The AI audit trail an auditor will actually read

An AI audit trail is only evidence if a non-engineer can read it. What NUADU records when AI assists a decision: approver, timestamp, override and its reason.

An AI audit trail is the record of what an AI system produced and what a named person decided with it, and it is only evidence if a non-engineer can read it. When AI assists a marking decision on NUADU, the record is small and specific: the result and whether it is signed, the approver by role and name, the timestamp of the review, and any override with its criterion, the score before and after, and a reason. This post covers those fields, the five logged events, the EU AI Act's logging duty and the readers.

What should an AI audit trail record?

The decision, with a name on it. Logs organised around system events show what a model did; a trail an auditor can use is organised around decisions, because the decision is what an appeal contests and what a QA office reviews.

That is the difference between AI audit logs and an AI audit trail. Logs record events for the people who run the system. An audit trail records decisions for the people who have to account for them: who approved a result, when, and what they changed on the way.

NUADU's homepage shows the shape in a sample decision record. It is an illustration of the interface, not a customer's data, and its values are evidence of nothing. The four fields are what matter:

  • Result — the outcome, and whether it has been signed.
  • Approver — the role, and the named educator who signed.
  • Reviewed — the timestamp of the review.
  • Override — if the marker changed a draft score: which criterion, the score before and after, and a reason.

The last field is the one an auditor reads first. An override without a reason is a decision nobody can later explain. An override with its criterion, its change and its reason can be defended on appeal and reviewed by QA. The rule underneath is the one the platform repeats at every stage: AI prepares the evidence; the educator decides, approves and signs off; every output is logged; there are no automated decisions about a named learner.

approver named · timestamp written · override recorded

Which events does an AI audit trail log?

On NUADU, five for one marking session, in the order the work happens. Each stage flows into the next, and at each one the AI's part is to prepare evidence for the human step that follows.

  • exam_submitted — marking starts: your rubric applied across the cohort.
  • draft_marks_ready — the educator reviews: borderline cases surface first.
  • marks_signed_off — feedback released: rubric-anchored, marker-approved.
  • results_released — analytics update: cohort patterns, programme indicators.
  • session_closed — the audit trail closes: every output, override and approval logged.

Read in order, the events answer an appeal in order: the rubric the candidate was marked against, the named educator who reviewed the draft marks, any override and its reason, the moment the marks became final, and what the institution learned afterwards. What it learned is cohort patterns and programme indicators, never a label on an individual.

Beside the per-decision record sits an aggregate view. Agreement rates, override rates and drift indicators are visible to QA, and they are the evidence a second-marking policy is adjusted on. The record explains one grade. The aggregate shows whether the marking as a whole is holding.

What does the EU AI Act require you to log?

For high-risk systems, the capability to record events automatically over the system's lifetime. Article 12 of the EU AI Act requires high-risk AI systems to technically allow the automatic recording of events (logs) over their lifetime, and assessment is in scope: Annex III, point 3(b) lists AI systems intended to evaluate learning outcomes among the high-risk uses.

Who carries that duty in practice is set out in NUADU's AI Transparency Notice. In an institutional deployment the notice splits the work: record-keeping and logging under Article 12 is shared by NUADU and the institution, while human oversight in operation, under Article 14, belongs to the institution, “using the review, override and sign-off capability NUADU provides”.

The notice calls the controls behind the trail — logging, review, override, sign-off — “EU AI Act-ready”, and is explicit that the label is not a certification claim. Which uses in education the EU AI Act treats as high-risk, and what else follows from that, is a post of its own.

Who reads an AI audit trail, and for what?

Four readers, one set of fields, four different questions.

  • The candidate's appeal reads the rubric-anchored feedback, the marker's notes and the criterion-by-criterion result, under the institution's own appeals process.
  • The QA office reads every AI output, override and approval, logged, exportable and reviewable, beside the agreement and override rates it already sees.
  • The regulator or board traces a decision to a named human, with the DPIA template, DPA template and sub-processor list from the documentation pack beside it.
  • The works council and the DPO, in enterprise deployments, read audit trails, configurable retention and exportable evidence packages, with a pack that says per AI module what it does, what data it uses, who decides and who can contest.

Nothing in the trail is a verdict. It is the material an exam appeals process reasons from. It is not a prediction about the candidate either; override rates are visible to QA as a quality view, and the marker's reason travels with every override.

The record also stops where NUADU's pages stop. Retention is configurable; where the trail is stored, in what format and for how long are questions for the DPA and the documentation pack, both on request. The audit trail behind exam results reporting and the AI in assessment page show where the record sits in the lifecycle.

Ask a vendor to show one decision record that a non-engineer can read. If it cannot, the logging claim is a claim about storage, not about evidence.

Frequently asked questions

What is an AI audit trail?

A record of each AI output and of the human decision taken with it: inputs, outputs, overrides, approvals and decisions. On NUADU it is written for the people who have to read it — the QA office, the DPO, the candidate's appeal and the external auditor.

What is the purpose of audit trails in AI systems?

To let someone outside engineering answer, later, how a decision was reached and who made it. An appeal, an audit, an accreditation panel and the EU AI Act all put that question, and the trail is where the answer is kept.

What does an audit trail check for?

For an AI-assisted result: whether the decision traces to a named human, and whether every change is explained. That means the draft, the reviewer, any override with its reason, and the moment the result was signed.

→ Talk to us about your assessment workflow (30 min)