Outsider Oversight: Designing a Third Party Audit Ecosystem for AI Governance

May combine two claims.

Google researchers on the internal AI team were dismissed and blocked from publishing critiques of large language models that later proved critical to the company's product roadmap.

“For instance, Google researchers on the internal AI team were dismissed and blocked from publishing critiques on the large-scale language models [18,27], which would later be revealed to be critical to the company's future product roadmap [23,85,87].”

• cites [18] sourcePartly supported

  The Introduction reports that Google fired Timnit Gebru and Margaret Mitchell and that they were told not to cast Google’s technologies in a negative light in research focused on large language models. It does not establish that these researchers were blocked from publishing the critiques. The later product-roadmap point has separate citations in the citing sentence and is not attributed to this source. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [27]Not checked yet

  Not checked in the free run.

• cites [23]Not checked yet

  Not checked in the free run.

• cites [85]Not checked yet

  Not checked in the free run.

• cites [87] sourceNot checked — no full text of the source

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

For an audit to successfully measure performance, its scope should be sufficiently delineated.

“For an audit to successfully measure performance, its scope should be sufficiently delineated.”

• cites [42] sourcePartly supported

  The sections on benchmarking and defining a peer group say that how a comparison group and performance range are defined is crucial to assessing a company's ESG performance. That supports the general need to define the basis of an assessment, but the paper does not address audits or state a requirement about an audit's scope (searched: audit, scope, delineated). No page number was available in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

May combine two claims.

Voluntary reporting can be unreliable and create reporting bias.

“Voluntary reporting can be unreliable and create reporting bias [4].”

• cites [4] sourceCould not verify

  The paper’s “Measuring and comparing audiences” section uses self-reported gender and race data and notes that voter-file information may be outdated. Neither that section nor the other passages found when searching for voluntary reporting and reporting bias says that voluntary reporting is unreliable or creates reporting bias. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

May combine two claims.

Similar biases have been found in recent consumer complaints to the CFPB, although early CFPB complaints appeared higher from minority neighborhoods.

“Similar biases have been found in recent consumer complaints to the Consumer Financial Protection Bureau (CFPB) [65]-although early CFPB complaints, if anything, appeared to be higher from minority neighborhoods [6].”

• cites [65] sourceCould not verify

  Section 2.1 discusses bias in automated facial analysis, not consumer complaints. Searches for CFPB, consumer complaints, and minority neighborhoods found no passage in our copy addressing the finding attributed to this source. The separate claim about early complaints is attached to citation [6], not this source. (LLM, conf=0.75)

  Experimental — accuracy not yet measured
• cites [6]Not matched

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

May combine two claims.

Resource constraints prompted development of syndromic surveillance in public health, adverse event reporting for pharmaceuticals, and financial fraud complaint databases to provide real-time signals and focus audit efforts.

“Similar resource constraints prompted the development of”

• cites [35]Not checked yet

  Not checked in the free run.

• cites [66]Not checked yet

  Not checked in the free run.

• cites [53] sourceCould not verify

  The passages on medical oversight (sections 2.3–2.4; page numbers unavailable) discuss professional review and accountability, not the development of surveillance or complaint databases. Searches for syndromic surveillance, pharmaceutical adverse-event reporting, financial fraud complaints, resource constraints, and real-time audit signals found no passage addressing the claim. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

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

Stricter teacher occupational licensing is associated with a wider distribution of student test scores, while other studies found no impact on standardized test scores.

“One study of occupational licensing laws for teachers found that stricter occupational licensing regimes is associated with a wider distribution of student test scores [52], though others concluded that state occupational licensing requirements for teachers had no impact on students' standardized test scores [28].”

• cites [52] sourceCould not verify

  The citation to this source is attached to the first clause, about stricter teacher licensing and the distribution of student test scores. Our copy discusses algorithmic impact assessments and accountability, including testing in a different context (section 4.1), but the passages found for searches on teacher occupational licensing and student test scores do not address that association. The second clause is attributed to a different citation. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [28]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).

Rotation requirements are employed in a minority of audit cases.

