Auditing the Black Box: Rethinking Audit Assurance and Professional Liability in the Age of Artificial Intelligence

May combine two claims.

Krieger, Drews, and Velte found that adoption within firms was uneven and that the obstacles were organizational and epistemic rather than technical, with engagement teams uncertain how regulators and internal quality reviewers would treat conclusions resting on analytical output.

“Krieger, Drews, and Velte (2021) found that adoption within firms was uneven and that the obstacles were organizational and epistemic rather than technical, with engagement teams uncertain how regulators and internal quality reviewers would treat conclusions resting on analytical output.”

• cites Krieger, Drews, and Velte (2021) sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

In forensic work, interpretability is necessary for testimony to be useful, not merely a matter of style, which supports Rudin's (2019) case for interpretable models in high-stakes settings.

“Interpretability is therefore not a stylistic preference in forensic work but a condition of testimonial usefulness, which is an applied vindication of Rudin's (2019) argument for interpretable models in high-stakes settings.”

• cites Rudin's (2019) sourceSupported

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Deterministic tools execute human instructions and raise no novel questions.

“Deterministic tools execute human instructions and raise no novel questions.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Neither banning opaque systems outright nor relying on them without challenge is defensible.

“the article argues that neither wholesale prohibition of opaque systems nor uncritical reliance on them is defensible.”

May combine two claims.

Artificial intelligence has shifted from the edge of audit practice to its operational center, while the profession's conceptual framework has lagged behind.

“Artificial intelligence has moved from the periphery of audit practice to its operational core, and the profession's conceptual apparatus has not kept pace.”

Probably not a substantial claim: may not pass S4 (based).

The indeterminacy in responsibility allocation falls hardest on jurisdictions, including many in Africa, whose regulators and courts have the least institutional capacity to resolve it case by case.

“this indeterminacy falls hardest on jurisdictions, including many in Africa, whose regulators and courts have the least institutional capacity to resolve it case by case.”

• AssumptionRests on an assumption: Regulators and courts in many African jurisdictions have the least institutional capacity to resolve responsibility allocation case by case.

May combine two claims.

An audit memorandum drafted by a generative system may contain citations to workpapers that do not exist, and unless the reviewing auditor checks every reference, the documentation trail relied on by liability and inspection regimes becomes unreliable at its source.

“An audit memorandum drafted by a generative system may contain citations to workpapers that do not exist, and unless the reviewing auditor checks every reference, the documentation trail that liability regimes and inspection regimes both depend upon becomes unreliable at its source.”

• AssumptionRests on an assumption: The documentation trail is relied upon by liability regimes and inspection regimes.

Probably not a substantial claim: may not pass S3 (load-bearing).

Appelbaum, Kogan, and Vasarhelyi noted that analytical techniques entering the audit had begun to outrun the evidential concepts available to describe them.

“Appelbaum, Kogan, and Vasarhelyi (2017) reached a similar conclusion from the direction of research design, noting that the analytical techniques entering the audit had begun to outrun the evidential concepts available to describe them.”

• cites Appelbaum, Kogan, and Vasarhelyi (2017) sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Burrell identified three kinds of algorithmic opacity: corporate secrecy, technical illiteracy, and opacity inherent in the scale and structure of machine learning.

“Burrell (2016) distinguished three forms of algorithmic opacity: opacity as corporate secrecy, opacity as technical illiteracy, and opacity intrinsic to the scale and structure of machine learning itself.”

• cites Burrell (2016) sourceNot checked — no full text of the source

May combine two claims.

A 2024 PCAOB survey of the largest audit firms found generative AI already in audit workflows, first in administrative and research tasks, with stated ambitions for risk assessment, scoping, and disclosure evaluation.

“When the Public Company Accounting Oversight Board surveyed the largest audit firms in 2024, it found that generative artificial intelligence had already entered the audit workflow, initially in administrative and research tasks but with declared ambitions extending to risk assessment, scoping, and the evaluation of financial statement disclosures (PCAOB, 2024).”

• cites PCAOB, 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The reasonable auditor of the coming decade is neither a technophobe who refuses algorithmic assistance nor an enthusiast who accepts machine output without challenge.

