From principles to criteria: a structured map of HTA guidance on generative AI

This report maps what guidance on generative AI issued by HTA bodies and HTA or HEOR societies specifies, and at what depth. Nine documents (2023–2026), forming six units, were coded on twelve normative requirements on a 0–3 scale, with a section reference for every code. Disclosure, accountability, human oversight, legal and security safeguards and representativeness of data are stated as principles by most units and operationalised by almost none. No document in the corpus specifies when AI-assisted evidence may be relied on, for any task, and none defines levels of risk. The components of such criteria exist — risk grading (FDA), verification (NICE HTA Lab), a three-outcome decision (RAISE) — but have not been assembled; the report outlines the assembly and the study that would set its thresholds. Files: the report (PDF) and the coding matrix (CSV: 72 cells with section references, the author's final codes and the model's preliminary codes). The use of a large language model in preparing the report is disclosed in §6 and Appendix D.