S2.07 - Contestability

S2.07 — Contestability

flowchart LR
    A[Background problem
disputed GenAI-supported outputs
generic policy gives weak recourse] --> B[RAIDT
run-level evidence framework]
    H[Practical artefacts and contexts
prompt, source material, model settings, output, review notes, escalation, public services] --> C[[Contestability
evidence-backed challenge of a specific run]]
    B --> C
    C --> D[Evidence pack]
    C --> E[RAIDT score profile]
    D --> F[Reviewer reconstruction and correction]
    E --> G[Governance readiness and organisational learning]

← Star S2 - Governance Meaning and Problem Context

Star context: Clarifies governance as a practical capacity for oversight, accountability, reviewability, contestability and improvement, showing that responsible GenAI governance requires challengeable decisions rather than broad ethical aspiration alone.

Academic picture

Definition / background

Contestability is the ability of a stakeholder to question, challenge, and seek review of a GenAI-supported output or decision on the basis of examinable evidence and an intelligible review path. In governance terms, it is the difference between merely asserting that oversight exists and making it possible for a disputed outcome to be revisited in a disciplined way. The concept draws on broader traditions of procedural fairness, accountable decision-making, and review rights, but in RAIDT it is translated into a practical requirement for evidence-backed challenge.

In GenAI governance, contestability matters because high-impact uses often produce outputs that are probabilistic, revisable, context-sensitive, and capable of influencing human judgement. A stakeholder may need to ask what information was used, how the output was generated, who relied on it, what checks were performed, and whether correction is possible. Without a structured basis for such questions, governance remains declarative and recourse becomes weak.

Contestability is closely related to, but distinct from, neighbouring terms. Reviewability means a run can be inspected; reconstructability means the event can be rebuilt after the fact; interpretability helps reviewers understand how an output arose; accountability identifies who is responsible for acting on the review. Contestability adds the practical capacity to challenge and potentially change an outcome. It therefore depends on those adjacent concepts, but it is not reducible to any one of them.

This item belongs inside RAIDT because RAIDT treats the run as the unit of governance. Contestability becomes operational only when a stakeholder can point to run-level evidence, use an evidence pack to support the review, and connect the result to a score profile and governance judgement. In that sense, contestability is one of the reasons RAIDT moves beyond principle-led governance towards evidence-led oversight.

Why this concept matters

Contestability solves a recurring governance problem: organisations may claim that humans remain in control, yet when a questionable GenAI-supported outcome appears, it is often unclear how that outcome can be challenged in practice. If there is no clear evidence trail, no route to re-examination, and no identified reviewer, the appearance of accountability masks a lack of operational recourse.

The concept also avoids confusion between dissatisfaction and governance. People can always disagree with an output, but governance-quality contestability means the disagreement can be examined against evidence, criteria, and decision ownership. This matters particularly in organisational settings where outputs may influence patient communication, case prioritisation, financial advice, internal investigations, or public-facing administrative actions.

If contestability is missing, several risks emerge: affected people may have no meaningful avenue for correction, reviewers may be unable to determine whether the problem arose from the model, the prompt, the source material, or the workflow, and organisations may struggle to defend their practice to supervisors, auditors, clients, or regulators. RAIDT uses contestability to convert abstract governance commitments into an evidence-based process of challenge, review, and improvement.

Key idea: Contestability matters because responsible GenAI governance requires not only oversight in principle, but a practical, evidence-backed way to challenge and correct disputed runs.

What this item enables

A meaningful route for questioning a specific GenAI-supported output or decision.
The use of run-level evidence to distinguish between model error, workflow failure, weak prompting, and poor human review.
Documented escalation, correction, override, or re-run decisions when an output is disputed.
More defensible governance claims because challenge and response are tied to evidence rather than assertion.
Organisational learning from contested cases, near misses, and recurrent points of failure.
Clearer links between stakeholder concern, reviewer action, and governance readiness across the RAIDT pillars.

Practical example / likely audience question

Audience question

Is contestability in RAIDT just another name for an appeals process, or for general human oversight?

Answer

The concern behind this question is that the term can sound broader than it really is. The direct answer is no: contestability is not simply the existence of a complaint channel, and it is not satisfied merely because a human was somewhere in the loop. A generic appeals process can exist without enough evidence to examine what happened, and nominal human oversight can exist without any realistic basis for correction.

A practical example is a local authority team using GenAI to draft an internal summary that informs a homelessness-priority assessment. If a supervisor or affected citizen later questions whether the summary omitted crucial context, contestability requires more than saying that staff can raise concerns. It requires the run to be reconstructable: what prompt was used, which case notes were supplied, what output was generated, who edited it, who relied on it, and how the review should proceed.

RAIDT handles this better than a generic AI governance approach because it connects challenge to one run, one evidence trail, and one governance judgement. Instead of treating contestability as a broad ethical aspiration, it asks whether the organisation can inspect the disputed event, assemble the evidence pack, justify the score profile, and determine whether correction or escalation is warranted.

Practical example in RAIDT terms

Consider a public-services setting in which a caseworker uses a GenAI assistant to draft a summary of an applicant's circumstances for a housing-support assessment. The GenAI use case is administratively attractive because it reduces drafting time, but the run-level issue is whether the generated summary mischaracterises vulnerability factors and therefore influences prioritisation unfairly.

The evidence needed includes the task purpose, the prompt template, the case notes and policy extracts supplied as input, the model or tool version, the generated summary, the caseworker's edits, the final summary sent forward for review, and a record of who approved or relied on it. Responsibility is affected because the organisation must show who owned the review and correction decision. Auditability is affected because a later reviewer must be able to reconstruct the run. Interpretability is affected because reviewers need to understand how the output related to the source material and prompt. Dependability is affected because repeated contested summaries may indicate unstable or poor-quality workflow performance. Traceability is affected because the case must be linked to time, actor, artefacts, and subsequent action.

In governance-readiness terms, contestability improves the organisation's position because a disputed outcome can be examined as a specific evidential event rather than as an anecdotal complaint. That supports fairer internal review, better incident handling, more credible assurance, and more disciplined organisational learning.

Detailed link to RAIDT

Contestability links to RAIDT in four ways.

First, it gives practical force to the RAIDT core idea that responsible governance should attach to real uses of GenAI in organisational work, not only to abstract principles or supplier claims.

Second, it depends on the run as the unit of governance, because a challenge becomes meaningful only when it can be tied to one identifiable GenAI event with sufficient context for re-examination.

Third, it relies on the evidence pack and the score profile: the evidence pack gathers the material needed for review, while the RAIDT score profile shows whether the run met an acceptable standard across the five pillars.

Fourth, it reinforces reviewability, audit readiness, and organisational learning by ensuring that problematic or disputed outputs can lead to structured scrutiny, correction, and improvement rather than informal disagreement alone.

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

Link to the five RAIDT pillars

Responsibility

Contestability strengthens Responsibility by identifying who must respond when a run is challenged and who has authority to uphold, revise, or escalate the outcome.