“Rotation requirements, which require auditors to rotate across audit targets to ostensibly foster independence, are employed in a minority of cases.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Nearly all audit programs have specific standards and rules that help scope the audit.

“First, we note that nearly all audit programs have specific standards and rules that help to scope the audit.”

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

Significant variation exists within third-party audits: they may be conducted by public agencies or private parties, which typically determines whether auditors are compensated by the audited party.

“Second, we note that significant variation exists within the third party audit category, which may be conducted by public agencies (e.g., nursing home inspections) or private parties (e.g., children's toys), which typically also determines whether auditors are compensated by the audited party.”

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

There are many varieties of audits.

“What this survey highlights most of all is that there are many varieties of audits.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Due to limited regulatory capacity, public inspections appear more likely to be initiated by complaint or risk-based heuristic rather than universally mandated.

“Due to limited regulatory capacity, public inspections appear more likely to be initiated by some complaint or risk-based heuristic, rather than being universally mandated.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In many cases, audit schemes contemplate post-audit actions such as recommendations for remediation and disclosure.

“Last, in many cases, the audit scheme contemplates post-audit actions, such as recommendations for remediation and disclosure.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Accreditation schemes exist for all instances where the auditee selects and compensates the third-party auditor.

“Fourth, accreditation schemes exist for all instances where the auditee selects and compensates the third party auditor.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Accreditation is widespread in instances where auditees choose auditors.

“As seen in the instances where auditees choose auditors in Table 1, accreditation is widespread.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Third-party audits invariably involve privileged access to information.

“As Table 1 shows, third party audits invariably involve privileged access to information.”

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

Audited companies may ignore or overlook audit results.

“Furthermore, audited companies could outright ignore or overlook audit results.”

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

Conflicts of interest can compromise the quality of third-party audits of algorithmic systems, as in ESG.

“Like in ESG, conflicts of interest can compromise the quality of third-party audits of algorithmic systems.”

May combine two claims.

Effective target identification, independence in selection and compensation, access to data and AI systems, clear audit standards, and post-audit transparency are critical for audits to promote rather than degrade accountability.

“Existing evidence from other regulatory fields suggests that effective target identification mechanisms, independence in selection and compensation, access to data and AI systems, clear and well-scoped audit standards, and post-audit report transparency will be critical to ensure that audits promote, rather than degrade, overall accountability.”

May combine two claims.

AIAs are often practically executed by similar stakeholders as internal audits, with broader regulatory implications and purpose of prompting institutional reflexivity and due diligence.

“Thus, AIAs are often practically executed by similar stakeholders as internal audits, with broader regulatory implications and a purpose framed towards prompting and gauging institutional reflexivity and due diligence, in addition to the audit goals of performance assessment with respect to articulated expectations [47].”

• cites [47]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).

Audits can be qualitative or quantitative.

“Audits can be qualitative or quantitative, as system expectations can be articulated and assessed in either mode.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Auditing automated decision systems of a large technology platform can be as complex as auditing human rights violations in a large supply chain.

“Auditing the automated decision systems of a large technology platform can rival the complexity of auditing for human rights violations in a large supply chain.”

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

AIA scope can be broader than a specific product feature and includes reflective articulations of trade-offs and decision-making.

“The scope of an AIA is thus potentially broader than a specific product feature, and the assessment goes beyond just evaluation measurements, to include more reflective articulations of the trade-offs and decision-making inherent in the release and development of a particular algorithmic system.”

May combine two claims.

Auditor access to algorithmic systems is difficult to mediate, and identification of audit targets and scope is uncertain.

“The mediation of auditor access is difficult, and the identification of audit targets and scope uncertain.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Audits evaluate performance relative to expected behavior as part of a broader accountability process.

“Audits evaluate performance relative to expected behavior, as part of a broader accountability process [62].”

• cites [62] sourceNot checked — no full text of the source

May combine two claims.

Third-party oversight work has led to voluntary moratoriums, lawsuits, regulatory action, congressional hearings, and recalls or redesigns of deployed algorithms.

“The impact of this work is undeniable -leading to voluntary moratoriums [30], and several successful lawsuits [75,79]; nation-wide waves of regulatory action and public outcry [74]; congressional hearings [3]; and the recalls or radical redesigns of the deployed algorithms [12,64,65,88].”

