• qa_deck_100#slide 52 · Responsibility and auditability

In RAIDT, Responsibility is one of the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and is assessed with the run as the unit of governance. The foundations paper defines it as the extent to which a run satisfies ethical, legal and organisational obligations in its context of use. That definition is deliberately broader than simple harm avoidance. A responsible run must respect decision rights, avoid inappropriate overreach, communicate uncertainty and limitations, and leave enough information for contestability. In other words, Responsibility asks whether the specific use was appropriate, bounded, safe, authorised and overseen, rather than whether the model looks responsible in the abstract.

RAIDT therefore links Responsibility to a reconstructable run-level evidence pack. The relevant evidence includes the policies and constraints applied to the run, safety and compliance checks, uncertainty or limitation disclosures, and documented human oversight decisions about whether and how the output was used. The same run may also contain influence methods as governance interventions, such as prompt constraints or alignment layers, but these do not remove the need for organisational oversight. RAIDT treats Responsibility as a joint property of system behaviour and governance process: what the system produced matters, but so do the controls, approvals and escalation pathways surrounding that output. This is why Responsibility connects output use directly to organisational accountability, rather than to technical performance alone.

In the healthcare example used in the foundations paper, a GenAI system drafts a summary of a high-risk clinical presentation. A responsible run would not simply produce fluent text. The run-level evidence pack would show conservative prompting, an instruction not to make autonomous diagnoses, uncertainty and escalation guidance, and an oversight flag requiring clinician review before the note is adopted.

If the summary were filed without any recorded review decision, or if red-flag symptoms were omitted without escalation, Responsibility would be weak even if the prose looked polished. By contrast, Responsibility is strong when the evidence pack shows the task boundary, the policy layer applied, the safety checks performed, and the clinician?s approval or rejection decision.

• 08-RAIDT_Foundations_M_V50