Skip to content

Why AI audit logs do not capture human judgement

Jump to bottom
Arsalan Shahid edited this page Aug 25, 2026 · 1 revision

Operational logs usually show that a person acted. Accountable decision evidence must also show what they reviewed, what they changed, why they changed it and under whose authority.

Short answer

Most AI audit logs capture system events: a model ran, a tool was called, a screen was opened, an approval button was clicked, or a record was updated. These facts are useful, but they are not a complete representation of human judgement.

Judgement is relational. It exists between a proposal, a person, a decision context and a resulting action. If the log retains only the final action, it loses the difference between passive confirmation, expert correction, risk-based rejection, inability to decide and escalation to another authority.

An audit log can be designed to capture those distinctions. CHAP provides a standard vocabulary and evidence model for doing so across applications.

An event is not yet a decision record

Consider this event:

{
  "timestamp": "2026-08-25T10:42:18Z",
  "user": "reviewer-4821",
  "item": "draft-991",
  "action": "approved"
}

It proves that an application recorded an approval action. It does not reveal whether the reviewer saw the same draft that is now stored, whether they changed it before approving, whether they were qualified to decide, or whether the application later rewrote the item.

A useful human-judgement record connects at least five elements:

  1. Proposal: the exact artefact or action presented for review.
  2. Reviewer: the participant and authority under which they acted.
  3. Decision: approve, reject, override, abstain or escalate.
  4. Reasoning evidence: rationale, tags, policy references and relevant citations.
  5. Outcome: the accepted artefact and any subsequent execution result.

The purpose is not to capture a person's private chain of thought. It is to retain the explicit evidence needed to understand and govern the decision.

What ordinary logs tend to lose

The original proposal

Many systems overwrite the agent's output with the human-edited version. The final record may be better, but the intervention disappears. Without both versions or a structured diff, there is no reliable way to measure where human expertise changed the result.

The meaning of approval

An approved Boolean often combines several cases:

  • The person accepted the proposal unchanged.
  • The person corrected minor details and accepted it.
  • The person replaced the core recommendation.
  • The person approved only because another system had already constrained the action.

These cases carry different accountability and learning value.

A reason for rejection

Rejected outputs are often logged as failures or abandoned runs. That discards the expert signal: unsupported evidence, policy conflict, unsafe consequence, missing information or a simple factual error.

Abstention and escalation

Good human oversight includes knowing when not to decide. A reviewer may lack subject knowledge, face a conflict of interest or need a higher authority. If the data model only permits approve or reject, it misrepresents responsible abstention as delay or failure.

Authority and context

The same click has different meaning in shadow, trial and production operation. It also differs when made by a trainee, domain specialist, system owner or emergency delegate. Identity without role and policy context is incomplete.

Evidence integrity

A database row can be mutable even when access is restricted. For high-consequence work, a verifier may need to detect deletion, reordering or alteration of decision events independently of the application that created them.

How CHAP represents judgement

CHAP links the proposal and decision through a task and artefact lifecycle.

  • review.request identifies the item being evaluated and the intended reviewer.
  • decide.approve records acceptance as-is.
  • decide.reject records refusal and can carry a reason.
  • decide.override records an accepted modification, including a structured diff, rationale, tags and whether the original intent was preserved.
  • abstain.declare records a deliberate inability or refusal to decide.
  • escalate.raise records the need for another authority or process.

Signed envelopes extend an append-only hash chain. Audit operations can read and verify that sequence. Tool calls and external agent exchanges can be cited without forcing every external payload onto the CHAP wire.

Example: a clinical summary

An agent drafts a clinical discharge summary. A clinician corrects a medication dose, removes an unsupported diagnosis and approves the revised document.

A typical application log may show document_updated followed by approved. A better decision record shows:

  • The agent-produced version and model identity.
  • The dose correction as a field-level modification.
  • The removed diagnosis and its reason.
  • The clinician's role and authority.
  • Whether the document was in trial or production mode.
  • The final approved artefact.

This record can support review of the individual case, evaluation of recurring model failure modes and verification of the oversight process. It still does not prove that the clinician's judgement was medically correct.

Human judgement is valuable data, not free training data

Structured overrides can reveal recurring patterns that generic logs hide. Tags such as unsupported-claim, policy-conflict or missing-context can make expert interventions measurable.

That does not mean every override should be sent directly into model training. The data may contain personal information, confidential business context, biased local practice or decisions made under exceptional conditions. Use a governed process for consent, access, quality review, de-identification and evaluation before treating it as a learning signal.

Privacy and proportionality

More evidence is not always better. Rationale fields can expose sensitive information, and immutable records can conflict with retention or deletion obligations if payloads are copied carelessly.

A proportionate design can store content in an external controlled system and place an opaque reference or cryptographic hash in CHAP. Retention rules, redaction and access control remain deployment responsibilities. Record enough to understand the accountable event, not every thought or keystroke.

What CHAP does not prove

A valid evidence chain can show that a declared participant sent a message in a declared sequence. Depending on the identity and signing profile, it can provide stronger attribution. It cannot by itself prove that:

  • The proposal was factually correct.
  • The reviewer was attentive or competent.
  • The policy was lawful or fair.
  • The interface presented the information without manipulation.
  • The final action matched the approved artefact.
  • The wider AI system is compliant with a regulation.

These require organisational controls, testing, monitoring and domain-specific assurance. CHAP makes the decision process more inspectable; it does not replace judgement about the process.

Related pages

Primary sources

If you have a real audit event that fails to express a human intervention, star CHAP and share a redacted example in Discussions.

Last reviewed: 25 August 2026.

Clone this wiki locally