• cites [30]Not matched

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

• cites [75]Not matched

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

• cites [79] sourceNot checked — no full text of the source
• cites [74]Not matched

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

May combine two claims.

Algorithmic audit standards and scope may be unclear, leading to unfocused and expensive assessments.

“The standards and scope of an algorithmic audit may be unclear -leading to assessments that are unfocused and expensive.”

May combine two claims.

Because ESG is ambiguous, ESG certifications risk becoming cheap talk that rubber stamps practices without promoting social responsibility.

“And due to the ambiguity of what is being audited, ESG certifications risk becoming”

• cites [8]Not matched

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

• cites [19] sourceNot checked — no full text of the source
• cites [37]Not matched

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

May combine two claims.

Regulatory and institutional interpretations of hate speech differ from those experiencing hate online, leading to moderation tools that marginalize and censor impacted users.

“In the same vein, regulatory and institutional interpretations of hate speech often differs dramatically from the understanding of those experiencing hate online [43], leading to deployed moderation tools that further marginalize and censor impacted users rather than protect them [16].”

• cites [43]Not matched

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

• cites [16]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).

Audits are conventionally classified into first-, second-, and third-party types.

“By convention, audits are classified into three types [14].”

• cites [14]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). May combine two claims.

Clearview AI claimed its facial recognition was 100% accurate across demographic groups under ACLU methodology but did not address ethical and legal challenges of its face data sourcing.

“The controversial facial recognition company Clearview AI, for example, claimed to be”

• cites [29]Not matched

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

May combine two claims.

There is precedent for auditors compromised by selection, payment, or employment by auditees; access struggles; diffuse performance standards; and auditees concealing findings.

“There is a clear precedent for situations of auditors compromised by selection, payment, or employment by auditees; auditors struggling to secure access; auditors clumsily navigating diffuse performance standards; and auditees that may elect to share or conceal findings otherwise hidden behind a nondisclosure agreement.”

• cites [8]Not matched

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

• cites [19] sourceNot checked — no full text of the source
• cites [37]Not matched

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

May combine two claims.

Internal accountability tends to prioritize at most regulators' or customers' concerns, primarily to reduce liability risk, which may not align with impacted populations.

“This tends to prioritize at most the concerns of regulators or customers, primarily to reduce liability risk, which may not align with core concerns of impacted populations.”

• cites [54]Not matched

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

• cites [66]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). May combine two claims.

Third-party oversight by non-profits, journalists, and academics has exposed AI weaknesses but is not conventionally considered audits.

“We note that while third party oversight by non-profit organizations, journalists, and academics have, as noted above, played critical roles in exposing weaknesses to AI systems, such forms of oversight are not conventionally considered audits per se.”

May combine two claims.

Unlike most audits, AIAs are less strictly defined and tend to include more open-ended inquiry into impacts.

“Unlike most audits, AIAs are not as strictly defined and tend to include more open-ended inquiry into the impacts of a particular algorithmic deployment [55].”

• cites [55]Not matched

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

May combine two claims.

These format and content developments restrict and define audit content, focusing on concerns preferable for internal auditors and requiring internal access or cooperation.

“All of these format and content oriented developments work to restrict and define what is to be included in the audit, focusing on concerns that are preferable for internal auditors to operationalize, and that requires a certain level of access or internal co-operation to achieve.”

May combine two claims.

The turn toward audits alone is unlikely to achieve actual algorithmic accountability; meaningful third-party involvement requires sustained focus on institutional design.

“We conclude that the turn toward audits alone is unlikely to achieve actual algorithmic accountability, and sustained focus on institutional design will be required for meaningful third party involvement.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Identifying audit targets and scope is critical to ensuring audits are applied to relevant parties and conducted for the right issues.

“Identifying audit targets and scope is critical to ensuring that audits are applied to relevant parties, and conducted for the right issues.”

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

Risk-based or complaint-based selection of audit targets can help allocate attention where needed most when resources are scarce.

