Towards AI Accountability Infrastructure: Gaps and Opportunities in AI Audit Tooling

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

Past research in HCI and social computing has developed and studied a host of fairness, explainability, and other toolkits that inform AI evaluations.

“Past research in human-computer interaction (HCI) and social computing has developed and studied a host of fairness, explainability, and other toolkits that inform such evaluations [5, 15, 39, 54, 69, 70, 72, 80, 82, 85, 104, 119, 125, 126, 132].”

• cites [5] [5] Allison Woodruff, Sarah E. Fox, Steven Rousso-Schindler, and Jeffrey Warshaw. 2018. A Qualitative Exploration of Perceptions of Algorithmic Fairness. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (CHI '18). Association for Computing Machinery, New York, NY, USA, (Apr. 2018), 1-14. isbn: 978-1-4503-5620-6. doi: 10.1145/3173574.3174230. - sourcePartly supported

  In “Workshop as Method” (page unavailable), the paper describes using a workshop as a research instrument to study perceptions of algorithmic fairness and inform future technology and policy. That supports the sentence’s broader point about HCI methods informing fairness work, but our copy does not describe this study as developing or studying a fairness or explainability toolkit for AI evaluations. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [15] [15] Rachel K. E. Bellamy et al. 2018. AI Fairness 360: An Extensible Toolkit for Detecting, Understanding, and Mitigating Unwanted Algorithmic Bias. (Oct. 2018). arXiv: 1810.01943 [cs] . doi: 10.48550/arXiv.1810.01943. - sourceSupported
• cites [39] [39] Michael A. Madaio, Luke Stark, Jennifer Wortman Vaughan, and Hanna Wallach. 2020. Co-Designing Checklists to Understand Organizational Challenges and Opportunities around Fairness in AI. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI '20). Association for Computing Machinery, New York, NY, USA, (Apr. 2020), 1-14. isbn: 978-14503-6708-0. doi: 10.1145/3313831.3376445. - sourceSupported

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

Auditors need tools that support key components of the accountability process, including target identification, standardization of practice, communication, and advocacy.

“A thorough evaluation alone is not sufficient to hold key stakeholders responsible for system-wide behavior [13, 88]-auditors also need tools that support key components of the accountability process, including target identification, standardization of practice, communication, and advocacy [136].”

• cites [136] [136] Sasha Costanza-Chock, Inioluwa Deborah Raji, and Joy Buolamwini. 2022. Who Audits the Auditors? Recommendations from a field scan of the algorithmic auditing ecosystem. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22). Association for Computing Machinery, New York, NY, USA, (June 2022), 1571-1583. isbn: 978-1-4503-9352-2. doi: 10.1145/3531146.3533213. - sourcePartly supported

  Section 4.1.1 reports that few auditors use standardized frameworks and tools, while section 5.2 recommends standards and disclosure of audit findings—supporting the need for better practice and communication. Section 4.2.1 also links public pressure to action on ethical AI issues. The paper does not clearly recommend tools for target identification and advocacy as components of an accountability process. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

Microsoft's AI Fairness checklist was co-designed with practitioners in an effort to be more compatible with practical challenges, but many other standards frameworks did not obviously consult practitioners and fewer involved affected stakeholders.

“Microsoft's AI Fairness checklist, for example, was co-designed with practitioners [39] in an effort to be more compatible with practical challenges, but many of the other standards frameworks we found did not obviously consult practitioners and fewer involved affected stakeholders.”

• cites [39] [39] Michael A. Madaio, Luke Stark, Jennifer Wortman Vaughan, and Hanna Wallach. 2020. Co-Designing Checklists to Understand Organizational Challenges and Opportunities around Fairness in AI. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI '20). Association for Computing Machinery, New York, NY, USA, (Apr. 2020), 1-14. isbn: 978-14503-6708-0. doi: 10.1145/3313831.3376445. - sourceSupported

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Open tooling practices can have complex interactions with power and oversight.

“but can have complex interactions with power and oversight that are worthy of further study [137].”