“The reasonable auditor of the coming decade is neither the technophobe who refuses algorithmic assistance nor the enthusiast who accepts machine output without challenge”

Probably not a substantial claim: may not pass S4 (based).

A classification model may err, but it errs within the space of its defined outputs.

“A classification model may err, but it errs within the space of its defined outputs.”

May combine two claims.

Both the assurance model and the liability regime presuppose a human evaluator whose reasoning can be examined, documented, and litigated.

“Both structures presuppose a human evaluator whose reasoning can be examined, documented, and, when necessary, litigated.”

May combine two claims.

PCAOB 2024 outreach confirmed both acceleration and caution: firms reported investment in generative tools alongside anxiety about auditability of underlying data and the propensity of such tools to produce false or misleading content.

“The PCAOB's 2024 outreach confirmed both the acceleration and the caution: firms reported investment in generative tools alongside explicit anxiety about the auditability of the underlying data and the propensity of such tools to produce false or misleading content (PCAOB, 2024).”

• cites PCAOB, 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Auditors increasingly use opaque systems as audit tools and also encounter opaque systems as objects of audit inside client estimation processes, credit models, and automated controls.

“Auditors increasingly use opaque systems as tools of the audit, and they increasingly confront opaque systems as objects of the audit, embedded in the client's own estimation processes, credit models, and automated controls.”

Classified as probably not a substantial claim.

In most common law jurisdictions, the professional liability regime measures auditor conduct against the care a reasonably competent member of the profession would have exercised.

“The second is the professional liability regime, which in most common law jurisdictions measures the auditor's conduct against the care that a reasonably competent member of the profession would have exercised in the circumstances.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The negligence standard in auditor liability needs reformulation, not abandonment.

“Second, the negligence standard in auditor liability requires reformulation, not abandonment.”

May combine two claims.

Neither the assurance model nor the liability regime was designed for a professional who delegates parts of reasoning to a statistical artefact.

“Neither was designed for a professional who delegates portions of that reasoning to a statistical artefact.”

Classified as probably not a substantial claim.

The common law of auditor negligence was built for human error, and its landmarks reflect that heritage.

“The common law of auditor negligence was built for human error, and its landmarks reflect that heritage.”

Probably not a substantial claim: may not pass S4 (based).

Skepticism toward a management representation has a known grammar: corroborate, test the underlying records, and consider incentive and opportunity.

“Skepticism toward a management representation has a known grammar: corroborate, test the underlying records, consider incentive and opportunity.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The assurance model does not need new foundations to accommodate the insight about differential model reliability, but it does need that insight stated authoritatively so that an engagement partner documenting differential reliance across the risk spectrum is following guidance rather than improvising.

“The assurance model does not need new foundations to accommodate this insight, but it does need the insight stated authoritatively, so that the engagement partner who documents differential reliance across the risk spectrum is following guidance rather than improvising.”

May combine two claims.

A sufficiency judgment that treats a model's aggregate accuracy as if it applied uniformly to every item the model touches commits a statistical error with legal consequences, because edge cases generate restatements and lawsuits.

“A sufficiency judgment that treats a model's aggregate accuracy as if it applied uniformly to every item the model touches commits a statistical error with legal consequences, because it is precisely the edge cases that generate restatements and lawsuits.”

• AssumptionRests on an assumption: A model's reliability is not uniform; it is strong in the center of its training experience and weak at the edges.
• AssumptionRests on an assumption: The audit transactions that matter most—unusual, period-end, and management-initiated items—are disproportionately located at the edges of a model's training experience.
• AssumptionRests on an assumption: Edge cases are what generate restatements and lawsuits.

May combine two claims.

Commerford, Dennis, Joe, and Ulla demonstrated experimentally that auditors discount contradictory evidence more readily when it comes from an AI system than from a human specialist, a phenomenon situated in the algorithm aversion literature.

“Commerford, Dennis, Joe, and Ulla (2022) demonstrated experimentally that auditors discount contradictory evidence more readily when it comes from an artificial intelligence system than when it comes from a human specialist, a phenomenon the authors situated within the broader literature on algorithm aversion.”