“In those instances, risk-based or complaint-based selection of audit targets can help to allocate attention where needed the most.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Companies can adopt incident reporting systems to raise user and public awareness of a system's use in consequential scenarios.

“Companies can themselves adopt incident reporting systems to raise user and public awareness of a system's use in consequential scenarios.”

May combine two claims.

A third-party audit where the auditee selects and compensates the auditor, determines scope and access, and limits post-audit actions may be functionally indistinguishable from a first-party audit and questionably effective for downstream accountability.

“A third party audit where (a) the auditor is selected and compensated by the auditee (b) to conduct an audit of a scope and access determined by the auditee, and (c) with limited post-audit actions may be functionally indistinguishable from firstparty audit, and questionably effective in garnering the expected downstream accountability outcomes.”

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

Serious limitations characterize the current AI audit landscape despite policy attention turning toward audits.

“In sum, while policy attention has been turning toward audits, serious limitations characterize the current AI audit landscape.”

May combine two claims.

Some proposals make regulatory agencies the primary external auditor, but regulatory capacity to conduct such audits is lacking.

“Other proposals put the onus on regulatory agencies to serve the primary role of external auditor [36], but regulatory capacity to conduct such audits is currently lacking [22].”

• cites [36]Not matched

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

• cites [22]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). May combine two claims.

Privacy disclosure pop-ups implemented in response to the EU's ePrivacy Directive increased consent fatigue without meaningfully improving user privacy.

“Some have noted, for example, that the privacy disclosure pop-ups implemented in response to the EU's ePrivacy Directive have increased”

• cites [67]Not matched

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

• cites [76]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).

Third-party reporting has a similar effect on small businesses.

“Third-party reporting has a similar effect on small businesses [2].”

• cites [2]Not matched

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

May combine two claims.

One study found restaurant compliance notices were not representative of actual food safety status and pushed inspectors to spend resources adjudicating accuracy rather than addressing safety hazards.

“One study found, for instance, that restaurant compliance notices were not representative of actual food safety status, and pushed inspectors to spend outsize resources on adjudicating the accuracy of these notices rather than addressing safety hazards directly [36].”

• cites [36]Not matched

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

May combine two claims.

Before the 2019 Gender Shades audit, NIST failed to measure demographic differences in its FRVT, overlooking for years that Asian and African American people were up to 100 times more likely to be misidentified than white men.

“Before 2019 and the intervention of the Algorithmic Justice League's Gender Shades audit [10], the U.S. National Institute of Technology (NIST) failed to measure demographic differences in their regular Facial Recognition Vendor's Test (FRVT), thus overlooking for many years that”

• cites [10]Not matched

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

• cites [26]Not matched

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

May combine two claims.

The Algorithmic Accountability Act of 2022 and the Algorithmic Justice and Online Platform Transparency Act of 2021 contemplate obligations on regulated parties with enforcement by the FTC.

“Similarly, the Algorithmic Accountability Act in 2022, and the Algorithmic Justice and Online Platform Transparency Act of 2021 [49] contemplate obligations on regulated parties with enforcement by the Federal Trade Commission (FTC).”

• cites [49]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).

When audit resources are scarce, it may not be possible to select all parties for audit.

“When audit resources are scarce, however, it may not be possible to select all parties for an audit.”

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

A future where AI audits are conducted only by company employees and consulting companies fails to learn lessons of how third parties have generated robust oversight.

“A future where AI audits are conducted by company employees and consulting companies alone fails to learn the lessons of how third parties have generated robust oversight of AI systems.”

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

Third-party oversight of AI products has faced challenging access, in sharp contrast to other audit systems.

“This stands in sharp contrast to third party oversight of AI products, where access has proven challenging.”

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

AI systems are diverse, impact a range of industries and stakeholders, and require more third-party participants with varied tools to assess harms for implicated stakeholders.

“Such systems are diverse, impacting a range of industries and stakeholders, and it will require more third party participants, representing a variety of communities and making use of a wide set of tools, to concretely assess a wider range of the harms involved for the implicated stakeholders.”

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

Independence is key to third-party audits.

“The key to third-party audits is independence.”

Classified as probably not a substantial claim.

