S1.02 - Responsible_AI

S1.02 ? Responsible AI

flowchart LR
    A[Responsible AI concerns:
fairness, transparency, accountability, safety, oversight]
    B[Traditional limitation:
principles without run reconstruction]
    A --> C[RAIDT - run-level evidence framework]
    B --> C
    C --> D[[Responsible AI operationalised in RAIDT]]
    D --> E[Run-level evidence pack]
    D --> F[Five-pillar score profile]
    D --> G[Evidence over assertion]
    E --> H[Reviewer reconstruction]
    F --> I[Governance readiness]
    G --> I
    J[Healthcare, finance, public services,
education, enterprise productivity] --> D
    H --> K[Organisational learning and policy alignment]
    I --> K

← Star S1 - Origins, Background and History

Star context: Explains why RAIDT begins with the Responsible AI agenda but moves beyond abstract principles towards run-level evidence, reconstructability, scoring, and governance readiness for organisational GenAI use.

Academic picture

Definition / background

Responsible AI is the broad governance and normative agenda concerned with ensuring that AI systems are designed, deployed, and used in ways that are fair, safe, accountable, transparent, and aligned with human and organisational values. It developed as a response to the risks of opaque automated decision-making, harmful bias, weak accountability arrangements, and the difficulty of contesting machine-supported outputs once they enter consequential practice.

In the GenAI context, Responsible AI matters because large models can produce plausible but unstable outputs, can be configured in many different ways, and can be used by staff for tasks that carry legal, operational, reputational, or ethical consequences. High-level Responsible AI principles remain necessary, but they do not by themselves tell an organisation what evidence must be kept to reconstruct one use of a model for one task at one time in one setting.

This is why the concept belongs inside RAIDT. RAIDT does not replace Responsible AI; it operationalises it. The framework treats the run as the unit of governance and asks what evidence is needed to demonstrate responsible practice at that level. The run-level evidence pack and the five-pillar score profile translate broad Responsible AI expectations into inspectable artefacts. Responsibility becomes attributable, auditability becomes recordable, interpretability becomes contextual, dependability becomes assessable, and traceability becomes reconstructable.

Responsible AI is therefore the normative origin of RAIDT, whereas RAIDT is the operational governance response. Similar terms such as AI ethics, trustworthy AI, or ethical AI often remain principle-centred. In RAIDT, the emphasis shifts from general commitments to evidence-backed governance of a specific GenAI run.

Why this concept matters

Responsible AI matters in RAIDT because it clarifies the governance objective before the framework specifies the evidence mechanism. Without this conceptual anchor, run-level logging and scoring could become technically precise but normatively shallow. The framework needs a clear account of what responsible use is meant to achieve.

At the same time, the concept matters because it exposes a practical gap. Many organisations already have AI principles, review boards, or policy statements, yet still cannot answer simple operational questions about one GenAI use: what configuration was used, what human review occurred, what limits were defined, what risks were anticipated, and how a contested output would be examined after the fact. Responsible AI identifies the need for good governance; RAIDT provides a disciplined way to evidence it.

If this concept is missing, governance can default to slogan-based assurance. Organisations may claim that a model is fair, accountable, or human-supervised without preserving the records needed to test those claims. RAIDT matters precisely because it converts Responsible AI from a declarative posture into an evidential one.

Key idea: Responsible AI sets the normative standard, but RAIDT makes that standard reviewable by tying it to run-level evidence, scoring, and reconstruction.

What this item explains

The normative commitments that RAIDT inherits from the Responsible AI agenda.
Why principle-level AI governance is necessary but insufficient for governing one GenAI run.
How abstract values such as fairness, accountability, and oversight are translated into inspectable run-level evidence.
Why RAIDT?s evidence pack and five-pillar score profile are operational responses to Responsible AI demands.
How organisations move from policy assertions to reviewable governance practice when using GenAI in real work.

Practical example / likely audience question

Audience question

If an organisation already has a Responsible AI policy, ethics principles, and an approval process, why is RAIDT still needed?

Answer

The concern behind this question is the belief that high-level governance commitments are enough to demonstrate responsible use. They are not. A policy can state that outputs must be fair, safe, and human-reviewed, but it does not automatically show whether one particular GenAI run actually met those conditions.

The direct answer is that Responsible AI principles define expectations, whereas RAIDT specifies the evidence needed to test whether those expectations were met in practice. A generic governance approach may document the existence of policy, training, or governance structures. RAIDT asks a different question: for this run, with this prompt, this model, this configuration, this user, this task, and this review pathway, what can be reconstructed and defended?

Consider a team using a GenAI assistant to draft a customer-facing policy summary. The organisation may already require transparency and human oversight. However, if a complaint arises, reviewers still need to know which model version was used, whether retrieval or external tools were enabled, what prompt instructions shaped the output, what review occurred before release, and what risks were recognised at the time. RAIDT handles this better than a generic AI governance approach because it converts Responsible AI into a run-level evidence structure that can support contestability, audit, and organisational learning.

Practical example in RAIDT terms

A public services team uses a GenAI system to draft responses to housing support enquiries. The use case appears aligned with Responsible AI principles because staff remain in the loop and the system is presented as an assistive tool rather than an autonomous decision-maker. However, a run-level issue emerges when one response gives misleading eligibility guidance after a prompt template is updated and a model parameter change alters output style and confidence.

In RAIDT terms, the evidence needed includes the prompt template used in that run, model and version details, temperature or other relevant configuration settings, the source material retrieved or relied upon, the staff member who reviewed the draft, the decision on whether to send or amend the output, and the rationale for deployment conditions. Responsibility is affected because review roles and escalation responsibilities must be clear. Auditability is affected because investigators need a reconstructable record. Interpretability is affected because reviewers must understand what factors shaped the output. Dependability is affected because output stability and suitability matter in a citizen-facing setting. Traceability is affected because the run must be linked to inputs, outputs, approvals, and downstream action.

This is where Responsible AI improves governance readiness through RAIDT. Instead of merely saying that the system is used responsibly, the organisation can assemble evidence showing how that claim is supported for the contested run. That is the difference between principle compliance in the abstract and reviewable governance in practice.

Detailed link to RAIDT

Responsible AI links to RAIDT in four ways.

First, it supplies the normative rationale for the framework. RAIDT exists because organisations need a practical way to enact fairness, accountability, safety, and human oversight in real GenAI use rather than leaving them at the level of aspiration.

Second, it connects directly to the run as the unit of governance. RAIDT asks how responsible practice can be evidenced for one configured use of a GenAI system for one task at one time in one context.

Third, it shapes both practical outputs of RAIDT. The evidence pack records the artefacts needed to examine whether a run met responsible-use expectations, while the score profile translates those expectations into a structured assessment across Responsibility, Auditability, Interpretability, Dependability, and Traceability.

Fourth, it strengthens reviewability, contestability, audit readiness, and organisational learning. When Responsible AI is operationalised through run-level evidence, organisations can compare cases, explain decisions, challenge weak practice, and improve governance arrangements over time.

Responsible AI ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

This chain is central to the project because RAIDT turns Responsible AI from a policy aspiration into an operational governance pathway.

Link to the five RAIDT pillars

Responsibility

Responsible AI most directly grounds the Responsibility pillar because it asks who is answerable for configuring, reviewing, approving, and acting on GenAI outputs in organisational settings.