Enter your search text in the box above

Select a result to preview

S10.13 - Crisis_and_emergency_response

#raidt/S10

S10.13 ? Crisis and emergency response

flowchart LR
    A[Fast-moving events
Incomplete information
Public impact] --> B[RAIDT
Run-level evidence framework]
    H[Generic AI governance is too abstract for urgent use] --> B
    B --> C[[Crisis and emergency response
High-stakes RAIDT domain playbook]]
    I[Emergency messaging
Incident summaries
Health and cyber alerts
Local authority briefings] --> C
    C --> D[Run-level evidence pack]
    C --> E[Five-pillar score profile]
    C --> F[Reviewer reconstruction
Contestability
Audit readiness]
    D --> G[Governance readiness]
    E --> G
    F --> G

? Star S10 - Empirical Programme, Domains and Sector Playbooks

Star context: Positions crisis and emergency response as a high-stakes domain playbook within RAIDT, showing how run-level evidence becomes especially important when outputs shape urgent decisions, public messaging, and time-critical organisational action.

Academic picture
Definition / background

Crisis examples include public communications and incident summaries. Responsibility, clarity and provenance are critical during fast-moving events.

Within RAIDT, crisis and emergency response refers to the use of generative AI in situations where outputs may influence urgent organisational decisions, public understanding, escalation pathways, or immediate protective action. The concept therefore covers more than disaster response in a narrow sense. It includes any high-pressure context in which time is compressed, consequences are material, and the tolerance for ambiguity, fabrication, or undocumented intervention is low.

This matters conceptually because crisis settings expose the limits of generic AI governance language. A broad statement such as ?human oversight should be maintained? is too vague when a system is drafting an evacuation update, summarising a live cyber incident, or producing a briefing for emergency coordinators. RAIDT places the emphasis on the run as the unit of governance, so that each crisis-related use can be evidenced, reviewed, and scored in relation to the exact task, timing, prompts, source inputs, and checking process.

The item belongs inside RAIDT because crisis response magnifies all five pillars at once. Responsibility is tested because accountability must remain clear under pressure. Auditability and Traceability are tested because later review may need to reconstruct why a message or recommendation was produced. Interpretability matters because users must understand what the model has done and where uncertainty remains. Dependability matters because instability, omission, or hallucination can have immediate operational consequences. In this sense, crisis and emergency response is a demanding application domain through which RAIDT demonstrates why run-level evidence is necessary for governance readiness.

Why this concept matters

Crisis and emergency response solves a practical governance problem: organisations increasingly want AI assistance in urgent workflows, but the very features that make generative AI attractive, such as speed and fluency, can become liabilities when people act on outputs before they are properly checked. This concept helps distinguish low-stakes automation from high-stakes support where provenance, reviewability, and role clarity are non-negotiable.

Without this concept, organisations can easily confuse fast output generation with effective emergency support. They may assume that a plausible draft is operationally safe, or that a human in the loop automatically guarantees responsible use. RAIDT avoids that confusion by asking what evidence exists for the specific run, what controls operated at the point of use, and how the run can be reconstructed if challenged later.

For organisations using GenAI, the concept matters because emergency environments compress decision time but increase the need for scrutiny. A framework that remains only principle-based is weakest precisely where it is most needed. RAIDT makes the concept operational by tying crisis uses to evidence packs, pillar-based scoring, and governance interventions that can be calibrated, tested, and improved across scenarios.

Key idea: Crisis and emergency response matters in RAIDT because urgent AI-assisted outputs must be governable at the level of the individual run, not trusted on the basis of general policy claims.

What this item captures
Practical example / likely audience question

Audience question

Why are crisis and emergency response use cases treated as especially high stakes in RAIDT rather than simply another application area for GenAI?

Answer

The concern behind the question is that many organisations already use drafting tools in routine communications, so crisis drafting can appear to be just a faster version of an existing task. RAIDT treats it differently because the combination of urgency, uncertainty, public impact, and compressed checking time changes the governance burden. In a crisis, a misleading summary, omitted qualification, or unsupported recommendation can shape behaviour before corrections are possible.

The direct answer is that crisis use is high stakes because output quality is not the only issue. What matters equally is whether the organisation can show who initiated the run, what evidence informed it, how uncertainty was handled, what review took place, and why the output was released or rejected. For example, if a local authority uses a model to draft a flood update, the draft may sound authoritative while still misrepresenting affected postcodes or exaggerating the certainty of timing. RAIDT requires that the run be evidenced so that the communication can be checked before release and reconstructed afterwards.