Although definitions are provided for recommender systems and automated decision systems, there is little guidance for translating them into specific performance expectations.

“Although definitions are provided for these terms, there is little guidance for how regulators, auditors, and companies are meant to translate such descriptions into specific performance expectations.”

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

The literature suggests benefits to establishing a national incident reporting system enabling third parties to file complaints about algorithmic systems.

“The literature suggests that there may be benefits to establishing a national incident reporting system to enable third parties to file complaints about algorithmic systems.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

An audit may more productively ask whether an advertising platform violates disparate treatment provisions of the Fair Housing Act rather than asking whether there is bias.

“Instead of asking,”

• cites [82]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).

System of Records Notices may alert parties to potential uses of algorithms.

“System of Records Notices (SORN), which are intended to notify the public whenever a federal agency creates, modifies, or eliminates records involving personally identifiable information, may alert parties to potential uses of algorithms.”

May combine two claims.

The status quo is that algorithm use is rarely disclosed transparently, making it hard to identify audit targets.

“The status quo is that the use of algorithms is rarely disclosed in a transparent manner, making it hard to identify audit targets in the first place [84].”

• cites [84]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).

Compensation consultants recommended higher CEO salaries when they also offered additional services to those executives.

“Murphy & Sandino (2009) found that compensation consultants recommended higher CEO salaries when those consultants offered additional services to those executives [59].”

• cites [59]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).

Some evidence suggests auditing can help standardize information for more consistent reporting.

“Some evidence suggests that auditing can help to standardize information for more consistent reporting [30].”

• cites [30]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). May combine two claims.

Cross-selling opportunities may lower statutory audit quality but did not affect the quality of independent expert reports for takeover targets, though they may lower fees.

“Another study, however, concluded that while cross-selling opportunities could lead to lower quality statutory audits, they did not have an impact on the quality of independent expert reports provided to takeover targets-though the incentive to cross-sell could have an impact on lowering fees [9].”

• cites [9]Not matched

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

May combine two claims.

No auditor can audit an entire VLOP single-handedly, and the definition of ADS is context-dependent and broad, doing little to help scope an audit.

“No auditor can audit an entire VLOP single-handedly, and the definition of ADS is so context-dependent and purposefully broad [68] that it does little to help scope an audit.”

• cites [68]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).

Private car emissions testing facilities with more chances to cross-sell services are more likely to pass cars that fail emissions standards.

“Another study found that private car emissions testing facilities with more opportunities to cross-sell services to customers were more likely to pass cars that did not meet emissions standards [63].”

• cites [63] sourceNot checked — no full text of the source

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

Regulators can impose consequences ranging from monetary sanctions to industry bars for violations of auditor independence rules.

“If an auditor violates these independence rules, regulators can impose a range of consequences ranging from monetary sanctions to a bar from the industry [1].”

• cites [1]Not matched

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

May combine two claims.

Financial auditing independence rules prohibit close auditor-client relationships and prevent auditors from helping prepare financial statements.

“One approach to reduce concerns regarding familiarity is seen in financial auditing, where independence rules prohibit close relationships between auditors and their clients, and prevent the auditor from assisting in the preparation of the financial statements.”

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

Opening audits to third-party participation could allow diverse stakeholders to bring precise inquiries specific to their context, focus, and organizational target.

“Opening audits up for third party participation could also allow for a variety of stakeholders to come in with diverse yet precise inquiries specific to the context, focus and organizational target of which they are most concerned.”

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

Regulators can use incident reporting information to identify priority targets and assess fees or penalties based on audit results.

“Regulators can use this information to identify specific priority targets and assess fees or penalties based on audit results.”

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

Minimizing conflicts of interest and undue influence helps ensure the reliability of audit results.

“Minimizing conflicts of interest and undue influence helps to ensure the reliability of audit results.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Regulators could incentivize third-party auditors to inspect repeatedly flagged cases through RFPs or bug bounties.

“They could then incentivize third party auditors to inspect repeatedly flagged cases in this database through various means, such as RFPs or bug bounties.”

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

Adjusting for differences in reporting propensities and actively soliciting minority perspectives will be key components of an incident reporting system's design.

