What do we mean when we say one thing “causes” another?
Published:
The language of causality has become increasingly prominent in statistics and machine learning. Methods are often presented not merely as tools for prediction or intervention, but as ways of recovering cause and effect from data. Our paper, “Cause” is Mechanistic Narrative within Scientific Domains: An Ordinary Language Philosophical Critique of “Causal Machine Learning,” asks a more basic question: what are scientists actually saying when they call something a cause?
The paper takes an ordinary-language approach. Rather than beginning by stipulating a universal mathematical definition of causality, we examine how causal language functions within scientific practice. Our central claim is that the grammar of causal explanation changes substantially from one scientific domain to another, while retaining a common role: a causal account supplies a mechanistic narrative about the forces or processes of influence that matter within that domain.
Causality is not the same object in every science
In physics and engineering, the relevant mechanisms can often be represented sufficiently completely by mathematical models. A differential equation, conservation law, constitutive model, or engineered system description can therefore carry much of the causal content itself.
Biology introduces a different difficulty: emergence and organization across scales. Molecular interactions, cellular processes, organisms, and populations can each support meaningful mechanistic descriptions, but a causal explanation may require showing how mechanisms remain coherent as one moves between those scales.
The social sciences present a still harder case. Human systems may resist models with the predictive precision expected in physics, while interpretation, context, and what the paper discusses through hermeneutics can nevertheless produce explanations that are instrumentally useful. Calling all three situations “causal inference” does not make their epistemic structures identical.
The criticism of causal machine learning
The target is therefore not mathematical causal modeling as such. Mathematical models can be extraordinarily useful, and in some domains they can give remarkably complete descriptions of causal structure. The concern is the stronger move from a formalism that answers a precisely stated inferential question to the claim that the formalism has thereby captured causation in a domain-independent sense.
A statistical distinction, graph, intervention operator, or counterfactual construction does not acquire scientific meaning in isolation. Its interpretation depends on the mechanisms, assumptions, measurements, and background knowledge of the domain in which it is being used. Formal machinery should sharpen that scientific reasoning rather than silently substitute a mathematical vocabulary for it.
Evidence across domains
This leads to a more pluralistic picture of scientific explanation. For sufficiently rich phenomena, strong causal claims may emerge from an agglomeration of mutually consistent evidence across domains and scales, rather than from a single privileged formal language. That creates difficult but productive questions about how different scientific “language games” can be reconciled and how explanations survive transitions between scales.
The practical implication is epistemic caution. Causal-learning methods can provide valuable evidence, but the strength of the scientific claim should track what the evidence and domain assumptions actually warrant. Overstating what a formal method establishes risks contributing to the broader credibility problems produced when scientific precision is communicated with greater certainty than the underlying evidence supports.
Why I think this matters
This argument connects to a broader theme in my work: mathematics is most powerful when it is explicit about the scientific or engineering structure it represents. The same principle appears in scientific machine learning, optimization, and control. The goal is not to weaken mathematical formalization, but to make the relationship between formal models, domain knowledge, data, and the claims we draw from them more precise.