• cites [137] [137] David Gray Widder, Meredith Whittaker, and Sarah Myers West. 2024. Why 'open' AI systems are actually closed, and why this matters. Nature , 635, 8040, (Nov. 2024), 827-833. doi: 10.1038/s41586-024-08141-1. - sourceSupported

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

Participation is an increasing focus for AI audit practitioners and audit tool developers.

“it is no surprise that participation is an increasing focus for AI audit practitioners and audit tool developers as well.”

• cites [90] [90] Kulynych, Bogdan, David Madras, Smitha Milli, Inioluwa Deborah Raji, Angela Zhou, and Richard Zemel. 2020. Participatory approaches to machine learning. International Conference on Machine Learning Workshop. (July 2020). -Not checked yet

  Not checked in the free run.

• cites [92] [92] AbebaBirhane, William Isaac, Vinodkumar Prabhakaran, Mark Diaz, Madeleine Clare Elish, Iason Gabriel, and Shakir Mohamed. 2022. Power to the People? Opportunities and Challenges for Participatory AI. In Proceedings of the 2nd - ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '22). Association for Computing Machinery, New York, NY, USA, (Oct. 2022), 1-8. isbn: 978-1-4503-9477-2. doi: 10.1145/3551624.3555290. -Not matched

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

• cites [121] [121] Fernando Delgado, Stephen Yang, Michael Madaio, and Qian Yang. 2023. The Participatory Turn in AI Design: Theoretical Foundations and the Current State of Practice. In Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '23). Association for Computing Machinery, New York, NY, USA, (Oct. 2023), 1-23. isbn: 979-84007-0381-2. doi: 10.1145/3617694.3623261. - sourcePartly supported

  The Introduction describes increasing calls for stakeholder participation in AI design and a growing consensus that stakeholders should participate more. It does not establish that participation is an increasing focus specifically for AI audit practitioners or audit tool developers. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

Fewer studies examine how practitioners evaluate criteria such as basic functionality, safety, privacy, or recourse, just as fewer tools exist for this purpose.

“While there are multiple studies on the use of tools for fairness evaluation [54, 70, 119] and explainability [1, 46, 72, 82, 85, 98], fewer studies examine how practitioners evaluate other criteria such as basic functionality [117], safety, privacy, or recourse, just as fewer tools exist for this purpose.”

• cites [54] [54] Wesley Hanwen Deng, Manish Nagireddy, Michelle Seng Ah Lee, Jatinder Singh, Zhiwei Steven Wu, Kenneth Holstein, and Haiyi Zhu. 2022. Exploring How Machine Learning Practitioners (Try To) Use Fairness Toolkits. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22). Association for Computing Machinery, New York, NY, USA, (June 2022), 473-484. isbn: 978-1-4503-9352-2. doi: 10.1145/3531146.353 3113. - sourceSupported
• cites [70] [70] Kenneth Holstein, Jennifer Wortman Vaughan, Hal Daumé, Miro Dudik, and Hanna Wallach. 2019. Improving Fairness in Machine Learning Systems: What Do Industry Practitioners Need? In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems . ACM, Glasgow Scotland Uk, (May 2019), 1-16. isbn: 978-1-4503-5970-2. doi: 10.1145/3290605.3300830. - sourceCould not verify

  The citation [117] is attached to “basic functionality.” In “Needs for More Proactive Auditing Processes” (p. 8), the paper discusses a checklist for accuracy only as a contrast to the lack of one for fairness. The passages examined do not establish that fewer studies examine how practitioners evaluate basic functionality or that fewer tools exist for it. (LLM, conf=0.77)

