Accountable Action End Points for Evaluating Artificial Intelligence in Digital Public Health.

Unlabelled: AI is increasingly embedded in digital public health systems, but model performance, system deployment, message delivery, alert volume, and user engagement do not establish whether an AI-supported workflow has changed public health practice. This viewpoint proposes a public health action end point-a prespecified, auditable, AI output-specific, actor-bound, time-bound, denominator-based way to specify and measure an existing proximal process or implementation end point. It begins with a defined AI output and records the corresponding action opportunity, accountable actor, action status, and evidence concerning the AI's role in the action. A complete specification includes rules for repeated outputs, completed action, justified nonaction, missed action, unresolved cases, denominator loss, and paired safety, workload, privacy, equity, and model drift monitoring. A public health action end point is not a new causal outcome, reporting guideline, or validated surrogate for a population health benefit. An observed action rate alone does not establish that AI initiated or caused the action; causal claims require a documented attribution strategy and an appropriate comparator or causal design. The framework applies to research and production deployments but complements rather than replaces ethics review, trial registration, and jurisdiction-specific regulatory and governance obligations. It is intended to help authors, implementers, reviewers, and editors align claims with what an evaluation has actually measured.