🤖 machine learning

"Cause" is Mechanistic Narrative within Scientific Domains: An Ordinary Language Philosophical Critique of "Causal Machine Learning"

This paper critiques the premise of "causal machine learning" by applying ordinary language philosophy to argue that true causality is fundamentally mechanistic, asserting that definitive causal claims are only valid in domains with complete mechanistic models (like physics) and require agglomerated evidence across disciplines for complex systems (like biology and social sciences), thereby urging greater caution against overstating causal certainty in scientific research.

Original authors: Vyacheslav Kungurtsev, Leonardo Christov Moore, Gustav Sir, Martin Krutsky

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Vyacheslav Kungurtsev, Leonardo Christov Moore, Gustav Sir, Martin Krutsky

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Detective Game: Why "Cause" Isn't Just One Thing

Imagine you are a detective trying to solve a mystery. In the world of science and computers, there is a huge trend right now called "Causal Machine Learning." It's like giving a super-smart computer a magnifying glass and asking it to find the real reason why things happen, rather than just noticing that two things often happen at the same time. For example, just because ice cream sales and shark attacks both go up in July doesn't mean ice cream causes shark attacks; they are both caused by the summer heat. Finding that "real reason" is the holy grail of science because it helps us fix problems, cure diseases, and build better technology.

But here is the tricky part: what does "cause" actually mean? In a physics lab, "cause" is like a billiard ball hitting another; it's a precise, mathematical push. In a hospital, "cause" might be a complex chain of events involving your genes, your diet, and your environment. In a city, "cause" could be a story about how a new law changed people's behavior. The paper we are about to explore asks a big question: Can we use the same computer tools to find causes in all these different worlds? The authors, a team of researchers, argue that the answer is a loud "no." They suggest that trying to force a single computer formula to explain everything is like trying to use a hammer to fix a watch, a paint a picture, and bake a cake all at once. It might work for the hammer, but it will break the watch and ruin the cake.

The Paper's Big Idea: One Size Does Not Fit All

This paper is a philosophical detective story that investigates the word "cause." The authors use a method called "Ordinary Language Philosophy," which is basically a fancy way of saying, "Let's look at how real people actually use words in real life." They argue that the word "cause" changes its meaning depending on which "tribe" of scientists you are talking to.

The Physics Tribe: The Perfect Machine
In physics and engineering, the world is like a giant, perfectly built clock. If you know the gears (the equations) and you push one (the cause), you can calculate exactly where the other gears will go (the effect). Here, "cause" is a mathematical force. If you drop a rock, gravity pulls it down. The math describes the whole story perfectly. In this world, computer models that look for patterns in data can actually find the "cause" because the system is closed and the rules are known.

The Biology Tribe: The Messy Jungle
Now, jump to biology and medicine. The world here is more like a wild jungle. It's an "open system," meaning things are constantly interacting with the outside world in ways we can't fully measure. A human body is so complex that you can't write a single math equation that explains every cell, every chemical, and every thought. The authors say that in biology, you can't just look at a graph and say, "A caused B." Instead, you have to build a "narrative." You need to combine statistical clues (like data from a clinical trial) with a story about how it works (like how a drug clogs arteries). You need evidence from many different angles to be sure.

The Social Science Tribe: The Storytellers
Finally, there are the social sciences (like psychology, economics, and sociology). This is where things get really tricky. Humans are messy, unpredictable, and full of stories. The authors argue that in these fields, "cause" is often a story that helps us understand ourselves. For example, a therapist might help a patient by telling a story about how their past shaped their present. This story isn't a math equation, but it is "true" in a way that helps the person feel better and act differently. The authors point out that social science is currently struggling because it tries to act like physics, using strict math to prove things that are actually about human meaning and interpretation.

The Problem with "Causal Machine Learning"

The paper takes a hard look at the new trend of "Causal Machine Learning." These are computer programs that try to draw a map (a graph) showing how variables are connected. They use rules like "if A happens before B, and they are linked, then A causes B."

The authors say this approach is dangerous because it assumes the world is a closed, perfect machine (like physics) when it is often a messy, open system (like biology or society). They point out that these computer models often fail in the real world because they can't account for all the hidden factors (like your mood, your history, or the weather) that influence the outcome.

The paper explicitly argues against the idea that a computer can look at a dataset and definitively say, "This is the cause." They say that unless the computer is working in a closed physics lab where the rules are known, these models are just guessing. They are "overconfident." The authors warn that if we trust these computer maps too much, we might make bad decisions in medicine or policy because the computer missed the messy, human parts of the story.

The Solution: A Team of Detectives

So, what is the answer? The authors suggest we need to be more humble. Instead of relying on one computer model to tell us the truth, we need to gather evidence from everywhere.