  Experimental — accuracy not yet measured
• cites [119] [119] Michelle Seng Ah Lee and Jat Singh. 2021. The Landscape and Gaps in Open Source Fairness Toolkits. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (CHI '21). Association for Computing Machinery, New York, NY, USA, (May 2021), 1-13. isbn: 978-1-4503-8096-6. doi: 10.1145 /3411764.3445261. - sourceSupported
• cites [1] [1] Sunnie S. Y. Kim, Elizabeth Anne Watkins, Olga Russakovsky, Ruth Fong, and Andrés Monroy-Hernández. 2023. "Help Me Help the AI": Understanding How Explainability Can Support Human-AI Interaction. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, (Apr. 2023), 1-17. isbn: 978-1-4503-9421-5. doi: 10.1145/3544548.3581001. - sourceSupported
• cites [46] [46] Danding Wang, Qian Yang, Ashraf Abdul, and Brian Y. Lim. 2019. Designing Theory-Driven User-Centric Explainable AI. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI '19). Association for Computing Machinery, New York, NY, USA, (May 2019), 1-15. isbn: 978-14503-5970-2. doi: 10.1145/3290605.3300831. - sourceNot checked — no full text of the source
• cites [72] [72] Harmanpreet Kaur, Harsha Nori, Samuel Jenkins, Rich Caruana, Hanna Wallach, and Jennifer Wortman Vaughan. 2020. Interpreting Interpretability: Understanding Data Scientists' Use of Interpretability Tools for Machine Learning. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI '20). Association for Computing Machinery, New York, NY, USA, (Apr. 2020), 1-14. isbn: 978-1-4503-6708-0. doi: 10.1145/3313831.3376219. - sourceNot checked — no full text of the source
• cites [82] [82] Alison Smith-Renner, Ron Fan, Melissa Birchfield, Tongshuang Wu, Jordan Boyd-Graber, Daniel S. Weld, and Leah Findlater. 2020. No Explainability without Accountability: An Empirical Study of Explanations and Feedback in Interactive ML. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI '20). Association for Computing Machinery, New York, NY, USA, (Apr. 2020), 1-13. isbn: 978-1-4503-6708-0. doi: 10.1145/33138 31.3376624. - sourceNot checked — no full text of the source
• cites [85] [85] Astrid Bertrand, Tiphaine Viard, Rafik Belloum, James R. Eagan, and Winston Maxwell. 2023. On Selective, Mutable and Dialogic XAI: a Review of What Users Say about Different Types of Interactive Explanations. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, (Apr. 2023), 1-21. isbn: 978-1-4503-9421-5. doi: 10.1145/3544548.3581314. - sourceSupported
• cites [98] [98] Q. Vera Liao, Daniel Gruen, and Sarah Miller. 2020. Questioning the AI: Informing Design Practices for Explainable AI User Experiences. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI '20). Association for Computing Machinery, New York, NY, USA, (Apr. 2020), 1-15. isbn: 978-1-4503-6708-0. doi: 10.1145/3313831.3376590. - sourceSupported
• cites [117] [117] Inioluwa Deborah Raji, I. Elizabeth Kumar, Aaron Horowitz, and Andrew Selbst. 2022. The Fallacy of AI Functionality. In 2022 ACM Conference on Fairness, Accountability, and Transparency . ACM, Seoul Republic of Korea, (June 2022), 959-972. isbn: 978-1-4503-9352-2. doi: 10.1145/3531146.3533158. - sourcePartly supported

  The Related Work section says little has been done to systematically discuss functionality problems, which supports the narrower point that basic functionality receives less research attention. It does not establish that fewer studies examine how practitioners evaluate functionality. The paper also says in §5.2.3 that many model-logging and evaluation tools are available, rather than establishing a shortage of such tools. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

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

Shapley values are prone to misuse and may be unsuitable for normative evaluation.

“But as Kumar et al. [95] argue, Shapley values are prone to misuse and may be unsuitable for normative evaluation.”

• cites Kumar et al. [95] [95] I. Elizabeth Kumar, Suresh Venkatasubramanian, Carlos Scheidegger, and Sorelle A. Friedler. 2020. Problems with Shapley-value-based explanations as feature importance measures. In Proceedings of the 37th International Conference on Machine Learning (ICML'20). Vol. 119. JMLR.org, (July 2020), 54915500. Retrieved Nov. 20, 2023 from. - sourceSupported

May combine two claims.

Several tools surfaced in the survey (e.g., from Holistic.ai, Credo.ai) were explicitly marketed for use for NYC Local Law 144 compliance, but studies of audit practice suggest the produced measures may not be reliable or legally compatible.

“For instance, several tools surfaced in our survey (e.g., from Holistic.ai, Credo.ai) were explicitly marketed for use for NYC Local Law 144 compliance, but studies of audit practice suggest the produced measures may not be reliable or legally compatible [30, 84, 86].”

