Practical guide
What evidence shows human accountability is operating
Human accountability is operating when a named person has real authority, receives the information needed to act, intervenes under defined conditions, and leaves evidence that the intervention changed or affirmed a decision. A policy, role title, approval button, or “human in the loop” claim alone does not demonstrate that control operation.
Human accountability is operating when a named person has real authority, receives the information needed to act, intervenes under defined conditions, and leaves evidence that the intervention changed or affirmed a decision. A policy, role title, approval button, or “human in the loop” claim alone does not demonstrate that control operation.
Define the control objective before collecting evidence
Write the objective as an observable outcome: “For each bounded AI-assisted decision, an authorized owner reviews the required evidence before release, records the disposition and uncertainty, and can pause, change, or reject the output.”
Then state the scope: system or workflow, decision type, affected population, risk tier, jurisdiction, review period, and exclusions. A human who can see an output but cannot understand, challenge, or change it is not evidence of meaningful authority.
NIST’s AI RMF Core calls for clear roles and communication, trained personnel, executive responsibility, differentiated human-AI oversight, documented risk processes, and ongoing review. It also says the framework actions are not a checklist. Those outcomes are useful design references, not certification that a control works. NIST AI RMF Core
Sources: [1]
Test the ACT evidence chain
ACT separates assigned responsibility from demonstrated control operation.
### A — Authority is assigned and usable
Inspect the approved role matrix, delegated limits, conflict rules, training record, and system permissions. The accountable person must know what they own, what evidence they must inspect, which decisions they may make, and when they must escalate.
Test the actual interface and operating conditions. Can the owner pause release, reject the output, request more evidence, correct the record, or switch to a fallback? If the system only records acknowledgement, describe it as acknowledgement rather than approval.
### C — Case evidence connects input to intervention
Sample real cases from the review period, including ordinary decisions, exceptions, overrides, and near misses. For each case, connect:
The strongest evidence shows what the person did, not merely that their account was present. Preserve rejected recommendations and unresolved uncertainty so the record is not biased toward clean approvals.
### T — Testing proves the path works under pressure
Run a designed challenge: missing evidence, conflicting sources, out-of-range recommendation, unavailable reviewer, or urgent deadline. Observe whether the workflow detects the condition, routes it to the authorized person, supports intervention, records the outcome, and invokes fallback when necessary.
GAO’s AI Accountability Framework organizes practices around governance, data, performance, and monitoring and emphasizes accountability processes across the system lifecycle. Use it to structure the test plan and evidence inventory, while keeping any conclusion bounded to the sampled control and period. GAO AI Accountability Framework
- the use-case and risk classification;
- the AI output and relevant source evidence;
- the human reviewer and timestamp;
- the checks actually performed;
- the decision, rationale, uncertainty, and escalation;
- any change made to the output or downstream action;
- the retained artifact and access record.
Sources: [2]
Evidence register
Use one row for every test or sampled case:
| Test ID | Control objective | Case or challenge | Accountable owner | Required evidence | Action observed | System effect | Exception | Disposition | Retest date | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | ACT-01 | authorized review before release | normal / exception / designed failure | named person and role | sources, output, risk route, prior decisions | inspect / question / change / reject / escalate | release / pause / correction / fallback | documented deviation | pass / repair / hold | date and owner |
Link each row to the source artifacts rather than copying sensitive content into the register. Record the retention rule, access boundary, sampling method, and reviewer. A dashboard count can summarize the register but cannot replace it.
Worked example: a procurement recommendation
The following example is synthetic. An AI-assisted workflow ranks three software proposals and recommends Vendor A. Policy requires a procurement owner to verify source completeness, conflicts, and evaluation criteria before the recommendation enters the approval packet.
The normal-case sample shows a reviewer identity and timestamp, but no evidence of source inspection. That is insufficient. During a designed challenge, the test team removes one security questionnaire and changes a cost assumption. The workflow still marks the packet ready, and the reviewer interface offers only “acknowledge.” The procurement owner notices the missing questionnaire in a separate folder, pauses the packet outside the tool, and requests correction.
ACT records mixed evidence: the person exercised judgment and prevented release, but the formal control path failed to surface the condition or provide a pause action. The disposition is repair, not pass. Required remediation adds a completeness gate, an in-product pause, a reason field, a backup approver, and a retest. The result does not establish broader compliance or overall system safety.
Sampling and interpretation rules
Select cases across risk level, team, reviewer, time, decision outcome, and exception type. Include rejected and escalated cases; a sample of approvals alone cannot show that challenge works. Define how missing logs, late approvals, shared accounts, or unavailable source artifacts affect the result.
Use three separate conclusions:
Do not convert a design review into operating evidence, or an operating sample into a universal claim. Report sample size, selection method, exceptions, missingness, reviewer independence, and residual uncertainty.
- Design present: policy, authority, interface, training, and evidence requirements exist.
- Operation observed: sampled cases show the required review and intervention during the stated period.
- Effect supported: the intervention changed, paused, corrected, or deliberately affirmed a decision for a documented reason.
Failure and escalation conditions
Hold or fail the conclusion when responsibility is shared without a final owner, authority exists only on paper, reviewer identity is unreliable, required source evidence is missing, intervention is impossible, overrides have no rationale, exceptions bypass the route, or the fallback has not been tested.
Escalate when the control concerns rights or safety, crosses jurisdictions, handles sensitive data, relies on a conflicted reviewer, or produces legal or regulatory implications. The EU AI Act, for example, contains requirements for human oversight in specified high-risk systems, but applicability and conformity require current system-specific and jurisdiction-specific analysis. This page does not make that determination. EU AI Act, Regulation (EU) 2024/1689
Sources: [3]
Review cadence and limits
Retest after material model, workflow, interface, policy, ownership, or data changes; after incidents and exceptions; and on the control’s scheduled review date. Preserve prior failures and remediation evidence rather than overwriting them with the latest result.
This method is educational guidance. It does not certify legal compliance, control effectiveness, system safety, or accountability across an organization. Two Prune may assess how a participant reviews evidence, uses AI, verifies risk, and communicates a decision in a realistic work simulation. It does not make employment decisions or replace the organization’s accountable owners.
Sources
National Institute of Standards and Technology: AI Risk Management Framework Core
- U.S. Government Accountability Office: Artificial Intelligence Accountability Framework for Federal Agencies and Other Entities
- EUR-Lex: Regulation (EU) 2024/1689 — Artificial Intelligence Act
Sources: [1] [2] [3]
Sources
- 1.NIST AI RMF Core · National Institute of Standards and Technology
- 2.GAO AI Accountability Framework · U.S. Government Accountability Office
- 3.EU AI Act, Regulation (EU) 2024/1689 · eur-lex.europa.eu