Imagine you are trying to figure out why a forest is dying.

The Physicist looks at the soil chemistry equations.
The Biologist looks at the insects and the trees.
The Sociologist looks at how the local people use the land.
The Storyteller listens to the elders about what the forest meant to them.

The paper suggests that a "cause" is only real when all these different perspectives agree. This is called a "preponderance of evidence." It's not about one perfect math proof; it's about a chorus of different voices telling the same story.

For the future, the authors suggest that we should stop trying to force computers to be the ultimate judges of cause and effect. Instead, we should use computers as tools to help us see patterns, but we must always check those patterns against real-world stories and expert knowledge. They propose a new way of thinking called "Hermeneutics," which is the art of interpretation. It means understanding that sometimes, the "truth" isn't a number; it's a story that makes sense to people.

The Takeaway

The main finding of this paper is that "cause" is not a single thing. It changes shape depending on whether you are studying a falling rock, a beating heart, or a human mind. The authors suggest that the current excitement about "Causal AI" is a bit of a trap. It tries to use a simple, rigid tool to solve complex, messy problems.

They don't say we should stop using computers. Instead, they say we should use them with caution. We need to respect the differences between the sciences. In physics, math is king. In biology, we need a mix of math and mechanism. In social science, we need math, stories, and human understanding all working together.

The paper concludes with a call for "Epistemic Virtue," which is a fancy way of saying "intellectual honesty." It asks scientists and AI developers to admit when they don't know something for sure. Instead of shouting, "We found the cause!", they should say, "Here is a pattern that suggests a possibility, but let's look at the whole story before we decide." It's a reminder that in a complex world, the most powerful tool we have isn't a super-computer; it's the ability to listen to different kinds of stories and weave them together into a truth that makes sense.

Technical Summary: "Cause" is Mechanistic Narrative within Scientific Domains

Problem Statement

The paper addresses the growing trend in statistics and machine learning (ML) to develop "Causal Learning" and "Causal Inference" techniques that claim to discover "true" causality from data, often relying on independence criteria and probabilistic graphical models (PGMs). The authors argue that this field suffers from a fundamental epistemological flaw: it attempts to apply a singular, axiomatic definition of causality (derived largely from physics and formal logic) to scientific domains where the semantics of "cause" are distinct, pluralistic, and deeply tied to domain-specific mechanisms.

The core problem is the "bewitchment of intelligence by language," where computational techniques (such as Directed Acyclic Graphs (DAGs) and conditional independence tests) are mistaken for definitive certification of causal phenomena. The paper posits that this overconfidence ignores the reality that:

Causal Sufficiency (the assumption that all common causes are observed) rarely holds in real-world open systems (biology, social sciences).
Inductive Bias is inherent in all model selection, yet causal learning often presents its specific choice of statistical representation (e.g., Markovian transitions) as a neutral, practice-independent certification of truth.
Domain Disparity: The mechanisms defining causality in closed physical systems differ fundamentally from those in open biological systems or social systems, where emergence, irreducibility, and narrative interpretation play central roles.

Methodology

The authors employ a methodology rooted in Ordinary Language Philosophy (inspired by Wittgenstein) and Anthropological Semiotics. Rather than proposing new algorithms or statistical tests, the paper conducts a philosophical and linguistic investigation into how the concept of "cause" is actually used across different scientific "language games" (tribes).

The approach involves:

Reviewing Formal Definitions: Analyzing contemporary causal ML frameworks, including Bayesian Networks (BNs), Dynamic Bayesian Networks (DBNs), Structural Causal Models (SCMs), and Causal Probabilistic Logic (CP-logic).
Epistemological Analysis: Examining the history of causality in philosophy (Descartes, Hume, Kant, Popper, Kuhn, Quine) to understand the role of inductive bias, paradigms, and the limits of falsifiability.
Cross-Domain Semiotic Investigation: Comparing the "language games" of causality in:
- Physics/Engineering: Where closed systems and differential equations provide a one-to-one correspondence between mechanism and mathematical representation.
- Biology/Medicine: Where open systems, emergence, and multi-scale interactions require a "preponderance of evidence" from multiple layers of abstraction (e.g., combining statistical trials with mechanistic narratives).
- Social Sciences: Where human behavior is an irreducible open system, and causality is often established through hermeneutic narratives, phenomenology, and the accumulation of consistent findings across diverse methodologies rather than definitive quantitative prediction.
Integration of Neuro-Symbolic Perspectives: Considering how neuro-symbolic AI and mechanistic interpretability might bridge statistical models with domain narratives, while acknowledging current limitations.

