S11.05 - Privacy_and_data_protection

S11.05 ? Privacy and data protection

flowchart LR
    A[Traditional AI governance problem:
need evidence but logging can expose sensitive data] --> B[RAIDT:
Run-level evidence framework]
    B --> C[[Privacy and data protection:
governs capture, minimisation, access, retention]]
    C --> D[Evidence pack:
redacted or controlled run record]
    C --> E[RAIDT score profile:
privacy-aware governance judgement]
    D --> F[Reviewer reconstruction]
    D --> G[Contestability]
    E --> H[Audit readiness]
    E --> I[Organisational learning]
    J[Healthcare] --> C
    K[Finance] --> C
    L[Public services] --> C
    M[Education] --> C
    N[Enterprise productivity] --> C

? Star S11 - Boundaries, Limitations and Future Questions

Star context: This item marks a practical boundary for RAIDT: evidence improves governance only if the evidence pack itself is handled in ways that protect sensitive data, respect proportionality, and avoid creating new organisational harms through logging.

Academic picture

Definition / background

Privacy and data protection in RAIDT refer to the disciplined governance of any sensitive information that may enter the run record, the evidence pack, or the surrounding review process. Because RAIDT treats the run as the unit of governance, it necessarily pays attention to the concrete artefacts generated by one configured use of a generative AI system for a specific task, at a specific time, in a specific context. Those artefacts may include prompts, outputs, attached sources, model settings, reviewer notes, user roles, and downstream actions. Any of these can contain personal data, confidential organisational material, or other restricted content.

Conceptually, this item sits at the intersection of information governance, responsible AI, records management, and risk control. It is concerned not only with whether data are lawfully processed, but also with whether evidence capture is proportionate to the governance purpose. That distinction matters. A system can generate rich evidence while still creating avoidable privacy risk if it logs too much, stores it too long, or grants overly broad access.

Within RAIDT, privacy and data protection therefore differ from a generic compliance statement. They become run-level design questions: what should be captured, what should be redacted, who should see it, how long should it be retained, and how can reviewers reconstruct a run without exposing unnecessary detail. This makes the concept structurally important to both the evidence pack and the five-pillar score profile.

The item belongs in RAIDT because responsible governance cannot rely on evidence collection alone. Evidence must itself be governed. Otherwise, a framework intended to improve accountability may introduce new harms by creating sensitive records without adequate minimisation, access control, or retention discipline.

Why this concept matters

Privacy and data protection matter because governance systems often fail in one of two opposite ways: either they capture too little evidence to support review, or they capture so much detail that the evidence pack becomes a new source of institutional risk. RAIDT is specifically designed to avoid this false choice. It seeks evidence that is sufficient for reviewability and contestability, but proportionate to the sensitivity of the run.

This concept prevents a common confusion in AI governance: the assumption that more logging is always better. In practice, indiscriminate logging can expose personal data, leak confidential business information, reveal protected case material, or create discoverable records that were never intended for broad internal circulation. If privacy controls are absent, the evidence pack may become legally difficult to manage, ethically questionable, or operationally unusable.

For organisations using generative AI, this item turns privacy from a broad principle into an operational governance discipline. It helps teams decide when raw prompts should be masked, when source excerpts should be abstracted, when access should be role-based, when retention should be shortened, and when evidence capture should be limited to metadata rather than content. In that sense, it moves AI governance from general aspiration to implementable control.

Key idea: Privacy and data protection matter in RAIDT because run-level evidence is only governance-ready when the evidence pack is informative enough to review, but restrained enough to avoid creating unnecessary exposure.

What this item controls

The scope of run evidence captured from prompts, outputs, source materials, user context, and system configuration.
The degree of data minimisation applied before evidence enters the pack.
The redaction, masking, pseudonymisation, or abstraction of sensitive fields.
The access rights granted to reviewers, auditors, managers, and system operators.
The retention period and deletion rules for evidence linked to completed runs.
The conditions under which raw content is stored versus when metadata-only capture is more appropriate.
The balance between accountability needs and exposure risk across different organisational settings.
The practical governability of RAIDT outputs in high-sensitivity domains.

Practical example / likely audience question

Audience question

Can logging create risk?

Answer

Yes. The misconception behind the question is that logging is inherently protective because it supports audit and review. In reality, logging can create a second-order governance problem: the records captured to make a run accountable may themselves contain sensitive personal, commercial, legal, or clinical information. If those records are over-detailed, widely accessible, or retained for too long, the governance mechanism becomes a source of risk.

In RAIDT terms, the direct answer is that evidence capture must be governed just as carefully as model use. For example, a team using a generative AI assistant to draft case summaries may wish to retain enough information to reconstruct why a specific output was accepted or challenged. However, storing the full prompt thread with names, identifiers, and attached case notes in a broadly accessible evidence repository would create privacy exposure. A better RAIDT approach is to capture the relevant run metadata, selected excerpts, redacted content, the decision rationale, and the reviewer trail needed for reconstruction.

RAIDT handles this issue better than a generic AI governance approach because it does not stop at a principle such as ?protect data?. It asks what evidence is needed for this run, who needs to inspect it, and what minimum record supports reviewability without unnecessary disclosure. That makes privacy and data protection operational rather than rhetorical.

Practical example in RAIDT terms

Consider a healthcare provider using a generative AI tool to help draft discharge summaries from clinician notes. The use case is valuable because it reduces administrative burden and speeds documentation. The run-level issue, however, is immediate: prompts and outputs may contain patient identifiers, diagnoses, medication details, and clinician annotations.

In RAIDT, the organisation should not assume that the full prompt-output exchange can simply be logged into a standard evidence repository. Instead, it would define the evidence needed for governance readiness: the task description, model version, prompt template class, timestamp, operator role, reviewer decision, flagged risks, and redacted excerpts sufficient to explain why the output was accepted, amended, or rejected. Where full content must be stored, access control and retention rules become part of the governed run record.

The pillars most affected are Responsibility, Auditability, Dependability, and Traceability, with Interpretability also relevant where reviewers must understand how sensitive context shaped the output. The item improves governance readiness because it allows the healthcare provider to reconstruct the run, justify its oversight process, and support audit or challenge without converting the evidence pack into an uncontrolled patient-data archive.

Detailed link to RAIDT

Privacy and data protection link to RAIDT in four ways.

First, they reinforce RAIDT's core idea that governance should be grounded in concrete evidence from specific runs rather than broad claims about systems in general.
Second, they shape what can responsibly be captured at the run level, including how sensitive prompts, outputs, and contextual metadata are minimised and governed.
Third, they determine whether the evidence pack and score profile can be used safely in review, assurance, and escalation processes.
Fourth, they support reviewability, contestability, audit readiness, and organisational learning by ensuring that evidence remains accessible to the right people without becoming unnecessarily exposed to the wrong ones.

Privacy and data protection ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

In other words, this item ensures that RAIDT's move towards evidence does not undermine the very governance quality it is meant to improve.

Link to the five RAIDT pillars

Responsibility

Privacy and data protection strengthen Responsibility by requiring those who design, approve, and operate GenAI use cases to define acceptable evidence practices in advance rather than treating data handling as an afterthought.