“Nonetheless, considerations for how to adjust for differences in reporting propensities and actively solicit minority perspectives for incident reporting will be a key component of the system's design.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Sarbanes-Oxley requires mandatory rotation of the lead audit partner.

“Sarbanes-Oxley, for example, requires mandatory rotation of the lead audit partner.”

May combine two claims.

Third parties often face great legal risk to gain access to audited AI systems and are vulnerable to corporate retaliation or obstruction.

“Currently, third parties are often forced to operate at great legal risk to gain access to the audited AI system, keeping them vulnerable to various forms of corporate retaliation or obstruction.”

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

The literature is mixed on licensing benefits, and accreditation through a professional organization may be preferred.

“The literature is much more mixed on the benefits of licensing schemes; accreditation through a professional organization may be preferred for three reasons.”

• AssumptionRests on an assumption: Demand for AI audit services will grow significantly in the near future.
• AssumptionRests on an assumption: Occupational licensing reduces the supply of services.

May combine two claims.

More training and greater team experience are linked to better audit results.

“Training improves audit quality and more experienced team members are associated with more effective audits [5].”

• cites [5]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). May combine two claims.

A shift in Belgian financial audit standards toward more rule-based procedures decreased auditing errors and increased auditor independence.

“One study, for example, found that a shift in Belgian financial audit standards towards more rule-based procedures decreased financial auditing errors and increased auditor independence [11].”

• cites [11]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).

More rigorous dental licensing standards did not improve average patient dental outcomes.

“Another analysis of dental licensing requirements found that more rigorous licensing standards did not improve average dental outcomes of patients [45].”

• cites [45]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).

Less empirical work has examined whether professional licensing affects quality than has examined costs and barriers to entry.

“While much empirical work has focused on the costs and barriers to entry imposed by occupational licensing, less empirical work has probed whether professional licensing has an impact on quality [44].”

• cites [44]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).

PCAOB inspections improved audit quality by prompting lower-quality auditors to exit the market.

“Another found that PCAOB inspections, which assess the compliance of registered public accounting firms with professional audit rules and standards, improved audit quality by prompting lower-quality auditors to exit the market [17].”

• cites [17]Not matched

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

May combine two claims.

Corporations can block access through paywalls, prohibitive Terms of Service, or product design that obscures test points.

“Corporations can also use non-legal strategies to block access, including setting up a paywall, adopting prohibitive Terms of Service conditions, or structuring the product to obscure any clear set of test points.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

Fears of legal repercussions or corporate retaliation can weaken audit inquiry, and professional standards can help determine limited conditions for liability.

“Fears of legal repercussions or corporate retaliation can weaken the audit inquiry, and professional standards can help determine limited conditions for liability.”

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

Increasing or decreasing CPA licensing requirements had no meaningful effect on CPA quality.

“Several studies have found that increasing or decreasing licensing requirements for CPAs had no meaningful effect on CPA quality [7,12,16].”

• cites [7]Not matched

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

• cites [12]Not matched

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

• cites [16]Not matched

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

May combine two claims.

HireVue commissioned and controlled the scope of its ORCAA audit, including mediating public communication through its PR channels.

“Similarly, Hirevue commissioned and controlled the scope of its audit by ORCAA, including by mediating any public communication of audit results through their own public relations channels.”

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

Unannounced inspections provided better information about nursing home care quality and could reduce regulatory burdens, though announced inspections could better identify organizational issues.

“Another study found that unannounced inspections provided better information about the quality of care at nursing homes, and could reduce the regulatory burdens of conducting inspections-though announced inspections could be better suited to identifying organizational issues [46].”

• cites [46]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).

Existing research on whether professional licensing affects quality is mixed.

“Existing research on this question is mixed.”

Classified as probably not a substantial claim.

The PCAOB experience supports creating an audit oversight board within a public agency or public interest institution to monitor auditors and set scope and standards.

“In particular, the PCAOB experience lends support to the creation of an audit oversight board, housed within a public agency or operating as a separate public interest institution, in order to monitor auditors and to determine the expected scope and standards for audit work.”