Key Contributions

1. The Pluralistic Definition of Causality

The paper argues that "cause" is not a monolithic concept with a single essential definition. Instead, it is a functional term within specific scientific language games:

In Physics, cause is defined by field equations and differential operators (e.g., $\frac{dX}{dt} = f(X)$ ), where the mechanism is fully captured by the equation.
In Biology/Medicine, cause is a "mechanistic narrative" supported by a convergence of evidence across scales (molecular, physiological, clinical). No single statistical model can capture the full causal mechanism of an open system.
In Social Sciences, cause is often a "Hermeneutic truth"—a compelling narrative that explains human behavior and meaning, which may be instrumentally useful for individuals or policy even without the predictive precision of physical laws.

2. Critique of "Causal Machine Learning"

The authors critique the premise that statistical independence criteria (e.g., $X \perp Y | Z$ ) can certify causality. They highlight:

The Failure of Causal Sufficiency: Real-world systems are rarely closed; latent confounders are ubiquitous, violating the core axioms of standard causal discovery.
The Misalignment of Abstraction: Applying Markovian DAGs to complex, emergent systems (like human psychology) is an inappropriate choice of abstraction that redefines "cause" to fit the tool rather than the phenomenon.
Overconfidence: The field suffers from an incentive structure that rewards definitive causal claims, leading to the overstatement of results that are merely statistical associations.

3. The Role of Hermeneutics and Narrative

The paper elevates Hermeneutics (the theory of interpretation) as a valid and necessary epistemological framework for social sciences and complex biological systems. It argues that:

Narratives are not merely "soft" explanations but are essential tools for processing meaning and establishing "instrumental truth" in domains where precise mechanistic modeling is impossible.
The "Hard Problem" of consciousness and the debate over free will are often linguistic confusions that can be resolved by recognizing that different language games (neuroscience vs. phenomenology) offer complementary, non-reductive perspectives.

4. Epistemic Virtue and the "Preponderance of Evidence"

The paper proposes a shift in scientific practice toward Epistemic Virtue. Instead of seeking definitive causal certification from a single statistical model, researchers should:

Exercise caution in communication, avoiding definitive causal claims when evidence is limited to statistical association.
Seek a preponderance of evidence by aggregating consistent findings across multiple domains, scales, and methodologies (e.g., combining statistical data, evolutionary theory, and phenomenological narratives).
Treat causal learning tools as methods for generating "interpretable representations of object relationships" rather than as definitive proofs of causation.

Results and Findings

Physics vs. Other Domains: The paper confirms that causal learning techniques (like SCMs) are most appropriate for physics and engineering, where systems are closed and mechanisms are fully describable by differential equations.
Inappropriateness for Open Systems: For biology and social sciences, the assumption of causal sufficiency is practically impossible to satisfy. Consequently, causal discovery algorithms often fail to identify valid graphs in real data, or they produce spurious results due to unmeasured confounders.
The Necessity of Multi-Domain Harmony: Valid causal claims in complex systems (e.g., the health effects of smoking) are not established by a single randomized trial or statistical model, but by the convergence of diverse evidence types: statistical associations, mechanistic biological narratives, and clinical observations.
Hermeneutic Utility: In social sciences, the "truth" of a causal claim often lies in its ability to provide a coherent, instrumentally useful narrative for human understanding (e.g., psychotherapy modalities), rather than in its ability to predict individual outcomes with high precision.

Significance and Claims

The paper claims that the current trajectory of "Causal Machine Learning" is fundamentally flawed because it attempts to reduce the rich, pluralistic concept of causality to a set of statistical independence tests.

Significance:

Correcting the "Bewitchment": It aims to clear up the confusion caused by the misuse of language, urging the scientific community to recognize that "cause" means different things in different contexts.
Promoting Intellectual Humility: It calls for a more modest approach to scientific claims, particularly in fields where systems are open and irreducible. It argues that the solution to the replication crisis in social sciences is not better statistics, but a greater willingness to accept uncertainty and rely on multi-disciplinary consensus.
Redefining Explainability: It suggests that true explainability in AI requires integrating statistical models with domain-specific narratives and mechanistic knowledge (Neuro-Symbolic approaches), rather than relying solely on feature importance or counterfactuals.
Methodological Guidance: It proposes that the path forward for interdisciplinary science is not the unification of methods under a single causal framework, but the harmonization of diverse "language games" to build a richer, more robust understanding of complex phenomena.

The authors conclude that while causal inference tools are powerful statistical methods, they must be stripped of their claim to definitive causal certification. Instead, they should be viewed as tools for generating hypotheses and identifying associations that must be validated through the rigorous, multi-layered, and often narrative-driven processes of domain-specific scientific inquiry.