• cites Commerford, Dennis, Joe, and Ulla (2022) sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

The competent hybrid auditor selects tools whose validation and limitations she understands at the level of an informed professional user, not a developer.

“She selects tools whose validation and limitations she understands at the level of an informed professional user, not a developer.”

May combine two claims.

None of these model-skepticism procedures is beyond the profession's competence, but none is currently required, described, or standardized, and quality review practices consequently vary in ways inspection findings are beginning to expose.

“None of this is beyond the profession's competence, but none of it is currently required, described, or standardized, and quality review practices consequently vary in ways that inspection findings are beginning to expose.”

May combine two claims.

Any framework for assurance or liability that fails to distinguish these three strata will either overregulate the harmless or underregulate the hazardous, and much current commentary fails in this way.

“Any framework for assurance or for liability that fails to distinguish these strata will either overregulate the harmless or underregulate the hazardous, and much of the current commentary fails in exactly this way.”

May combine two claims.

At the time of writing, the ISA 500 revision remains in progress within a broader project on audit evidence and risk response, meaning the profession is conducting audits with generative systems under an evidence standard drafted before such systems existed in usable form.

“At the time of writing, the revision remains in progress within a broader project on audit evidence and risk response, which means that the profession is conducting audits with generative systems under an evidence standard drafted before such systems existed in usable form.”

May combine two claims.

If auditors systematically discount algorithmic warnings in complex estimate settings, a plaintiff who shows that the firm's own model flagged the misstatement and that the engagement team overrode the flag without documented justification has a powerful case, and firms should assume discovery will target model logs because such logs now exist.

“If auditors systematically discount algorithmic warnings in complex estimate settings (Commerford et al., 2022), then a plaintiff who shows that the firm's own model flagged the misstatement and that the engagement team overrode the flag without documented justification has a powerful case, and firms should assume that discovery will target model logs precisely because such logs now exist.”

• cites Commerford et al., 2022 sourceNot checked — no full text of the source

• AssumptionRests on an assumption: Model logs now exist and discovery will target them.

May combine two claims.

The PCAOB placed technology on its research agenda and stated that existing standards are not viewed by firms as impediments to generative AI, while emphasizing that due professional care and professional skepticism cannot be delegated to a tool.

“The PCAOB has placed technology on its research agenda and has stated that its existing standards are not viewed by firms as impediments to the use of generative artificial intelligence, while simultaneously emphasizing that due professional care and professional skepticism cannot be delegated to a tool (PCAOB, 2024).”

• cites PCAOB, 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable).

The competent hybrid auditor preserves independent grounds for the conclusions that matter most, so that no material judgment rests on algorithmic output alone.

“And she preserves independent grounds for the conclusions that matter most, so that no material judgment rests on algorithmic output alone.”

May combine two claims.

Post hoc explanation methods produce approximations of a model's behavior, and an approximation of a reasoning process is not the reasoning process, a distinction that matters when the explanation is later used in litigation as though it were a contemporaneous record of professional judgment.

“Post hoc explanation methods produce approximations of a model's behavior, and an approximation of a reasoning process is not the reasoning process, a distinction that matters enormously when the explanation is later deployed in litigation as though it were a contemporaneous record of the basis for a professional judgment.”

Classified as probably not a substantial claim.

If algorithmic tools become ordinary professional equipment, an auditor who declines to use them may fall below the standard, just as a doctor who ignores available diagnostic imaging cannot plead fidelity to the stethoscope.

“If algorithmic tools become ordinary professional equipment, then the auditor who declines to use them may herself fall below the standard, in the same way that a doctor who ignores available diagnostic imaging cannot plead fidelity to the stethoscope.”

Classified as probably not a substantial claim. May combine two claims.

The reliability of information can be assessed by reference to its source and nature, which is why evidence obtained directly by the auditor has traditionally ranked above evidence from the entity, and documentary evidence above oral representation.

“It presupposes, further, that the reliability of information can be assessed by reference to its source and its nature, which is why evidence obtained directly by the auditor has traditionally ranked above evidence obtained from the entity, and documentary evidence above oral representation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Caparo Industries plc v Dickman (1990) performed a similar confinement in the English tradition through proximity and purpose, building on the assumption of responsibility analysis in Hedley Byrne & Co Ltd v Heller & Partners Ltd (1964).

