S11.04 - Proportionality

S11.04 ? Proportionality

flowchart LR
    A[One-size-fits-all governance fails] --> B[RAIDT - run-level evidence framework]
    A2[Low-risk uses can be over-burdened] --> B
    A3[High-stakes uses can be under-governed] --> B
    B --> C[[Proportionality - calibrates governance effort to each run]]
    H1[Healthcare discharge summaries] --> C
    H2[Public-service casework] --> C
    H3[Educational feedback] --> C
    H4[Internal drafting] --> C
    C --> D[Right-sized evidence pack]
    C --> E[Meaningful RAIDT score profile]
    C --> F[Reviewer reconstruction]
    C --> G[Governance readiness]
    D --> I[Audit readiness]
    E --> J[Reviewability and contestability]
    F --> K[Organisational learning]

? Star S11 - Boundaries, Limitations and Future Questions

Star context: Prevents overclaiming and explains what RAIDT can and cannot solve by matching governance effort to the seriousness, context, and consequences of each run.

Academic picture

Definition / background

Proportionality means that the depth, form, and scrutiny of governance evidence should match the level of risk, consequence, and organisational dependence associated with a particular run of a generative AI system. In RAIDT, this is not merely a policy preference; it is a design principle for deciding how much documentation, review, justification, and follow-up should accompany a run-level evidence pack.

The concept has roots in law, regulation, risk management, and public administration, where interventions are expected to be commensurate with the seriousness of the issue being addressed. In AI governance, proportionality matters because generative AI is used across tasks with very different stakes. An internal brainstorming prompt, a citizen-facing decision support tool, and a clinical summarisation workflow should not all be governed identically. Treating them as equivalent either creates wasteful process or leaves important risks unmanaged.

Proportionality differs from simple minimalism. It does not mean capturing as little evidence as possible. It means capturing enough evidence to justify trust, enable review, and support accountability for the actual use context. For RAIDT, that distinction is crucial because the framework is built around the run as the unit of governance. The relevant question is not whether a model is generally risky in the abstract, but what happened in this configured use, at this time, for this task, in this organisational setting.

This is why proportionality belongs centrally within RAIDT. Run-level evidence packs and five-pillar score profiles are only useful if their depth reflects the significance of the run they describe. A proportionate approach keeps RAIDT credible as a governance method: serious enough to support audit readiness and contestability, but flexible enough to be deployable in real organisational work.

Why this concept matters

Without proportionality, AI governance frameworks often fail in one of two ways. They either become too generic to guide action, or they become too burdensome to use consistently. RAIDT is explicitly designed to avoid both outcomes. Proportionality helps translate governance ambition into a usable operating logic for real tasks, teams, and decision environments.

This concept matters because organisations need a defensible basis for deciding when light-touch oversight is acceptable and when enhanced evidence capture is necessary. It helps prevent confusion between low-stakes experimentation and high-stakes operational reliance. It also avoids the false reassurance that a standard template or a single policy statement is sufficient for every case.

If proportionality is missing, three risks appear. First, staff may resist RAIDT as administratively excessive. Second, high-impact uses may be governed too weakly because the same limited process is applied everywhere. Third, review bodies may be unable to see why a run was treated as acceptable, especially when harm, challenge, or audit occurs later.

Key idea: Proportionality makes RAIDT usable and credible by matching governance effort to the actual stakes of each run.

What this item controls

How much run-level evidence must be captured for a given use case.
How much human review, sign-off, or escalation is required before use.
How much justification is needed to explain model choice, prompt design, safeguards, and intended use.
How detailed the evidence pack should be for later reconstruction and challenge.
How scoring under the five RAIDT pillars should be interpreted in relation to risk and context.
How organisations distinguish exploratory, advisory, and consequential uses of generative AI.
How governance resources are allocated so that attention is concentrated where harm or reliance is greatest.

Practical example / likely audience question

Audience question

Is RAIDT too heavy?

Answer

Audience question: Is RAIDT too heavy? Answer: not if implemented proportionately according to task risk.

The concern behind the question is understandable. Many AI governance schemes appear to assume that every use of a model should produce the same volume of records, approvals, and review materials. That becomes impractical in organisations where generative AI is used across a wide range of tasks, from low-stakes drafting support to high-stakes decision preparation.

The direct answer is that RAIDT is not intended to be uniformly heavy. Its run-level structure allows governance effort to scale with the seriousness of the task. A low-risk internal ideation run may need basic metadata, prompt and output capture, and a short note on intended use. By contrast, a run that informs a welfare eligibility recommendation or a clinical summary would require stronger evidence, clearer reviewer roles, and a more developed explanation of safeguards, limitations, and verification.

RAIDT handles this better than a generic AI governance approach because it does not rely only on abstract policy tiers or model-level claims. It ties proportionality to evidence at the level where work actually happens: the individual run. That makes the justification inspectable. A reviewer can see not only that a lighter or heavier process was used, but why that level of governance was judged appropriate.

Practical example in RAIDT terms

Consider a healthcare organisation using a generative AI system to draft discharge summaries for clinicians. The use case is not fully automated decision-making, but it sits close to patient safety and continuity of care. The run-level issue is therefore not simply whether the model can generate fluent text; it is whether this particular run, for this patient context, was properly bounded, checked, and documented.

A proportionate RAIDT approach would require stronger evidence than would be expected for ordinary internal brainstorming. The evidence pack might include the task purpose, user role, model and version, prompt template, source data boundaries, human review confirmation, known error modes, and a record of whether any material corrections were needed before the summary was used. The most affected pillars would be Responsibility, Dependability, and Traceability, with Auditability also becoming important because later reconstruction may be necessary if a concern arises.

Proportionality improves governance readiness here because it justifies enhanced scrutiny without claiming that every healthcare-related run must be treated identically. A low-risk educational mock exercise for staff training and a live patient-facing summary workflow would sit at different evidence depths even if they use similar tools. RAIDT makes that distinction visible and defensible.

Detailed link to RAIDT

Proportionality links to RAIDT in four ways.

First, it supports RAIDT's core idea that governance should move from broad principle statements to inspectable evidence tied to actual use.
Second, it operates at the level of the run, because the amount of assurance required depends on the task, context, users, and consequences of that particular configured use.
Third, it shapes both the evidence pack and the score profile by determining what should be captured, how deeply it should be assessed, and how scores should be interpreted.
Fourth, it strengthens reviewability, contestability, audit readiness, and organisational learning by making the reasoning for lighter or heavier governance visible after the fact.

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

In this chain, proportionality is the calibration logic. It prevents the evidence pack from becoming either superficial or unnecessarily bloated, and it makes the resulting score profile more meaningful because the assessment is tied to the stakes of the run rather than to a one-size-fits-all template.

Link to the five RAIDT pillars

Responsibility

Proportionality helps define what level of human responsibility is needed for a run and what justification is expected from those deploying it. Higher-stakes uses require clearer ownership, stronger role definition, and more explicit acceptance of residual risk.