Agents should not make intervention claims from predictive confidence

An agent should not recommend an intervention merely because its prediction is confident. Predictive confidence says the model expects an outcome under the distribution it has observed. An intervention claim says changing one variable will change that outcome. Those are different questions, and the gap between them is where automated recommendations become confidently harmful.

Pearl’s ladder gives the clean separation. Association asks what is likely when something is observed: P(Y|X). Intervention asks what happens when the system sets X: P(Y|do(X)). Counterfactual reasoning asks what would have happened to this case under a different action. Most machine-learning scores live on the first rung. An agent that sees customers who received a retention call cancel less often cannot infer that making more calls will reduce cancellation. The call may have been offered only to customers already judged likely to stay, or another variable may drive both selection and outcome.

Structural causal models make the missing commitments explicit. Variables, directed mechanisms, and unobserved causes define which paths carry causal influence. A backdoor adjustment is valid only if the chosen variables block confounding without conditioning on descendants of the intervention. Conditioning on a collider can create an association that was not there. Front-door criteria and instrumental variables help only under their own assumptions. The point is not that every agent needs to execute do-calculus. It is that every consequential recommendation needs to name why the effect is identifiable rather than borrowing certainty from a predictor.

I would encode that distinction in the action contract. A response schema should label a claim as descriptive, predictive, interventional, or counterfactual; attach the population, outcome, time horizon, evidence source, adjustment set, and assumptions; and refuse to promote a prediction into an action when those fields are absent. The runtime can then route high-impact intervention claims to an approved causal model, randomized evidence, or human review. Language quality cannot fill a missing identification argument.

This also changes evaluation. Accuracy and calibration are necessary for predictive claims, but an intervention system needs tests for confounding, transport across environments, positivity, sensitivity to hidden variables, and whether the chosen action produced the intended outcome. The eval contract should score the claim type itself: did the agent state an association honestly, or did it narrate it as a cause? Human feedback must preserve disagreements about assumptions rather than averaging them into one apparently objective recommendation.

Counterfactual explanations deserve equal caution. Saying “your application would have been approved if income were higher” implies a model of how income changes while other variables respond. Holding every correlated feature fixed may describe a classifier boundary, not a feasible world. The formal abduction-action-prediction procedure exists because individual counterfactuals require more structure than feature perturbation. A fluent alternative is not automatically an actionable one.

There is one precise concession: when a randomized experiment directly matches the target population, treatment, outcome, and operating conditions, an agent can often relay the estimated effect without rebuilding a full causal graph. Even then it must preserve uncertainty, eligibility, and external-validity limits. The concession is evidence-specific, not a license to treat every high-confidence model score as experimental proof.

I want agents to be decisive about what the evidence permits and restrained about what it does not. Unsupported factual claims are systems defects; unsupported causal claims are the same defect with an action attached. And causal attribution matters after execution for the same reason causal identification matters before it: observation alone cannot tell us what produced the outcome. Predict well when prediction is the job. Before changing the world, require a causal claim the system can defend.