“Caparo Industries plc v Dickman (1990) performed a similar confinement in the English tradition through the requirements of proximity and purpose, building on the assumption of responsibility analysis in Hedley Byrne & Co Ltd v Heller & Partners Ltd (1964).”

• cites Caparo Industries plc v Dickman (1990)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Hedley Byrne & Co Ltd v Heller & Partners Ltd (1964)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

The standards have long required auditors who use the work of an expert to evaluate the expert's competence, objectivity, and methods.

“The standards have long required auditors who use the work of an expert to evaluate the expert's competence, objectivity, and methods.”

Probably not a substantial claim: may not pass S2 (checkable).

Treating those factors as evidentiary attributes, assessable and documentable like any others, brings the black box inside the audit rather than leaving it as an unexamined input.

“Treating those factors as evidentiary attributes, assessable and documentable like any others, is the move that brings the black box inside the audit rather than leaving it as an unexamined input.”

May combine two claims.

Algorithmic output is neither inadmissible nor self-certifying.

“Algorithmic output is neither inadmissible nor self-certifying.”

Classified as probably not a substantial claim. May combine two claims.

Under the international framework, information becomes audit evidence when the auditor evaluates it and uses it to support the conclusions on which the opinion is based, and the auditor must judge both sufficiency and appropriateness.

“Under the international framework, information becomes evidence when the auditor evaluates it and uses it to support the conclusions on which the opinion is based, and the auditor must judge both its sufficiency, a measure of quantity, and its appropriateness, a measure of relevance and reliability (IAASB, 2019).”

• cites IAASB, 2019Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable).

The competent hybrid auditor calibrates her reliance to the demonstrated reliability of the tool for the specific population and assertion at issue.

“She calibrates her reliance to the demonstrated reliability of the tool for the specific population and assertion at issue.”

May combine two claims.

An auditor cannot delegate the opinion, but that does not determine how losses are shared if an algorithm's failure contributes to an audit failure.

“The auditor's non-delegable responsibility for the opinion does not settle the further question of how losses should ultimately be distributed when an algorithmic failure contributes to an audit failure.”

May combine two claims.

An auditor can interrogate a model's validation history, probe its performance on cases with known outcomes, examine whether its training population resembles the population to which it is applied, and design procedures to detect its characteristic failure modes.

“She can, however, interrogate its validation history, probe its performance on cases with known outcomes, examine whether the population on which it was trained resembles the population to which it is applied, and design procedures that would detect its characteristic failure modes.”

Probably not a substantial claim: may not pass S2 (checkable).

A risk score produced by a neural network is, in evidentiary character, closer to the opinion of an expert whose reasoning is undisclosed than to the recalculation it superficially resembles.

“A risk score produced by a neural network is, in evidentiary character, closer to the opinion of an expert whose reasoning is undisclosed than to the recalculation it superficially resembles.”

Classified as probably not a substantial claim.

The negligence inquiry asks how a reasonably competent auditor would have acted.

“The negligence inquiry asks how a reasonably competent auditor would have acted.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

An auditor cannot cross-examine a gradient boosting machine.

“An auditor cannot cross-examine a gradient boosting machine.”

Probably not a substantial claim: may not pass S2 (checkable).

The behavioral literature gives the competent hybrid auditor formulation empirical content that courts will eventually need.

“The behavioral literature gives this formulation empirical content that courts will eventually need.”

Probably not a substantial claim: may not pass S4 (based).

Sufficiency has always been understood as a function of risk and of the quality of individual items of evidence, with lower quality demanding greater quantity.

“Sufficiency has always been understood as a function of risk and of the quality of individual items of evidence, with lower quality demanding greater quantity.”

Classified as probably not a substantial claim.

An auditor cannot assess a system she cannot access.

“An auditor cannot evaluate what she cannot access”

May combine two claims.

Interpretable models sometimes purchase transparency at the price of predictive power, and an auditor who chooses a weaker but explainable fraud detection model over a stronger but opaque one has made a trade-off that the standards currently give her no vocabulary to justify or record.

