AI Standard: Transparency & Explainability
1. Purpose
Section titled “1. Purpose”Ensure AI systems can be understood, evaluated, and contested at levels appropriate to risk, while preventing deceptive or obscured operation.
2. Applicability
Section titled “2. Applicability”- Applies to AI systems used in Tier 1–3 contexts; Tier 0 is encouraged as feasible.
- Strongest requirements apply to high-impact decisions and Tier 2–3 deployments.
3. Ethical Mapping
Section titled “3. Ethical Mapping”A3 Justice: contestability and meaningful explanationA4 Trustworthiness: truthful capability/limitation representationA2 Dignity: transparency supports consent and agency
4. Requirements (Normative)
Section titled “4. Requirements (Normative)”AI-T-1 (Disclosure Packet). For Tier 1–3 deployments, operators MUST maintain an AI system disclosure packet including:
- intended purpose and prohibited uses
- system boundaries and dependencies (data, models, services)
- known limitations and uncertainty characterization
- training / fine-tuning data provenance at an appropriate granularity
- evaluation results relevant to the deployment context
AI-T-2 (Affect Notice). When an AI system materially influences an affected party, the operator MUST provide timely notice that AI influence occurred, unless doing so would create a demonstrable safety risk that is documented and mitigated.
AI-T-3 (Explanation Standard by Tier).
- Tier 1: operator SHOULD provide an explanation suited to operational debugging and user understanding.
- Tier 2: operator MUST provide an explanation sufficient for independent review and contestation by affected parties (directly or via representatives), including key factors and uncertainty.
- Tier 3: operator MUST provide an explanation and evidence package sufficient for third-party audit, including causal analysis where feasible and comprehensive uncertainty treatment.
AI-T-4 (Explainability Debt Register). Tier 2–3 operators MUST maintain an “explainability debt” register that:
- identifies components with low interpretability
- documents compensating controls (testing, constraints, human oversight)
- defines a time-bound plan to reduce debt or justify permanence
AI-T-5 (Anti-Deception). Systems MUST NOT intentionally misrepresent whether outputs are AI-generated, or the level of confidence/uncertainty, in ways that materially affect decisions.
5. Compliance Evidence
Section titled “5. Compliance Evidence”- disclosure packet artifacts and version history
- affected-party notice templates and delivery logs
- explanation reports and appeal outcomes (Tier 2–3)
- explainability debt register with mitigation tracking
- product/UI audits for deception patterns
6. Rationale (Non-normative)
Section titled “6. Rationale (Non-normative)”Transparency is not a single feature; it is a bundle of artifacts and practices enabling accountability, remedy, and safe operation. Explainability debt makes “we can’t explain it” a managed, measurable risk instead of an excuse.
7. Failure Modes & Abuse Cases
Section titled “7. Failure Modes & Abuse Cases”- hidden model updates changing behavior without notice
- “confidence theater” (fake certainty cues)
- explanations that are technically correct but practically unusable by affected parties
8. Change Log
Section titled “8. Change Log”- v0.1: Initial draft (filename:
02_ai_standards/transparency_and_explainability.md).
Traceability Table (Requirement → Axiom → Evidence)
Section titled “Traceability Table (Requirement → Axiom → Evidence)”| Requirement / Control | Axiom(s) | Evidence Artifacts |
|---|---|---|
| AI-T-1 | A4 | disclosure packet artifacts and version history |
| AI-T-2 | A3, A2 | affected-party notice templates and delivery logs |
| AI-T-3 | A3 | explanation reports and appeal outcomes (Tier 2–3) |
| AI-T-4 | A4 | explainability debt register with mitigation tracking |
| AI-T-5 | A4, A2 | product/UI audits for deception patterns |