separation).
When an algorithm calculates a causal effect or a counterfactual, it uses the graph to determine the correct "adjustment set"—which variables to condition on to block confounding pathways without accidentally blocking the causal mechanism itself or introducing selection bias via colliders. If a human expert or an automated discovery algorithm provides an incorrect graph, the mathematical equations derived from it become fundamentally unanchored from reality.
For instance, if a graph omits a latent confounder, the model will misinterpret a shared correlation as a direct causal link. Conversely, misidentifying a variable as a mediator instead of a collider can cause an algorithm to inadvertently inject bias where it intends to remove it. As Baohong Li et al. demonstrate in Two-Stage Shadow Inclusion Estimation: An IV Approach for Causal Inference under Latent Confounding and Collider Bias, correcting for complex real-world phenomena like selection and confounding bias is entirely contingent on having an accurate prior specification of the data-generating process.
The Downstream Collapse of Causal Applications
Because the graph dictates the mathematical formulas used for inference, structural misspecification ripples through every downstream application of causal AI.
Explainability and Recourse
Modern Explainable AI (XAI) frameworks use causal graphs to move beyond statistical feature importance. Frameworks like do-SHAP, analyzed by Tomas Garriga in Practical do-Shapley Explanations with Estimand-Agnostic Causal Inference, and CAGE by Nils Ole Breuer et al. in CAGE: Causality-Aware Shapley Value for Global Explanations, simulate physical interventions (using the do-operator) to attribute credit to features. Similarly, counterfactual explanation models like LEWIS (Romila Pradhan et al., Explaining Black-Box Algorithms Using Probabilistic Contrastive Counterfactuals) suggest actionable recourse for users. If the underlying DAG is wrong, these interventions simulate a fictional physics, generating explanations and recourse recommendations that are fundamentally unfaithful to the true system.
Algorithmic Fairness
Causal fairness methods attempt to eliminate bias by decomposing a model's predictions along specific paths from a sensitive attribute to the outcome. As Weishen Pan et al. show in Explaining Algorithmic Fairness Through Fairness-Aware Causal Path Decomposition, if the graph is incorrectly specified, the decomposition of model disparity into "fair" and "unfair" paths becomes misleading, potentially masking severe biases or introducing new ones.
Mathematical Bounds and Counterfactuals
When exact causal effects cannot be identified due to hidden variables, researchers calculate mathematical upper and lower bounds. In Unit Selection Based on Counterfactual Logic, Ang Li 0009 et al. derive bounds to identify optimal target individuals. However, these bounds are strictly valid only under the assumed structural model; a misspecified graph invalidates the constraints, rendering the derived bounds useless for decision-making.
The Asymmetry of Graph Errors: Omission vs. Commission
An intuitive and critical insight emerging from recent research is that not all graph errors are created equal. Causal architectures exhibit a distinct asymmetry in how they handle different types of structural misspecification.
Specifically, models are far more vulnerable to omitting true causal edges (under-specification) than they are to including extra, spurious edges (over-specification). This phenomenon is highlighted in the paper Neural Bayesian Network Understudy, which notes that missing edges severely degrade a model's ability to mimic causal properties.
Why does this asymmetry exist? When you omit an edge between two variables, you are asserting a strict, dogmatic claim: these two variables have zero direct causal relationship and are conditionally independent given their parents. This false independence constraint forces the model to ignore real dependencies, acting as active misinformation that corrupts the inductive bias (Shahine Bouabid et al., Returning The Favour: When Regression Benefits From Probabilistic Causal Knowledge).
On the other hand, adding a spurious edge merely allows for a relationship that happens to have a strength of zero. While an over-connected graph increases computational complexity and might slightly weaken statistical power, it rarely invalidates the underlying adjustment sets. As Ferdinand Rewicki observes in Leveraging Causal Information for Multivariate Timeseries Anomaly Detection, while missing true causal links causes significant failures in detecting subtle anomalies, the inclusion of spurious relationships is significantly less harmful.
Navigating Uncertainty in Real-World Systems
Given this extreme sensitivity, how can causal machine learning be safely deployed in high-stakes domains like biomedicine (Jie Xu 0012 et al., Protocol for the development of a reporting guideline for causal and counterfactual prediction models in biomedicine) or intensive care (JH Krijthe et al., The future of artificial intelligence in intensive care: moving from predictive to actionable AI)?
Researchers are actively developing methods to soften the rigid dependency on perfect graphs. One approach is to model the graph itself as a probabilistic object. In Causal Effect Identification in Uncertain Causal Networks, Sina Akbari et al. frame causal inference over networks where edges exist only with certain probabilities, algorithmically searching for the most plausible identifiable subgraphs. Others, like Tom Claassen in A Bayesian Approach to Constraint Based Causal Inference, integrate Bayesian reliability estimates directly into constraint-based discovery to quantify the confidence of inferred structures.
Ultimately, understanding structural sensitivity shifts the focus of causal ML from blindly trusting algorithmic outputs to rigorously validating and softening the structural priors that generate them.