Examples
Worked examples
- Is an instance
An external auditor producing a SMACTR-compliant audit report for a hiring-screening model.
- Is an instance
A regulatory pre-market conformity assessment for an EU AI Act high-risk system.
Counter-examples
Looks similar, but isn't
- Not an instance
A model card written by the developing team (self-reporting, not an audit).
- Not an instance
A code review of training scripts.
Editorial commentary
A model audit is a structured, independent assessment of a machine-learning model against pre-specified criteria — typically performance, robustness, fairness, security, privacy, and conformance with a stated policy or regulatory requirement. Audits range from an internal pre-deployment review by a separate team within the same organisation, to a formal third-party audit commissioned for regulatory or contractual purposes.
What distinguishes an audit from a card or an evaluation
An AI evaluation card or a model card is documentation the model’s own developer publishes describing what was measured and found — a disclosure, not an independent check. A model audit is the independent-verification step: someone other than the model’s developer checks the model, or checks the developer’s own claims, against defined criteria, and the audit’s findings become the artefact of record. A bias audit is a model audit scoped specifically to fairness/bias criteria rather than the full set.
Regulatory context
Under the EU AI Act, general-purpose AI models presumed to carry systemic risk (cumulative training compute above 10^25 FLOPs, per Articles 51 and 55) face model evaluation and adversarial-testing duties that function as a mandated audit regime; most models never reach that threshold and remain subject only to whatever contractual or institutional audit an adopting organisation chooses to require.
Why it matters for research administration
Before an institution deploys a third-party model for a consequential use — admissions screening, grant triage, clinical decision support — procurement and research-integrity offices increasingly ask for audit evidence, not just the vendor’s own model card, as part of due diligence.
References
- See also: AI assurance, Bias audit (model), EU AI Act.
Also known as
AI model audit
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Model audit"
vocab-term-identifier="https://casrai.org/dictionary/term/model-audit" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/model-audit",
"name": "Model audit",
"identifier": "https://casrai.org/dictionary/term/model-audit",
"description": "A structured assessment of a machine-learning model by an independent party against pre-specified criteria covering performance, robustness, fairness, security, privacy, and conformance with stated policy.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/model-audit",
"sameAs": [
"AI model audit"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:50",
"dateModified": "2026-08-22T15:43:53",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