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

The Department of Justice and the Department of Housing and Urban Development have engaged in audits to analyze disparate outcomes and civil rights compliance for decades.

“Similarly, the Department of Justice and the Department of Housing and Urban Development have engaged in audits to analyze disparate outcomes and compliance to civil rights guidelines for decades [78].”

• cites [78]Not matched

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

May combine two claims.

Third-party AI auditors have faced cease-and-desist orders, legal threats, and subpoenas.

“For instance, those making use of crowd-sourcing data collection tools can be interrupted abruptly by claims of violations to privacy or intellectual property rights -NYU researchers studying the Facebook ad delivery algorithm were presented with a cease-and-desist order [58]; German auditors at Algorithm Watch were legally threatened by Instagram [9]; and civil society watch dogs were subpoenaed”

• cites [58]Not matched

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

• cites [9]Not matched

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

• cites [46]Not matched

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

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

Auditors subject to PCAOB inspections produced higher quality audits than auditors the PCAOB was not allowed to inspect, suggesting access to regulators may provide additional benefits.

“Another study found that auditors who were subject to PCAOB inspections produced higher quality audits compared to auditors (typically auditors of foreign SEC registrants) that the PCAOB was not allowed to inspect, suggesting that providing access to parties beyond auditors, such as regulators, may provide additional benefits [51].”

• cites [51]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). May combine two claims.

There is no clear sense of expectations for what a quality AI audit entails or who is qualified to execute it.

“These auditors make use of a range of tools and methodologies to execute their work, and there is no clear sense of expectations as to what a quality audit entails or who is qualified to execute audit work in the algorithmic context.”

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

Non-academic groups can use a broader range of strategies to conduct audits and understand harms to distinct communities.

“Non-academic groups can also take advantage of a broader range of strategies to conduct audits and understand harms to distinct communities.”

May combine two claims.

In food safety inspections, frequent interactions between inspectors and facilities may improve expertise but can increase social coercion and bribery.

“This tradeoff has also been documented in food safety inspections, where frequent interactions between inspectors and facilities may improve inspector expertise, but can increase the likelihood of social coercion and even bribery [49,68].”

• cites [49]Not matched

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

• cites [68]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). May combine two claims.

All audit systems provide some form of privileged access to auditors, and disclosure need not be direct or absolute.

“Protecting proprietary information is not a proper response, as all audit systems provide some form of privileged access to auditors, and disclosure does not need to be direct nor absolute.”

• AssumptionRests on an assumption: All audit systems provide some form of privileged access to auditors.
• AssumptionRests on an assumption: Disclosure of audit information need not be direct or absolute.

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

Training, standardization, and accreditation of auditors are key to ensuring audits are accurately and consistently conducted.

“Training, standardization, and accreditation of auditors are also key to ensuring that audits are accurately and consistently conducted.”

May combine two claims.

A systematic analysis of PwC monitoring found that the resulting audits omitted major violations and had serious deficiencies.

“One systematic analysis of the monitoring methods employed by the company PricewaterhouseCoopers (PwC), for example, found that the resulting audits omitted several major violations and contained serious deficiencies [62].”

• cites [62] sourceNot checked — no full text of the source

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

For third-party audits to be effective, auditors must be granted access to enough data for a robust review.

“For third-party audits to be effective, auditors must be granted access to enough data to conduct a robust review.”

May combine two claims.

The literature strongly supports training, standardization, and accreditation for third-party AI auditors.

“The literature strongly supports training, standardization, and accreditation for third-party AI auditors.”

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

The FDA may withdraw an auditor's accreditation if the auditor certifies a supplier later linked to a foodborne illness outbreak or refuses FDA access to records.

“The FDA, for example, may withdraw an auditor's accreditation if that auditor certifies a supplier that is later linked to an outbreak of foodborne illness or if it refuses to give the FDA access to its records [54].”

• cites [54]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).

The PwC shortcomings were attributed in part to management pre-screening of information and interviews for auditors.

“The author attributed these shortcomings in part to management pre-screening of information and interviews for auditors.”

• cites [62] sourceNot checked — no full text of the source

May combine two claims.

The audit approach is tangible and can reveal concrete issues and repeated product failures that general standards can later resolve.