• cites [30] [30] Lara Groves, Jacob Metcalf, Alayna Kennedy, Briana Vecchione, and Andrew Strait. 2024. Auditing Work: Exploring the New York City algorithmic bias audit regime. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24). Association for Computing Machinery, New York, NY, USA, (June 2024), 1107-1120. doi: 10.1145/3630106.3658959. - sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [84] [84] Lucas Wright, Roxana Mika Muenster, Briana Vecchione, Tianyao Qu, Pika (Senhuang) Cai, Alan Smith, Comm 2450 Student Investigators, Jacob Metcalf, and J. Nathan Matias. 2024. Null Compliance: NYC Local Law 144 and the challenges of algorithm accountability. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24). Association for Computing Machinery, New York, NY, USA, (June 2024), 1701-1713. doi: 10.1145/3630106.3658998. - sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [86] [86] Alice Xiang and Inioluwa Deborah Raji. 2019. On the Legal Compatibility of Fairness Definitions. (Nov. 2019). arXiv: 1912 . 00761 [cs, stat] . doi: 10.48550/arXiv.1912.00761. - sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Interpretability and explainability methods promoted by popular tools vary widely in their goals and may encourage false confidence in their users.

“Interpretability and explainability methods promoted by popular tools vary widely in their goals [120] and, like many 'snake oil' AI products [18, 59, 91], may encourage false confidence in their users [118].”

• cites [120] [120] Zachary C. Lipton. 2018. The Mythos of Model Interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue , 16, 3, (June 2018), 31-57. doi: 10.1145/3236386.3241340. - sourceSupported
• cites [18] [18] Arvind Narayanan and Sayash Kapoor. 2024. AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference . Princeton University Press, Princeton. isbn: 978-0-691-24913-1. -Not checked yet

  Not checked in the free run.

• cites [59] [59] Frederike Kaltheuner, (Ed.) 2021. Fake AI . Meatspace Press. isbn: 978-1-91382403-7. -Not checked yet

  Not checked in the free run.

• cites [91] [91] Luke Stark and Jevan Hutson. 2021. Physiognomic artificial intelligence. Fordham Intell. Prop. Media & Ent. LJ , 32, 922. - sourceNot checked — no full text of the source
• cites [118] [118] Marzyeh Ghassemi, Luke Oakden-Rayner, and Andrew L. Beam. 2021. The false hope of current approaches to explainable artificial intelligence in health care. The Lancet Digital Health , 3, 11, (Nov. 2021), e745-e750. doi: 10.1016/S25 89-7500(21)00208-9. - sourceSupported

May combine two claims.

Existing tools for data scraping were helpful but often required extensive adaptation.

“Existing tools for data scraping were helpful but often required extensive adaptation:”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

In the dataset, goal-articulating principle statements (N=86) and self-assessment checklists (N=49) were most common, while participatory standard-setting methods (N=5) were rare.

“Most commonly found were goal-articulating 'principle statements' ( 𝑁 = 86), selfassessment checklists ( 𝑁 = 49), and similar documents, while methods for participatory standard-setting ( 𝑁 = 5) were comparatively rare in our dataset.”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Participation also helps instill confidence in the audit process and foster engagement with subsequent accountability efforts.

“Participation also helps instill confidence in the audit process and foster engagement with subsequent accountability efforts.”

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

Some participants thought assessment templates and checklists were focused too narrowly on fairness assessment.

“Some participants also thought that assessments templates and checklists were focused too narrowly on fairness assessment.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Auditors emphasized the importance of standardized evaluation frameworks that provide clarity and consistency.

“Auditors emphasized the importance of standardized evaluation frameworks that provide clarity and consistency: ' I think standardization is a big [concern]... ' (P23)”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Tools for Transparency Infrastructure (N=12) are interfaces and databases hosted by model operators that allow controlled access to relevant data.

“Tools for Transparency Infrastructure ( 𝑁 = 12) are interfaces and databases hosted by model operators that allow controlled access to relevant data.”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Auditors envisioned tools to help them access high-quality, uncompromised data, apply consistent and holistic standards and methods, and ensure audit integrity.

