Systematic independent examination against criteria; algorithmic auditing inherits and strains this.
In algorithmic accountability journalism and media research, an audit is an independent empirical investigation of a system's behavior conducted from the outside, usually without the operator's cooperation or consent: reporters and researchers probe recommenders, ad delivery, or generative models with controlled inputs, scraped outputs, or sock-puppet accounts to detect skew, suppression, or harm. Authority comes from method transparency and public interest, not from engagement letters or agreed criteria; publication, not a management letter, is the deliverable, and platform pushback is expected.
In practice: Design a methodologically defensible probe of a platform or model, document the method for public scrutiny, and report behavioral evidence of skew or harm without relying on operator access.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
For bank internal audit functions, an audit is a periodic, independent, evidence-based assessment of whether the model risk management framework operates as designed: the model inventory is complete, validations were performed by qualified independent staff at the required frequency, findings were remediated on schedule, and documentation supports supervisory review. Under SR 11-7 internal audit does not re-validate models; as the third line of defense it assures the process around them, and its findings are reportable to the board and available to examiners.
In practice: Plan and execute audits of the model risk management framework, test inventory completeness and validation quality, track remediation of findings, and report assurance conclusions to the board and examiners.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
In fair-lending and conduct analytics teams, an audit is a statistical testing exercise on decision outcomes: approval, pricing and default data are disaggregated by protected class or its proxies, disparities are estimated with regression controls for legitimate credit factors, and residual gaps are sized against materiality thresholds that anticipate examiner methodology. The audit object is the outcome distribution, not the governance process; a finding is a quantified, unexplained disparity that survives controls and demands business justification or remediation.
In practice: Assemble outcome data with protected-class estimates, run controlled disparity tests that replicate examiner methodology, and quantify residual gaps requiring business justification or remediation.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
In hospital quality practice, audit denotes the clinical-audit cycle extended to algorithms: care or algorithm performance is measured against explicit, locally agreed standards, gaps are fed back to the service, changes are made, and measurement repeats. For a deployed prediction model this means periodic re-checks of accuracy, calibration and alert burden on the institution's own patients, because vendor-reported performance is not trusted to transfer across sites. An audit finding is actionable when it shows deviation from the standard large enough to affect patient safety or clinical workflow.
In practice: Define local performance standards for a deployed model, schedule recurring measurement on the institution's own patient data, and escalate deviations that threaten patient safety or overload clinicians.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
In medical-device quality and regulatory affairs, an audit is a certification-oriented examination of the quality management system behind an AI-enabled device: notified bodies and certification auditors check that development, data management, validation, post-market surveillance and change control follow documented procedures required by the MDR and, for high-risk AI, the AI Act's quality-management and conformity-assessment provisions. The audit's product is certification or nonconformities to close; it examines the paper trail and process discipline more than model outputs themselves.
In practice: Maintain audit-ready quality-system documentation for an AI-enabled device, host certification and surveillance audits, and close nonconformities on schedule to keep the device on the market.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
For supreme audit institutions and government internal auditors, an audit of an algorithm is a formal examination, under a statutory mandate, of whether an automated system used in administration complies with law, established norms and stated policy objectives: legal basis, procurement rules, data protection, effectiveness and non-discrimination. Findings are addressed to the responsible minister or agency and to parliament, and the audited body must respond. Independence is institutional, and criteria must be traceable to law or published frameworks because conclusions must survive political and judicial scrutiny.
In practice: Derive audit criteria from statute and published frameworks, examine deployed government algorithms against them, and report findings that ministers and agencies are formally obliged to answer.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
Communities disagree over what qualifies an examination as an 'audit'. The assurance tradition (internal audit, supreme audit institutions, certification bodies) reserves the term for independent examinations against defined criteria under a mandate, with auditee cooperation, professional standards and formal reporting duties. The investigative tradition in journalism and research applies it to adversarial external probes of system behavior conducted without access, agreed criteria, or the operator's consent, grounding authority in method transparency and public interest instead.