RAIDT handles this better than a generic AI governance approach because it does not stop at broad commitments such as safety, oversight, or transparency. It asks for run-level evidence: the prompt, the source material, the model configuration, the human review, the final decision, and the resulting score profile. That makes the governance claim inspectable rather than merely aspirational.

Practical example in RAIDT terms

A public services resilience team uses a generative AI system to draft an emergency public update after a chemical spill near a residential area. The use case is a time-critical communication that must summarise the incident, state protective advice, and remain aligned with verified operational information.

The run-level issue is that the model may produce a coherent message that blends confirmed facts with inferred details, for example implying that evacuation is mandatory when the current instruction is only to shelter indoors. The evidence needed includes the exact prompt, the verified incident log or source briefing provided to the model, the model version and settings, the time of generation, the human reviewer identity, the edits made before release, and the approval record showing whether the message was published.

The most affected RAIDT pillars are Responsibility, Dependability, and Traceability, with Auditability also critical for post-incident review. This item improves governance readiness because it turns an urgent communication workflow into an evidencable process: reviewers can test whether the system remained within source bounds, whether responsibility for release was clear, and whether the organisation can defend its use of AI if the communication is later challenged.

Detailed link to RAIDT

Crisis and emergency response links to RAIDT in four ways.

First, it expresses RAIDT's core idea that governance must be attached to specific uses of generative AI rather than to the system in the abstract. A model may appear acceptable in general, yet still be unsuitable or insufficiently controlled in a live emergency communication task.

Second, it makes the run central. In crisis settings, what matters is the exact run carried out at a specific time, by a specific actor, for a specific purpose, against specific source material and constraints. RAIDT makes that run visible and assessable.

Third, it connects directly to RAIDT's practical outputs. The evidence pack gathers the artefacts needed to understand and review the crisis run, while the five-pillar score profile helps translate those artefacts into a structured view of governance strengths and weaknesses.

Fourth, it supports reviewability, contestability, audit readiness, and organisational learning. Crisis use is one of the clearest settings in which an organisation may later need to explain, defend, or improve an AI-assisted action. RAIDT provides a disciplined route from event-time usage to retrospective scrutiny and future refinement.

Crisis and emergency response ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

Link to the five RAIDT pillars

Responsibility

Responsibility is central because crisis outputs can trigger action, reassure or alarm the public, and shape inter-agency coordination. RAIDT asks who owned the run, who checked it, and who authorised any downstream use.

Example evidence / implication:

Auditability

Auditability matters because crisis uses are likely to be scrutinised after the event by managers, regulators, investigators, or affected stakeholders. The run must therefore be reconstructable in enough detail to support review.

Example evidence / implication:

Interpretability

Interpretability matters because users need to understand the status of the output: what came from source material, what is model phrasing, and where uncertainty or inference may have entered. In a crisis, unclear interpretation can lead to over-trust.

Example evidence / implication:

Dependability

Dependability is strongly affected because crisis workflows cannot tolerate erratic behaviour, hidden brittleness, or unstable output quality under pressure. Dependability in RAIDT is about whether the run can be relied upon for the task in context, not whether the model performs well in general.

Example evidence / implication:

Traceability

Traceability is critical because organisations may need to show how a public message or internal recommendation emerged from source evidence, model use, human judgement, and final release. This is especially important where accountability may later be contested.

Example evidence / implication:

This item affects all five pillars strongly, but its pressure is especially acute on Responsibility, Dependability, and Traceability because emergency contexts magnify the consequences of ambiguity, instability, and poor ownership.

Why this item is more than a generic concept

In general AI governance, crisis and emergency response may be treated as a high-risk domain requiring caution, human oversight, and escalation protocols. In RAIDT, it means something more operational: each emergency-related use is treated as a governable run that must produce evidence about context, source-grounding, review, and accountability. The RAIDT meaning is therefore more concrete because it does not rely on broad risk labelling alone. It ties crisis use directly to run-level evidence, evidence packs, score profiles, and the organisation's ability to justify and improve practice after the event.

Common misunderstanding

Misunderstanding

If a human reviews the AI output before release, crisis use is already governed adequately.

Correction

Human review is necessary but not sufficient. A reviewer under time pressure can still miss unsupported claims, ambiguous wording, or misleading emphasis, and without run-level evidence the organisation may be unable to show what the reviewer actually saw or why the output was accepted. For example, if an emergency briefing omits a key uncertainty about casualty numbers, saying only that ?a human checked it? does not reveal whether the uncertainty was absent from the source, introduced by the model, or overlooked during review. RAIDT corrects this by combining human oversight with evidence about the run itself.