“Auditors envisioned tools to help them access high-quality, uncompromised data, apply consistent and holistic standards and methods, and ensure audit integrity.”

Classified as probably not a substantial claim.

Auditors recognized that to comprehensively identify AI-related harms, they must engage with those directly impacted.

“Auditors recognized that to comprehensively identify AI-related harms, they must engage with those directly impacted.”

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

With recent datasets scaling up to staggering sizes, the concern about data quality has become more acute, and auditors commented that manual analysis was no longer feasible.

“With recent datasets scaling up to staggering sizes, this concern has become more acute and auditors commented on how manual analysis was no longer feasible:”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Multiple participants felt there was a culture or communications gap between legal and compliance people on one side and engineers on the other.

“Multiple participants felt there was a ' culture or communications gap between the legal and compliance people on one side and the engineers on the other ' (P12).”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

Some tools made advances in one or more dimensions of holistic, context-specific, inclusive, and practice-compatible standards, but few accomplished all three.

“Some of the tools we found made advances in one or more of these dimensions, but few accomplished all three.”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

May combine two claims.

NIST's AI RMF includes safety, security, reliability, transparency, explainability, and privacy in addition to fairness, but its guidance for specific evaluation techniques remains fairly broad.

“while NIST's AI RMF includes safety, security, reliability, transparency, explainability, and privacy, in addition to fairness, its guidance for specific evaluation techniques remains fairly broad [17].”

• cites [17] [17] Elham Tabassi. 2023. AI Risk Management Framework: AI RMF (1.0). Tech. rep. error: NIST AI 100-1. National Institute of Standards and Technology, Gaithersburg, MD, error: NIST AI 100-1. doi: 10.6028/NIST.AI.100-1. - sourceNot checked yet

  Not checked in the free run.

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

Data quality concerns intersected with concerns about audit integrity.

“Data quality concerns intersected with concerns about audit integrity.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Some external auditors worried that external data collection tools, particularly data scraping or data donation, may violate terms of service set by platforms and result in legal liability or retaliation.

“Some external auditors also worried that external data collection tools-particularly tools for data scraping or data donation-may violate terms of service set by platforms and result in legal liability or retaliation.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Many auditors wished to streamline the auditing process by offering predefined structures and templates for assessment.

“Many wished to streamline the auditing process by offering predefined structures and templates for assessment, which are essential for conducting audits effectively and facilitating communication.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

While we did find multiple popular databases for recording and collating incidents of harm, these databases record harms after they have already occurred, often rely on second-hand reports, and may not delve deeply into causes of harms or impacts.

“While we did find multiple popular databases for recording and collating incidents of harm (such as the AI Incident Database), these databases record harms after they have already occurred, often rely on second-hand reports, and may not delve deeply into causes of harms or impacts [139].”

• cites [139] [139] Violet Turri and Rachel Dzombak. 2023. Why We Need to Know More: Exploring the State of AI Incident Documentation Practices. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society (AIES '23). Association for Computing Machinery, New York, NY, USA, (Aug. 2023), 576-583. doi: 10.1145/3600211.3604700. - sourceNot checked — no full text of the source

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

Participants recognized the challenge of designing methodologies and tools without exploitation, tokenization, or other forms of 'participation washing'.

“Participants also recognized the challenge of designing methodologies and tools without exploitation, tokenization, or other forms of 'participation washing' [89], highlighting the need for fair compensation and participatory algorithmic development, in addition to participatory auditing.”

• cites [89] [89] Mona Sloane, Emanuel Moss, Olaitan Awomolo, and Laura Forlano. 2022. Participation Is not a Design Fix for Machine Learning. In Proceedings of the 2nd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '22). Association for Computing Machinery, New York, NY, USA, (Oct. 2022), 1-6. isbn: 978-1-4503-9477-2. doi: 10.1145/3551624.3555 285. - sourceNot checked — no full text of the source

May combine two claims.

While there currently exist many tools to support audit work, particularly for evaluating AI systems and managing standards, these tools often fell short of helping auditors achieve accountability in practice.

