C0.09 - Core_value_contestability

C0.09 ? Core value: contestability

flowchart LR
    A[Opaque or disputed GenAI output] --> B[Need for meaningful challenge]
    B --> C[RAIDT
Run-level evidence framework]
    C --> D[[Core value: contestability]]
    D --> E[Run-level evidence]
    E --> F[Evidence pack]
    E --> G[RAIDT score profile]
    D --> H[Reviewer reconstruction]
    F --> I[Correction or override]
    G --> J[Governance readiness]
    H --> J
    K[Healthcare, finance, public services, law, enterprise work] --> D
    I --> L[Organisational learning]
    J --> L

? Star C0 - RAIDT Core, Definition, Values, Claims and Innovation

Star context: Positions contestability as one of the core values that gives RAIDT its governance logic: GenAI use in organisational work should not only produce outputs, but produce outputs that can be challenged, reviewed, reconstructed, and, where necessary, corrected through run-level evidence.

Definition / background

Contestability means that a stakeholder can question, challenge, or seek review of a generated output or decision-support artefact because the basis of that output can be inspected in a structured way. In the context of GenAI governance, contestability is the practical condition that makes disagreement actionable. It is the difference between saying that a person may object and giving that person, reviewer, or organisation enough evidence to understand what happened in a specific AI-assisted run.

Conceptually, contestability sits close to reviewability, accountability, transparency, and procedural fairness, but it is not identical to any of them. Transparency may reveal general information about a system. Explainability may help someone understand why an output appears plausible. Accountability may assign responsibility for decisions. Contestability goes further by supporting an actual challenge process: a person can inspect evidence, identify grounds for concern, and ask for correction, escalation, or override.

This is why contestability belongs inside RAIDT. RAIDT is built around the run and run-level evidence, not only around policy statements. Because RAIDT generates an evidence pack and a score profile, contestability becomes operational rather than rhetorical. The ability to challenge a run is linked to what can actually be reconstructed, reviewed, and evidenced.

Within the RAIDT model, contestability also has a strong relationship to the five pillars. It depends especially on Responsibility, Auditability, and Traceability, but it also relies on Interpretability and Dependability. A contestable run is one in which an output can be traced to relevant inputs, configuration, context, and review records, and in which the organisation has enough governance structure to act on that evidence.

Why this concept matters

Contestability addresses a central governance failure in GenAI deployment: organisations may permit AI-assisted outputs to influence work without creating a realistic mechanism for those outputs to be challenged. In practice, many systems are reviewable only in theory. A stakeholder may be told that a human remains in the loop, yet the human reviewer may have no reliable record of the prompt, model version, retrieved material, parameters, edits, or surrounding task context. Without that evidence, challenge becomes weak, slow, and inconsistent.

The concept therefore prevents a common confusion. It is not enough to say that someone can appeal a decision or ask for reconsideration. A challenge process is only meaningful if there is sufficient evidence to reconstruct the run that produced the contested artefact. RAIDT matters here because it turns contestability into an operational governance requirement attached to the run itself.

If contestability is missing, organisations face several risks: uncorrected errors, poor defensibility in regulated settings, weak learning from failures, and false confidence that human oversight is working. Missing contestability can also damage legitimacy, because affected stakeholders may perceive AI-assisted processes as unchallengeable even when formal governance documents say otherwise.

For organisations using GenAI, contestability helps move governance from principles to practice. It supports justified intervention, reviewer confidence, structured escalation, and evidence-based improvement. It also aligns with RAIDT's broader aim of moving from assertion to evidence, and from static compliance language to usable governance mechanisms.

Key idea: Contestability matters because responsible GenAI governance requires not only the possibility of challenge, but the evidence needed to make challenge meaningful at the level of a specific run.

What this item enables

It enables stakeholders, reviewers, and organisations to challenge a specific GenAI-assisted output on evidential rather than purely subjective grounds.
It enables reconstruction of how a contested artefact was produced, including task context, configuration, inputs, outputs, and review activity.
It enables justified correction, override, escalation, or retention of an output after structured examination.
It enables learning from contested runs so that repeated weaknesses can inform prompts, controls, workflows, policies, and score thresholds.
It enables RAIDT to connect governance values to practical artefacts such as the evidence pack and score profile.

Practical example / likely audience question

Audience question

If a human already reviews the output, why is contestability needed as a separate value?

Answer

The concern behind this question is the assumption that human review automatically makes an AI-assisted process governable. In reality, a human may review an output without having enough evidence to understand how it was produced or why it should be trusted. A reviewer who only sees the final summary, recommendation, or draft is not yet in a strong position to challenge it.

Contestability matters because it strengthens the quality of review. It gives the reviewer or affected stakeholder a basis for saying, for example, that a crucial document was omitted, a prompt framed the task incorrectly, retrieved context was misleading, or a later human edit introduced a distortion. The direct answer is that human review without evidential support may be superficial, whereas contestability requires the ability to inspect the run in a structured way.

A practical example is a clinician reviewing a GenAI-generated discharge summary. The clinician may sense that the summary is incomplete, but contestability allows them to do more than rely on intuition. They can inspect the run record, see what source material was used, examine prompt instructions, identify whether the model omitted a medication change, and record the grounds for correction.

RAIDT handles this issue better than a generic AI governance approach because it ties contestability to run-level evidence rather than to a generic statement that outputs should be reviewed. That makes challenge reproducible, documentable, and usable for governance learning.

Practical example in RAIDT terms

Consider a public-services use case in which a caseworker uses GenAI to summarise a housing-support application and highlight potential safeguarding concerns. One run produces a summary that fails to mention a disclosed disability adjustment request. The applicant later challenges the handling of the case because the omission may have influenced the next stage of support assessment.

In RAIDT terms, the run-level issue is not simply that the output was imperfect. The issue is whether the organisation can inspect the specific run closely enough to determine what happened. Useful evidence would include the prompt or instruction template, the documents supplied to the model, retrieval context if any retrieval-augmented workflow was used, model and configuration details, timestamps, the identity or role of the operator, the generated output, any post-editing by staff, and the review notes associated with the case.

The most affected RAIDT pillars are Responsibility, Auditability, and Traceability, with Interpretability also playing a major role. Responsibility matters because someone must act on the challenge. Auditability matters because the run must be reconstructable. Traceability matters because the omitted issue must be linked back to available source material and workflow steps. Interpretability matters because the reviewer must understand enough about the run to decide whether the omission resulted from prompt design, missing evidence, model weakness, or human handling.

By making the run contestable, RAIDT improves governance readiness. The organisation is better placed to explain what occurred, correct the record, review whether safeguards worked, and learn whether the problem was isolated or systematic. Without that structure, the challenge would likely collapse into unsupported disagreement.

Detailed link to RAIDT

Contestability links to RAIDT in four ways.

First, it supports RAIDT's core idea that responsible GenAI governance should be grounded in inspectable evidence rather than broad claims of responsible use.
Second, it depends on the run and run-level evidence because a challenge must refer to a specific configured use in a specific context.
Third, it becomes practical through the evidence pack and the score profile, which organise what can be inspected and how governance strengths or weaknesses are signalled.
Fourth, it strengthens reviewability, audit readiness, and organisational learning by turning disputed outputs into analysable governance events rather than isolated complaints.

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

This chain matters because RAIDT does not treat contestability as an abstract user right. It treats it as an evidential property of AI-assisted work that can be examined, scored, improved, and defended.

Link to the five RAIDT pillars

Responsibility

Contestability requires clear ownership for receiving, assessing, and acting on a challenge. If no person or role is responsible for reviewing contested outputs, contestability remains nominal.