• workshop_dense_100#slide 59

Interpretability in RAIDT is defined pragmatically rather than aesthetically. It asks whether a GenAI output, together with its limits, is understandable enough for the intended user and task. The concept is broader than technical model interpretability because RAIDT focuses on situated use: a clinician, caseworker, analyst, or manager must be able to see what the output means, what evidence or criteria it draws on, where uncertainty remains, and when further review is needed. Because RAIDT treats the run as the unit of governance, interpretability is evaluated from the run-level evidence pack and expressed in the score profile alongside the other four pillars.

It matters because fluent text can be dangerously over-persuasive. The papers argue that generative systems can sound credible even when ungrounded, so governance cannot equate readability with trustworthy use. RAIDT therefore links interpretability to structured prompts, output schemas, source linkage, and uncertainty communication, and it requires those design choices to be recorded so they can be assessed later. This also explains why interpretability matters to RAIDT as a governance outcome: without it, reviewers cannot judge appropriate reliance, affected people cannot contest reasoning, and organisations cannot learn which configurations produce clearer or more misleading outputs. In short, interpretability helps convert a plausible-looking answer into a governable artefact whose meaning, evidential basis, and limits are visible enough for responsible use.

In healthcare, a hospital may use GenAI to draft a discharge summary from clinician notes. A weakly interpretable run would output a confident narrative with no indication of missing test results, no separation between extracted facts and generated bridging text, and no explanation of which internal guideline excerpt shaped the wording. A clinician may read it quickly and assume the system has stronger evidence than it actually has.

RAIDT would improve this by using a structured discharge-summary prompt that requires sections for confirmed findings, pending information, uncertainty, and limitations, potentially with a retrieved internal guideline excerpt preserved by identifier. Those prompt choices are logged because influence methods as governance interventions must be reviewable. The resulting run-level evidence pack lets the clinician and a later auditor understand what the draft means, what it omits, and why reliance should remain provisional. That is why interpretability matters directly to safe governance rather than only to model explanation.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10