Boundary and limitation

This item does not prove that a crisis response output is true, safe, or legally defensible in every case. It also does not replace incident command structures, subject-matter expertise, communications policy, or statutory emergency procedures. In highly novel or chaotic events, even well-documented runs may still perform poorly because source information is unstable or incomplete.

RAIDT handles this limitation by treating crisis and emergency response as a governance domain that requires both evidence and escalation. The framework can show how a run was conducted, where controls held, and where weaknesses remain, but it cannot remove the underlying uncertainty of the event itself. Its value is that it makes those uncertainties visible, reviewable, and actionable rather than hidden behind polished output.

Implementation levels

Manual implementation

A researcher or small team can apply this item manually by saving prompts, source notes, outputs, reviewer comments, and release decisions for each crisis-related run. A simple checklist can require confirmation of source provenance, uncertainty wording, human sign-off, and whether the output is internal drafting support or externally releasable text.

Semi-automated implementation

Semi-automated implementation can use structured templates, metadata forms, and workflow prompts that require users to record task type, urgency level, source inputs, reviewer identity, and approval status. Versioned evidence folders and standard scoring rubrics can support more consistent reconstruction and comparison across crisis scenarios.

Fully automated implementation

At scale, a platform or orchestration layer can automatically log prompts, model versions, timestamps, linked source documents, reviewer interventions, release states, and pillar-level scoring cues. Governance dashboards can flag high-risk emergency runs for mandatory review, preserve evidence packs automatically, and feed post-incident lessons into policy updates, playbooks, and retraining.

Practical use in the RAIDT project

This item is useful across the RAIDT project because it provides a concrete, high-stakes domain in which the framework's claims can be tested. In Paper 08 Foundations, it helps explain why run-level evidence is needed when principle-based governance is too abstract for urgent organisational action. In Paper 09 Empirical Validation, it offers a demanding scenario family for comparing configurations, repeated runs, and pillar-level outcomes under pressure.

In Paper 10 Policy Pathways, crisis and emergency response is especially valuable because it speaks directly to public accountability, institutional defensibility, and operational assurance. Within sector playbooks, it helps connect RAIDT to healthcare incidents, environmental warnings, cyber incident management, and public-service emergency communications. It is also a strong item for viva defence and supervisor explanation because it makes the practical stakes of RAIDT immediately legible: the framework is not only about documenting AI use, but about making urgent use contestable, reviewable, and improvable.

Key audience questions to prepare for

Q1. Why is crisis response a particularly useful test case for RAIDT?

Because it concentrates the exact governance pressures RAIDT is designed to address: time pressure, uncertainty, accountability, public impact, and the need for retrospective review. If RAIDT can structure evidence in this domain, it demonstrates value beyond routine low-stakes use.

Q2. Does RAIDT assume GenAI should be trusted in emergencies?

No. RAIDT does not presume trust; it creates conditions for conditional, evidencable, reviewable use. The framework is designed to show when a run is insufficiently controlled as much as when it is acceptable.

Q3. What kind of evidence is most important in an emergency communication run?

The most important evidence usually includes the source material used, the exact prompt, the generated output, the reviewer's edits, the approval decision, and timestamps. Together, these show whether the output was grounded, checked, and released responsibly.

Q4. How does this differ from ordinary communications governance?

Ordinary communications governance often focuses on policy, approval, and messaging standards. RAIDT adds run-level reconstruction of AI involvement, which is crucial when generated wording may have influenced meaning, confidence, or timing under pressure.

Q5. What does good governance readiness look like for this item?

Good governance readiness means the organisation can show that crisis-related AI runs are bounded by clear roles, grounded in verified sources, reviewable after the fact, and comparable across scenarios using evidence packs and pillar-based scoring.

Suggested citation concepts to support this item
Short explanation for presentation

Crisis and emergency response is a strong RAIDT domain because it shows why generic AI principles are not enough for organisational governance. In urgent situations, generative AI may draft public updates, incident summaries, or internal briefings, but the key question is not only whether the output sounds plausible. The key question is whether the organisation can evidence how that specific run was conducted, what sources informed it, who reviewed it, what uncertainties remained, and why the output was accepted or rejected. RAIDT addresses this by treating the run as the unit of governance, producing an evidence pack and a five-pillar score profile. That makes emergency use more reviewable, contestable, auditable, and ultimately more useful for continuous improvement across sector playbooks.

One-line takeaway

Crisis and emergency response is a high-stakes RAIDT domain because urgent GenAI use must be governed through run-level evidence rather than trusted through policy assertions alone.

Related items in empirical programme, domains and sector playbooks
Anchored questions
Powered by Forestry.md