This is checklist is for ministries, regulators, hospitals, councils, universities, platforms, or large organizations assessing their own AI use or an incident involving a vendor system.

Purpose:
To determine whether an AI-related incident reflects:

  • a localized operational error, or
  • a systemic governance failure requiring structural intervention.

Each item is Yes / No / Unknown.
“Unknown” is a finding: it signals loss of institutional visibility.

I. Authority & Mandate

  1. Is there a documented mandate specifying what epistemic role this system is allowed to play (e.g. advisory vs authoritative)?
    ☐ Yes ☐ No ☐ Unknown
  2. Are there explicit boundaries defining where the system must not provide conclusive outputs (e.g. health, legal, eligibility decisions)?
    ☐ Yes ☐ No ☐ Unknown
  3. Is a specific institutional role or unit accountable for epistemic harm caused by this system’s outputs?
    ☐ Yes ☐ No ☐ Unknown
  4. Do staff understand whether they are expected to defer to, challenge, or merely consider the system’s outputs?
    ☐ Yes ☐ No ☐ Unknown

II. Interface & Use Context

  1. Does the system present single synthesized outputs by default in high-stakes contexts?
    ☐ Yes ☐ No ☐ Unknown
  2. Are uncertainty, disagreement, or evidence strength visible to users at the point of decision?
    ☐ Yes ☐ No ☐ Unknown
  3. Does the interface nudge users toward acceptance (speed, tone, placement) rather than reflection?
    ☐ Yes ☐ No ☐ Unknown
  4. Have interface design choices been reviewed as governance decisions, not just UX choices?
    ☐ Yes ☐ No ☐ Unknown

III. Compression & Information Loss

  1. Is the system used to summarize, triage, or condense complex case material?
    ☐ Yes ☐ No ☐ Unknown
  2. Are there safeguards to prevent loss of urgency, severity, or minority cases during summarization?
    ☐ Yes ☐ No ☐ Unknown
  3. Is there a documented process for users to inspect what was omitted or downweighted?
    ☐ Yes ☐ No ☐ Unknown
  4. Have false reassurance or premature closure risks been explicitly assessed?
    ☐ Yes ☐ No ☐ Unknown

IV. Bias & Differential Impact

  1. Has the system been tested for differential effects across populations relevant to its use (e.g. gender, age, vulnerability)?
    ☐ Yes ☐ No ☐ Unknown
  2. Are bias assessments model-specific and task-specific (not generic assurances)?
    ☐ Yes ☐ No ☐ Unknown
  3. Is bias monitoring ongoing rather than one-time?
    ☐ Yes ☐ No ☐ Unknown
  4. Is there a mechanism to surface harms that arise through omission or framing, not just explicit error?
    ☐ Yes ☐ No ☐ Unknown

V. Institutional Comprehension & Control

  1. Can the institution clearly identify the model(s) in use, including versioning and update cadence?
    ☐ Yes ☐ No ☐ Unknown
  2. Is there internal capacity to audit or independently evaluate system behavior?
    ☐ Yes ☐ No ☐ Unknown
  3. Are changes in model behavior detectable by users or oversight staff?
    ☐ Yes ☐ No ☐ Unknown
  4. Was reliance on the system introduced before governance and oversight structures were in place?
    ☐ Yes ☐ No ☐ Unknown

VI. Governance & Enforcement

  1. Are safeguards enforceable at the same speed and scale as deployment?
    ☐ Yes ☐ No ☐ Unknown
  2. Do public statements or policies correspond to operational constraints?
    ☐ Yes ☐ No ☐ Unknown
  3. Is there a defined escalation pathway when harm is identified (pause, rollback, containment)?
    ☐ Yes ☐ No ☐ Unknown
  4. Does delay in review or regulation effectively allow continued operation?
    ☐ Yes ☐ No ☐ Unknown

VII. Scale, Speed, and Irreversibility

  1. Does the system operate or propagate at a scale that exceeds the institution’s corrective capacity?
    ☐ Yes ☐ No ☐ Unknown
  2. Are retraction, correction, or containment mechanisms slower than dissemination?
    ☐ Yes ☐ No ☐ Unknown
  3. Have downstream effects persisted after partial fixes or public acknowledgment?
    ☐ Yes ☐ No ☐ Unknown
  4. Is full reversal of harm realistically achievable?
    ☐ Yes ☐ No ☐ Unknown

Interpretation Guide (Internal Use)

  • Multiple “Unknown” responses → loss of institutional visibility
  • Clusters of “Yes” in I–II → authority/interface risk
  • Clusters in III–IV → harm and equity risk
  • Clusters in V–VII → systemic governance failure

A critical signal is misalignment:

  • High reliance + low comprehension
  • Strong statements + weak enforcement
  • Large scale + slow correction

What this checklist is designed to do

  • Force explicit answers where ambiguity is often tolerated
  • Distinguish operational risk from structural governance failure
  • Create an internal record of what the institution knew, when

It is not a compliance box-ticker.
It is a visibility and accountability instrument.