Co-Defining Transparency for Clinical AI: A Replicable Participatory Methodology

Transparency is the most invoked safeguard for clinical AI, yet frameworks (e.g.\ the EU AI Act, FUTURE-AI) are specified top-down by regulators and technical experts, not bottom-up by the frontline users whose trust and adoption they should calibrate. This disconnect exacerbates the central deployment failure of clinical AI, \emph{miscalibrated reliance and adoption} (such as automation bias, deskilling, or disuse), which technical explanation alone does not resolve. We argue that ``how much transparency, of what kind, and for whom'' is an empirical, context-dependent question, and present a replicable participatory methodology to answer it, instantiated at a large European university hospital. A cumulative chain of six workshops engages an inter-professional panel (such as clinicians, patient and caregiver partners, and students) in agnostic, open-ended elicitation: 275 stakeholder verbatims and external benchmarks (eCH-0272, EU AI Act, Stanford FMTI) yield 20 indicators (18 prioritised and 2 recovered at later stage), rated across five clinical scenarios of graduated intervention before and after technical briefings, discussed with legal experts, and translated into concrete implementation features. As an existence proof, the use case yields decision-useful signals: the right to human review of automated decisions is the single most critical safeguard (84\% of participants), and required transparency rises monotonically with the depth of intervention (AI informational tools to AI treatment tools), corroborating World Health Organisation (WHO) risk-proportionality; a short explainability briefing did not substantially shift participants' transparency requirements. Of 151 evaluated solutions, 79 were accepted as-is. We position the six-workshop protocol (not its single-site outcomes) as the transferable artefact and release all instruments to support replication.