S5.12 - Trade-offs

S5.12 ? Trade-offs

flowchart LR
    A[Background tensions
privacy, cost, latency, workflow burden] --> B[RAIDT
run-level evidence framework]
    H[Practical fields
logging depth, redaction, human review, retention rules] --> C[[Trade-offs
visible, evidence-based balancing]]
    B --> C
    C --> D[Evidence pack]
    C --> E[Score profile]
    D --> F[Reviewer reconstruction
contestability]
    E --> G[Governance readiness
organisational learning]

? Star S5 - RAIDT Pillars and Scoring

Star context: Defines how the five RAIDT governance dimensions should be read together so that scoring remains measurable without hiding the practical tensions between privacy, accountability, usability, reliability, and organisational cost.

Academic picture

Definition / background

Trade-offs are the practical tensions that arise when improving one governance-relevant property of a generative AI run places pressure on another property, resource, or organisational objective. In RAIDT, trade-offs are not treated as excuses for weak governance, but as conditions that must be surfaced, evidenced, and judged explicitly. A run may become more traceable through richer logging, for example, while also creating more privacy exposure, more storage burden, or more operational friction.

Conceptually, this item sits between scoring logic and governance interpretation. A trade-off is not the same as a failure, because some tensions are unavoidable in sociotechnical systems. It is also not the same as a compromise in the casual sense. In RAIDT, a trade-off becomes meaningful only when it is tied to a specific run, a concrete workflow, and reviewable evidence showing what was gained, what was constrained, and why the balance was considered acceptable.

This matters in GenAI governance because many governance conversations assume that every desirable quality should rise together. In practice, stronger controls may reduce speed, fuller records may increase data-handling risk, and more intensive human review may improve responsibility while affecting scalability. RAIDT therefore treats trade-offs as part of the reality of governing organisational GenAI use rather than as noise to be ignored.

Within RAIDT, the concept belongs in Pillars and Scoring because the five-pillar profile is designed to preserve unevenness. A run-level evidence pack can document the reasons for those tensions, while the score profile makes their pattern visible across Responsibility, Auditability, Interpretability, Dependability, and Traceability. Trade-offs therefore link the evidential basis of a run to the judgement of governance readiness.

Why this concept matters

Trade-offs solve a recurrent problem in AI governance discussions: the false expectation that responsible governance means maximising every desirable property simultaneously. In real organisational settings, that is rarely possible. RAIDT helps reviewers see where tensions are structural, where they result from poor design, and where they may be justified only under certain controls.

The concept also prevents scoring from becoming misleading. If RAIDT reported only a single composite impression, reviewers might miss the fact that a run is highly auditable but weak on privacy-sensitive traceability controls, or dependable in output consistency but poorly interpretable to frontline staff. Making trade-offs explicit keeps the framework honest about uneven governance performance.

If this concept is missing, organisations are likely to drift into one of two errors: either they overclaim readiness by hiding tensions behind broad assurances, or they reject useful controls because every new burden is seen as evidence that governance has failed. RAIDT offers a middle position in which trade-offs are examined, bounded, and justified through evidence.

Key idea: Trade-offs matter because RAIDT treats governance readiness as an evidence-based judgement about tensions between pillars, not as a simplistic claim that every governance objective can be maximised at once.

What this item explains

Why RAIDT uses a score profile rather than relying only on a flattened overall impression.
How gains in one pillar can create pressure on another pillar or on wider organisational constraints such as privacy, cost, speed, or workflow burden.
Why uneven scores are often analytically useful rather than signs that the framework has failed.
What kinds of evidence are needed to justify a claimed balance between competing governance objectives.
How reviewers, supervisors, and organisations should interpret tensions in run-level governance rather than hiding them.
Why trade-offs are central to audit readiness, contestability, and continuous improvement in GenAI use.

Practical example / likely audience question

Audience question

Can stronger traceability create a new governance burden by increasing privacy risk or administrative overhead?

Answer

Yes. The concern behind the question is that governance improvements are not free. A more detailed trace can make a run easier to reconstruct, audit, and challenge, but it may also capture more sensitive information, require stricter access controls, and impose extra work on staff. The direct answer is therefore that stronger traceability can create real burdens, but RAIDT treats those burdens as governable trade-offs rather than as reasons to avoid evidence altogether.

A practical example is a team that stores detailed prompts, contextual notes, reviewer comments, and output histories for every GenAI-assisted case summary. That improves auditability and traceability because a reviewer can see exactly what happened. However, if those records contain personal data or sensitive case details, the organisation must also manage retention limits, redaction, access rights, and lawful handling. The issue is not whether traceability is good in the abstract; it is whether the organisation can evidence a proportionate balance.

RAIDT handles this issue better than a generic AI governance approach because it works at run level. Instead of making a broad claim such as ?we keep sufficient logs?, RAIDT asks what was captured for this run, what privacy implications followed, what safeguards were applied, and how the resulting score profile should be interpreted. That makes the trade-off inspectable rather than rhetorical.

Practical example in RAIDT terms

Consider a public-services casework team using a GenAI drafting assistant to prepare first-pass response letters for citizens. The use case is administratively valuable because it speeds up routine drafting and helps staff standardise tone. The run-level issue is that detailed evidence capture would make the drafting process easier to reconstruct, but the same documentation may contain sensitive personal details, internal case reasoning, and staff comments that require careful handling.

The evidence needed includes the task definition, prompt template, case-data minimisation rules, model or tool version, output draft, reviewer edits, approval record, retention rule, and access-control decision for stored artefacts. Responsibility is affected because the team must show who approved the use and who remained accountable for the final letter. Auditability and Traceability are affected because the run must be reconstructable. Dependability is affected because the quality and consistency of the drafting process matter over time. Interpretability is affected because supervisors need to understand why a particular prompt structure and review workflow produced the final text.

Trade-offs become visible when the team decides how much contextual material to store. Capturing too little weakens reconstruction and challenge. Capturing too much may create privacy exposure and workflow burden. RAIDT improves governance readiness by requiring the team to evidence that choice, justify it in the evidence pack, and reflect its consequences in the score profile rather than pretending there was no tension.

Detailed link to RAIDT

Trade-offs link to RAIDT in four ways.

First, they reinforce the RAIDT core idea that governance should be based on what happened in real organisational use rather than on abstract claims that controls are simply ?in place?.

Second, they link directly to the run because trade-offs become meaningful when attached to one configured GenAI use in a particular context, with a particular task, at a particular time.

Third, they shape both the evidence pack and the score profile. The evidence pack records the reasons, controls, and consequences behind a governance tension, while the score profile shows how that tension appears across the five pillars.

Fourth, they support reviewability, contestability, audit readiness, and organisational learning by allowing reviewers to ask whether the balance struck in one run was justified, proportionate, and improvable.

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

Link to the five RAIDT pillars

Responsibility

Trade-offs affect Responsibility because someone must remain answerable for how competing governance objectives were balanced in a run. The key question is not only what trade-off existed, but who had authority to accept it and under what organisational rationale.