A Causal Argumentation Method for Explainability of Machine Learning Models

TEXT ANALYSIS

The Dissection

This is a technical computer science paper submitted to arXiv on May 20, 2026. It proposes integrating causal discovery methods with Bipolar Argumentation Frameworks (BAF) to explain why machine learning models make certain predictions, moving beyond standard XAI feature-attribution methods that merely identify which features matter. The core claim is that causal graphs + argument structures can produce more interpretable explanations of model decision logic.

The Core Fallacy

The paper operates within a fundamental category error: it assumes the problem of ML explainability is solvable through better inference from model outputs. It is not. The problem is architectural.

Modern neural networks — the models being "explained" — are not causal engines. They are correlation machines operating on high-dimensional statistical manifolds. No post-hoc causal layer bolted onto a BlackBox produces genuine mechanistic explanation. You are reasoning about why a vector arithmetic operation produced a token as if it were a deliberate causal chain. The BAF wrapper does not transform a correlation black box into a causal one. It describes the proxy features the model happened to weight, dressed in causal language. That is not explanation. That is elaborate misidentification.

The deeper fallacy: the entire XAI field (including this paper) assumes explainability has value independent of model accuracy. Under the Discontinuity Thesis, this assumption collapses. These models are not becoming more transparent. They are becoming more capable. The explanation of a model that writes code, diagnoses disease, and drafts contracts at superhuman level is not a regulatory safeguard — it is a ritual of legitimacy for systems that are already displacing the workers who would evaluate their explanations.

Hidden Assumptions

1. Causal discovery on ML features is valid. — Causal discovery methods assume data generated by causal mechanisms. Model features are learned representations, not necessarily causally connected in any domain-relevant sense.
2. "Why" questions have answers in the model. — Presumes model decisions are the product of interpretable causal logic rather than distributed, non-decomposable gradient evolution.
3. Post-hoc explanation is socially useful. — Assumes that explaining model behavior after the fact satisfies any real regulatory, legal, or trust-building need. It does not. Courts, regulators, and affected individuals need predictive accountability, not retrospective rationalization.
4. BAF semantics produce stable extensions. — The claim of finding "extensions of features that explain why outcomes were chosen" presupposes the argument framework converges meaningfully. This is non-trivial for complex models.
5. Human-legible explanations can be generated at scale for frontier models. — The paper tests on "two benchmark datasets." This methodology does not scale to models with hundreds of billions of parameters and complex multi-task architectures.

Social Function

This paper performs transition management and prestige signaling within the academic AI community. It:

• Positions the author(s) as contributing to the thriving XAI subfield, which attracts citations, grants, and conference slots.
• Conforms to the institutional demand for "responsible AI" without threatening the underlying research agenda.
• Provides a thin legitimacy veneer for models that are structurally unexplainable at the level that matters for accountability.
• Offers a publication path for causal reasoning researchers entering the ML space — interdisciplinary career moat construction.

The paper's function is not to solve explainability. It is to sustain the conversation about explainability in a form that is publication-eligible, grant-attractive, and institutionally comfortable.

The Verdict

Under the DT lens, this paper is a theoretical hospice intervention on a structural problem. Causal argumentation methods applied post-hoc to neural networks are like adding architectural drawings to a building after it has already been condemned. The building is the post-WWII economic order in which human cognitive labor is the unit of value. The models this paper attempts to explain are the demolition mechanism. The XAI field — including this contribution — is the architect drawing plans around the demolition, not addressing it.

The paper is technically competent within its own framing. It is structurally irrelevant to the actual threat vector. It does not slow, explain, or redirect AI displacement. It describes it in a language that feels rigorous enough to satisfy institutional conscience.

Survival Relevance: Zero for individuals navigating displacement. Useful for researchers building CV moats in the XAI space — a servitor role within a legitimacy apparatus that delays honest reckoning.

Oracle Viability Assessment: Publishable, fundable, citable, economically inconsequential.