“Interpretable models, for their part, sometimes purchase transparency at the price of predictive power, and an auditor who chooses a weaker but explainable fraud detection model over a stronger but opaque one has made a trade-off that the standards currently give her no vocabulary to justify or even to record.”

May combine two claims.

Concentrating liability on the audit firm preserves clarity of the assurance promise but leaves the vendor, which has the greatest capacity to prevent model defects, with the least exposure to their consequences.

“Concentrating liability on the firm preserves the clarity of the assurance promise, but it also means that the party with the greatest capacity to prevent model defects, the vendor, bears the least exposure to their consequences.”

• AssumptionRests on an assumption: Audit software is usually licensed on terms that deny suitability for a specific engagement and limit damages to the license fee.

May combine two claims.

ISA 315 as revised in 2019 modernized the risk identification standard and acknowledged automated tools and techniques, and the quality management standards effective in 2022 obliged firms to establish governance over technological resources used in engagements.

“ISA 315 as revised in 2019 modernized the risk identification standard and acknowledged automated tools and techniques, and the quality management standards that took effect in 2022 obliged firms to establish governance over the technological resources used in engagements (IAASB, 2019, 2020).”

• cites IAASB, 2019, 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable).

Management deploying machine learning in financial reporting is obliged, on a sensible reading of its representations, to supply the auditor with records, validation evidence, and access required to grasp how the system affects the figures.

“management that deploys machine learning in financial reporting processes assumes, on any sensible reading of its representations, an obligation to provide the documentation, validation records, and access necessary for the auditor to understand the system's role in the figures.”

Probably not a substantial claim: may not pass S2 (checkable).

When the opaque system belongs to the client, the analysis moves from contribution to cooperation.

“Where the opaque system belongs to the client, the analysis shifts from contribution to cooperation.”

May combine two claims.

A vendor's failure to comply with the AI Act can evidence a defect and may ground contribution claims.

“since compliance failures by a vendor will furnish evidence of defect and perhaps anchor contribution claims”

• cites European Union, 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionRests on an assumption: Audit software is usually licensed on terms that deny suitability for a specific engagement and limit damages to the license fee.

Probably not a substantial claim: may not pass S4 (based).

Auditing has outlasted prior technologies by absorbing them into a stable promise: a competent, independent person reviewed the evidence and reached an opinion.

“The auditing profession has survived every previous technology by absorbing it into an unchanged promise: an independent, competent human examined the evidence and formed an opinion.”

May combine two claims.

The EU AI Act sets up a horizontal, risk-tiered framework, and its duties on transparency, human oversight, and technical documentation cover some systems used in or near financial reporting.

“The European Union's Artificial Intelligence Act establishes a horizontal, risk-tiered regime whose obligations of transparency, human oversight, and technical documentation will reach some systems used in and around financial reporting”

• cites European Union, 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Courts, regulators, insurers, and audit committees are already assigning algorithmic-failure costs by default, via boilerplate contracts and doctrinal inertia.

“Courts, regulators, insurers, and audit committees are already allocating the costs of algorithmic failure by default, through contract boilerplate and doctrinal inertia”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Transparency about technology's role will not close the expectation gap, but it moves debate from imagined machine capability to what the engagement record shows was done.

“Transparency of this kind does not eliminate the expectation gap, but it relocates the argument from what the jury imagines the technology could do to what the engagement actually documented it doing”

• AssumptionRests on an assumption: Public discussion treats AI as superhuman at detection, encouraging the belief that an AI-assisted audit should miss nothing, while vendors do little to correct that belief.

May combine two claims.

Ghana's audit profession is governed by the Companies Act, 2019 (Act 992) and the Data Protection Act, 2012 (Act 843), yet lacks any instrument aimed at algorithmic systems in financial reporting or assurance.

“Consider Ghana, whose audit profession operates under a modern companies statute, the Companies Act, 2019 (Act 992), and a data protection regime, the Data Protection Act, 2012 (Act 843), but without any instrument addressed to algorithmic systems in financial reporting or assurance.”

