FormulationGPT Is Not a Formulator: Governing Generative Artificial Intelligence across Excipient Selection, Process Design, and Product Decisions

Generative artificial intelligence can retrieve, reorganize, and synthesize formulation knowledge, but fluent output is not equivalent to the situated expertise required to design a pharmaceutical product. Formulation decisions depend on dosage-form purpose, material attributes, excipient functionality, compatibility, stability, process conditions, scale, bioavailability, patient needs, and the quality of supporting evidence. This article develops an original, non-empirical governance architecture for generative artificial intelligence used across excipient selection, process design, experimental planning, and product decisions. The architecture separates bounded generative assistance from actions requiring pharmaceutical evidence and accountable authorization. It integrates an intended-use gate, data-provenance and knowledge-grounding controls, formulation and compatibility constraints, process-design and scale-up reasoning, evidence-warrant classification, human review, experiment authorization, product-decision boundaries, and lifecycle monitoring. Uncertainty is treated as a decision-relevant property that must be traced to evidence gaps, model limitations, conflicting sources, or unresolved experimental questions rather than represented by an unqualified confidence statement. Human involvement is defined as an accountable review function with authority to reject, revise, escalate, or authorize proposed work, rather than nominal confirmation of model output. Monitoring and change control extend governance to changes in data, retrieval sources, prompts, models, materials, equipment, and development context. The original contribution is an evidence-to-authorization architecture designed to prevent generated recommendations from crossing directly into pharmaceutical consequences. Its propositions require computational, experimental, manufacturing, human-factors, and implementation validation. The framework does not establish model validity, formulation suitability, clinical benefit, regulatory acceptability, universal applicability, or deployment readiness.