From plausible menus to trustworthy nutrition: a risk-proportionate VALID-Diet framework for artificial intelligence-generated dietary advice.

Abstract

Generative artificial intelligence can produce menus, recipes, shopping lists, and dietary explanations that appear individualized and clinically informed. Fluency alone, however, provides little evidence of nutritional accuracy, safety, feasibility, or benefit. We propose VALID-Diet, a staged framework for evaluating systems that generate actionable dietary recommendations as sociotechnical interventions. Its domains are Verification of nutritional targets and composition; Adversarial safety and boundary testing; Linkage between user inputs and recommendation changes; Implementation feasibility and inclusion; and Documentation, versioning, and reproducibility. Stage 0 defines intended use, input integrity, and a structured risk profile based on harm severity, exposure, autonomy, reversibility, and enforceable oversight. Stage 1 separates target validity, menu achievement, and arithmetic agreement and, for integrated systems, tests retrieval, calculation, and generation as distinct components. Stage 2 evaluates safety, substantive personalization, feasibility, and inclusion using pre-specified case blueprints. Stage 3 evaluates human oversight, workflow utility, and prospective outcomes only after earlier gates are passed. A retrospective worked application to a published retrieval-augmented personalized-nutrition system illustrates risk assignment and progression decisions. VALID-Diet is not a regulatory standard or consensus guideline; it is a testable proposal intended to align research claims with evidentiary strength.