• cites Companies Act, 2019 (Act 992)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Data Protection Act, 2012 (Act 843)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

UNESCO's AI ethics recommendation, adopted by consensus, sets out broad principles of transparency, accountability, and human oversight.

“UNESCO's recommendation on the ethics of artificial intelligence, adopted by consensus of its member states, articulates principles of transparency, accountability, and human oversight at the broadest level of generality”

• cites UNESCO, 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Fraud investigations into financial statements will more often face records that algorithms produced or filtered.

“Investigations of financial statement fraud will increasingly encounter records produced or filtered by algorithmic systems”

The quote only partly states this claim.

Algorithmic tools will make that coarseness worse, since hindsight bias flourishes when the defendant possessed or could have possessed such a tool.

“Algorithmic assistance will sharpen the coarseness, because outcome bias is easiest to indulge when the defendant had, or could have had, a tool.”

• AssumptionRests on an assumption: Public discussion treats AI as superhuman at detection, encouraging the belief that an AI-assisted audit should miss nothing, while vendors do little to correct that belief.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Professional bodies in developing economies have a harder preparation task: their members will face imported technology before local guidance appears, and for a time the qualification syllabus may be the only preparation instrument available.

“Professional bodies in developing economies face the sharper version of this task, since their members will encounter imported technology before local guidance exists, and the qualification syllabus may be, for a period, the only instrument of preparation the jurisdiction possesses.”

• AssumptionRests on an assumption: Many African audit oversight bodies are new, poorly funded, and burdened by ordinary inspections, so independent model inspection is unlikely soon.

Probably not a substantial claim: may not pass S4 (based).

The auditing profession now works with machines whose reliability it must establish itself.

“The profession that claims the public's trust in the reliability of information now works alongside machines whose reliability it must itself establish.”

May combine two claims.

Forensic analysts who deploy their own machine learning tools face a sharper courtroom explainability problem: expert evidence regimes demand disclosed and testable reasoning, and an opinion that merely reports the model's output may not survive admissibility review.

“forensic analysts deploying their own machine learning tools face, in the courtroom, an intensified version of everything this article has said about explainability, because expert evidence regimes in most jurisdictions require the expert's reasoning to be disclosed and testable, and an opinion that reduces to the model said so may not survive a competent admissibility challenge.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The behavioral findings underlying parts of the article await replication outside North American and European samples.

“the behavioral findings on which parts of this article rest await replication outside North American and European samples”

• AssumptionRests on an assumption: No empirical literature exists on how reliance on imported audit technology affects audit quality in importing countries.

Classified as probably not a substantial claim.

An investigator will first need to reconstruct the inputs shown to the systems, the alerts they raised, and how people responded to those alerts.

“the investigator's first analytical task will be to reconstruct what the systems were shown, what they flagged, and what humans did with the flags.”

May combine two claims.

A fraud detection model trained mainly on North American and European ledgers encodes ideas of normality, and it is most fragile at the edges where local conditions differ from its training data.

“A fraud detection model trained principally on North American and European ledgers embeds assumptions about what normal looks like, and the edge-case fragility described in Section 4 is most acute exactly where local conditions diverge from the training distribution.”

May combine two claims.

AI challenges the audit promise more deeply than earlier tools because it joins the examination itself, not only the collection of evidence, and in its most powerful forms its role is opaque to the human it helps.

“Artificial intelligence tests that promise more searchingly than its predecessors because it participates in the examining, not merely in the gathering, and because its participation is, in its most powerful forms, opaque to the participant it assists.”

May combine two claims.

The US has favored sectoral guidance and voluntary architecture, especially NIST's risk management framework, which gives a shared vocabulary of trustworthiness characteristics without binding force.

“The United States has preferred sectoral guidance and voluntary architecture, most prominently the National Institute of Standards and Technology's risk management framework, which supplies a shared vocabulary of trustworthiness characteristics without binding force”

• cites NIST, 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The framework fits current standards: its stages can be worded using existing ISA terms and inserted into the ISA 500 revision or quality management guidance.

“It is standard-compatible, in that every stage can be expressed in the existing vocabulary of the ISAs and, with modest drafting, incorporated into the ISA 500 revision now in progress or into application material under the quality management standards.”