This paper studies how a communicator can persuade someone about a causal link. The sender can choose which real variables to reveal, give the true joint distribution of those variables, and propose a causal story that links them. The receiver accepts the story only if the disclosed data and the proposed model make the causal link unambiguous.

The authors build a formal model in which the true causal structure is represented by a directed acyclic graph (DAG). A DAG is a simple diagram that shows which variables directly cause others. The sender is assumed to observe the true DAG and the full data. The receiver may start with no prior model or with a prior subjective model. The paper analyzes two types of receivers: a naive one who accepts any model consistent with the data, and a sophisticated one who accepts a model only when the data conclusively identifies the causal link.

Their main finding is an asymmetry. When the true causal link exists, the sender often needs to disclose only one or two well-chosen variables to make that link provable to a sophisticated receiver. But to convince a receiver that a perceived causal link does not exist, the sender must reveal every common cause (also called a confounder) that could explain the correlation. Because there can be many confounders, disproving a causal link can be much more demanding than establishing one. The paper shows formally that persuading a sophisticated receiver that causality runs the opposite way from the truth is impossible (Theorem 1), and that debunking a false model can sometimes be done by revealing at most two appropriate variables but may require arbitrarily large disclosures otherwise (Theorem 2).

The paper gives a concrete example about Master of Business Administration (MBA) degrees and earnings. Suppose education and earnings are correlated because of ability and social skills, not because the MBA causes higher pay. A business school could run a campaign called “Experience+MBA=More Pay!” and disclose a variable such as work experience. By revealing how experience links to education and pay, the school can make the data look like it supports a causal effect of the MBA. This illustrates how selective disclosure of real variables and a proposed causal model can persuade, even though no false data are produced.