AI Causal Graphs: 4 Checks Before You Act on One
On this page
An AI model can read a short list of variable names, say TV spend, brand awareness, website visits, sales, and return a full cause-and-effect graph without seeing a single row of your data. The output looks decisive: arrows, direction, a written explanation for every edge. It is still a hypothesis. Before it moves a budget or a strategy, it needs to clear four checks.
What a proposed causal graph actually claims
A causal graph represents each variable as a node and each cause as a directed arrow: TV spend → brand awareness, brand awareness → website visits, website visits → sales. Each arrow makes a specific, testable claim. Change the variable at the tail and the variable at the head may change. In an acyclic model with no reverse path, intervening on the head does not cause the tail through that arrow. Real systems can contain feedback over time, so define the time scale before drawing the graph.
An arrow gives direction of dependence only. It does not say whether the effect is positive or negative, how large it is, or whether it is the same across segments and spend levels. Those come from a model and evidence.
How an AI model builds that graph
Given variable names, short descriptions, and a working definition of causality, a large language model can propose a plausible graph from general knowledge. A survey of the field by Wan et al., "Large Language Models for Causal Discovery: Current Landscape and Future Directions," arXiv 2402.11068v2, revised February 2025 reviews how such models extract causal relationships from text and bring domain knowledge into statistical methods, and it discusses their limits. A model working from names alone reasons from general market knowledge, not your numbers. That makes it a low-effort way to generate a first hypothesis. It is not evidence that the pathway holds in your market.
Four checks before an AI-proposed causal graph drives a decision
1. Does the direction match how the intervention would actually run?
An edge like "brand awareness → website visits" implies that changing awareness may change visits, not the reverse. Whether the change raises or lowers visits is a separate hypothesis to state and test. Ask whether your team could run the change the arrow implies. If the intervention that would move the tail variable does not exist, or moves several variables at once, the direction is a guess dressed as a finding.
2. How sensitive is the result to the graph's assumptions?
A graph is a set of edges that were accepted and a set that were deliberately left out. A single doubtful edge, such as routing all of TV's effect through awareness and none directly to visits, changes what the whole graph implies about where a dollar should go. Check the edges the model chose not to draw, not only the ones it kept.
3. Would a domain expert draw the same arrows from the same variable names?
Because the model reasons from variable descriptions and general world knowledge, two plausible-sounding graphs can disagree on direction for the same variables. Recent work treats this as a search problem. Zanna and Sano propose and score edges with active learning, and they test the approach on a semi-synthetic benchmark built from the UCI Adult dataset with embedded fairness pathways (Zanna and Sano, "Uncovering Bias Paths with LLM-guided Causal Discovery," arXiv 2506.12227v1, June 2025). That benchmark is a fairness-path test, not a validation on a marketing graph. Treat disagreement between the model and a domain expert as a signal to test the edge, not a tiebreaker to resolve by re-prompting.
4. What evidence identifies the effect?
A graph that reads well is not the same as a graph that has been checked against how people respond to the intervention it describes. A randomized intervention is one way to test an effect. Observational identification can also be justified, for example with a credible instrument, discontinuity or adjustment set under stated assumptions. A total-effect test does not automatically verify each mediating edge in the TV-to-sales pathway.
From hypothesis to a tested effect
The fourth check is the one a proposed graph cannot answer on its own. The test has to match the edge. Take two edges from the example graph:
| Edge | A test that fits | What the result supports |
|---|---|---|
| Message variant → stated choice among a defined audience | A randomized discrete choice experiment comparing the variants on a stated choice, with an interval | An estimated effect within the simulated or surveyed population, subject to the usual limit that stated choice is not a live purchase |
| TV spend → website visits | A live geographic holdout or media lift test that measures visits | Visits attributable to the spend change in the tested markets |
Subconscious runs controlled experiments on simulated populations to estimate the effect of a defined intervention on a stated-choice endpoint. That stated-choice endpoint does not measure website visits from TV spend; a media-budget conclusion needs matched visit evidence. For human transfer, specify a matched human comparison and confirm its availability. See how a study moves from hypothesis to replicated evidence and the kinds of decisions it has tested in past studies, and the design in how a study is run.
Where this still needs care
An AI-proposed graph is a hypothesis-generation step, not a validated causal effect. Multi-agent setups can widen the hypothesis space. Le et al. combine statistical methods with LLM reasoning and an adversarial debate step to refine candidate graphs (Le et al., "Multi-Agent Causal Discovery Using Large Language Models," arXiv 2407.15073v4, revised May 2026). Even so, a wider or better-ranked set of plausible graphs is still a set of hypotheses until data and a justified identification strategy support the relevant effect. Treat every edge the same way regardless of how it was generated: as a claim to test before it changes what your team does next.
Next step
Before reallocating budget on the strength of a proposed causal graph, pick the one edge the decision depends on, decide which test fits that edge, and run it. See the published evidence and its limits on the research page.
