• integrated_82#Q3.11

The Responsibility pillar in RAIDT is the pillar that tests whether a specific GenAI use was governed in a way that is ethically, legally and organisationally appropriate for its context. It is not a general statement that the organisation values responsible AI; it is a run-level judgement about whether the use was bounded, authorised, monitored and contestable. Both the foundations and evidence-review papers emphasise that responsibility becomes meaningful only when grounded in artefacts and socio-technical arrangements rather than left as an abstract principle. This is why RAIDT treats the run as the unit of governance and requires a run-level evidence pack for review.

The evidence supporting the Responsibility pillar is therefore concrete and reviewable. At minimum, reviewers need context-of-use information, user role and intended audience, task constraints, policy references, prompt or policy-layer controls, the output itself, communicated uncertainties or limitations, and an oversight trace showing approvals, edits, exceptions or escalations. The evidence-review paper?s design requirements reinforce this by highlighting contextual completeness, configuration provenance, output integrity, and especially oversight trace as core features of a governance-ready evidence object. The foundations paper adds that Responsibility must combine behavioural evidence with process evidence. In RAIDT terms, the pillar is supported when reviewers can show not only what the model said, but under what authority, under what controls, for what purpose, and with whose recorded decision the output entered the workflow.

In an HR performance-appraisal workflow, the Responsibility pillar would be supported by evidence that the run was for draft assistance only, that the manager using the tool had the right role, that the relevant policy wording or template version was fixed, and that prohibited grounds were excluded from consideration. The run-level evidence pack should also show the generated text, any edits made by the manager, and the final approval decision.

If a dispute arises, the organisation can then demonstrate that the appraisal was not an unbounded AI judgement but a bounded, overseen use within a governed process. If those approvals, constraints and policy references are missing, the Responsibility pillar is weak even if the final appraisal appears reasonable.

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