“A major advantage of the audit approach we articulate is its tangibility: audits can help reveal concrete issues and repeated product failures that can later be resolved by general standards.”

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

Restaurant grading across ten jurisdictions suffered from grade inflation and grading inconsistencies.

“Another analysis of restaurant grading across ten jurisdictions found that the practice-meant to protect consumers and improve food sanitation via targeted disclosures-suffered from several serious deficiencies, including grade inflation and grading inconsistencies.”

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

AI policy can foster third-party audits through deliberate institutional design informed by other industries' audit schemes.

“AI policy can foster third party audits with more deliberate institutional design, informed by the audit schemes from other industries.”

May combine two claims.

Public visibility of audit results can improve accountability by preventing concealment of undesirable outcomes and incentivizing better behavior.

“Public visibility of audit results can improve accountability in several ways.It can prevent companies from being able to disguise or hide undesirable audit outcomes [21] and incentivize better behavior.”

• cites [21]Not matched

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

May combine two claims.

Confidential PwC audits were shielded from verification by other researchers or NGOs and limited improvement of regulated entity conditions.

“An analysis of PwC's auditing practices also found that the confidential nature of PwC's audits shielded them from verification by other researchers or NGOs and limited their ability to improve the conditions of the regulated entity [62].”

• cites [62] sourceNot checked — no full text of the source

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

The paper's survey is necessarily selective and incomplete.

“As just a first step in the direction of learning from cognate audit schemes, and because the evidence draws from so many different areas, our survey is necessarily selective and incomplete.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

IEEE P7013 standard for automated facial analysis is directly informed by the Gender Shades audit.

“For instance, IEEE P7013 Inclusion and Application Standards for Automated Facial Analysis Technology industry standard is directly informed by the Gender Shades audit.”

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

The paper's case studies and evidence are not exhaustive.

“First of all, our case studies and evidence offered above are not exhaustive.”

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

Publicly traded companies spend an average of $2.2 million to comply with Sarbanes-Oxley financial reporting requirements.

“One study found that publicly traded companies spend an average of $2.2 million to comply with the Sarbanes-Oxley Act's financial reporting requirements [50].”

• cites [50]Not matched

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

May combine two claims.

A registry of audit results avoids bias from pre-publication review or NDA secrecy and ensures audits provide market incentives, inform enforcement, and provide evidence for standard setting.

“A registry of audit results avoids the bias from pre-publication review or the secrecy of findings under NDAs and ensures that audits can provide market incentives for companies to correct behavior, inform enforcement actions, and provide an evidence base for standard setting.”

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

Mandated disclosure of food inspection results increased food business compliance with food safety regulations, though other studies found similar schemes ineffective.

“One study found that mandated disclosure of food inspection results increased the compliance of food businesses with food safety regulations [8], though others have found similar food safety disclosure schemes to be ineffective [31,36].”

• cites [8]Not matched

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

• cites [31]Not matched

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

• cites [36]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 S2 (checkable), S4 (based).

The AI community has largely ignored what other audit schemes teach about fostering policy environments for third-party participation.

“Our point in this specific text is more simple: as the AI community turns toward audits, it has largely ignored what other audit schemes have to teach us on fostering policy environments conducive to third party participation.”

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

A restaurant grading scheme shifted inspection resources away from more significant safety hazards.

“The implementation of a restaurant grading scheme shifted valuable inspection resources away from more significant safety hazards [36].”

• cites [36]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 S2 (checkable).

Specific interventions are necessary to allow effective participation of third parties in AI audits.

“The mere insistence for audits is not enough -in particular, specific interventions will be necessary to allow for the effective participation of third parties, who play a critical role yet continue to face serious and often debilitating challenges when engaged in their work.”

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

Limited disclosure of audit results would balance proprietary information protection against external accountability benefits.

“Limited disclosure of audit results would balance the protection of proprietary information against the benefits of external accountability.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

There is tension between AI transparency and enforcing intellectual property rights.

“Other commentators have also noted the tension that exists, particularly in the context of algorithmic accountability, between the desire for AI transparency and the desire to enforce existing intellectual property rights.”