“We find that while there currently exist many tools to support audit work, particularly for evaluating AI systems and managing standards, these tools often fell short of helping auditors achieve accountability in practice.”

May combine two claims.

Practitioners struggled to independently access high-quality data about system behavior, apply consistent and holistic standards and methods, ensure audit integrity, involve affected stakeholders, and collaborate across disciplines.

“Practitioners struggled to independently access high-quality data about system behavior, apply consistent and holistic standards and methods, ensure audit integrity, involve affected stakeholders, and collaborate across disciplines.”

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

Auditors expressed concerns about legal risks from existing laws and regulations, particularly the Computer Fraud and Abuse Act (CFAA) and other privacy laws.

“Auditors expressed concerns about legal risks from existing laws and regulations, particularly the Computer Fraud and Abuse Act (CFAA) and other privacy laws.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

Practitioners found some tools-such as open source tools for data collection-empowering, but tools for tasks beyond evaluation, such as discovering harms, communicating audit results, and advocating for subsequent changes, were much less common.

“Practitioners found some tools-such as open source tools for data collection-empowering, but tools for tasks beyond evaluation, such as discovering harms, communicating audit results, and advocating for subsequent changes, were much less common.”

May combine two claims.

Despite court orders and regulations like Article 40 of the Digital Service Act that require construction of transparency tools, key APIs used for auditing are becoming more costly and undependable.

“Despite court orders and regulations like Article 40 of the Digital Service Act [47] that require the construction of transparency tools, one participant noted that key APIs used for auditing are becoming more costly and undependable:”

• cites [47] [47] Paddy Leerssen. 2023. Digital Services Act: Summary Report on the Call for Evidence on the Delegated Regulation on Data Access. Tech. rep. European Commission, (Nov. 2023). Retrieved Oct. 22, 2024 from https://digital-strateg y.ec.europa.eu/en/library/digital-services-act-summary-report-call-eviden ce-delegated-regulation-data-access. -Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

Transparency Infrastructure tools provide external auditors with controlled access to models and data, especially for online platforms in the dataset, but they require investment from model operators.

“Transparency Infrastructure tools provide external auditors with controlled access to models and data-especially for online platforms, in our dataset (Fig. E.4)-but they require investment from model operators.”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

May combine two claims.

While there exist many tools for aiding AI auditing, practitioners found existing tools inadequate in multiple ways.

“While there exist many tools for aiding AI auditing, practitioners found existing tools inadequate in multiple ways.”

May combine two claims.

Tools to facilitate Audit Communication were the rarest in the dataset (N=8), consisting mostly of tools for Dataset Visualization and Audit Reporting.

“Tools to facilitate Audit Communication were the rarest in our dataset ( 𝑁 = 8), and consist mostly of tools for Dataset Visualization (e.g. Google's FACETS [56]) and Audit Reporting (e.g., the ACLU's repository of NYC Local Law 144 hiring bias audit reports [127]).”

• cites [56] [56] 2023. Facets - Know Your Data. (Dec. 2023). Retrieved Dec. 8, 2023 from https://pair-code.github.io/facets/. -Not checked yet

  Not checked in the free run.

• cites [127] [127] Melissa Gerchick and Brooke Watson Madubuonwu. 2024. Tracking Automated Employment Decision Tool Bias Audits. aclu-national. (Aug. 2024). Retrieved Sept. 10, 2024 from https://github.com/aclu-national/tracking-ll144 -bias-audits. -Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Auditors hoped greater communication could help unite the profession around policy developments, shared language, standards, and goals that could improve the impact of their work.

“Auditors hoped greater communication could help unite the profession around policy developments, shared language, standards, and goals that could improve the impact of their work.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

AI auditors care deeply about accountability but struggled to achieve it.

“This echoes a theme repeated several times throughout our interviews-AI auditors care deeply about accountability but struggled to achieve it.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

The most popular Performance Analysis tool found on Github is SHAP, a game-theoretic method for measuring feature importance in a model.

“For example, the most popular Performance Analysis tool we found on Github is SHAP (SHapley Additive exPlanations), a gametheoretic method for measuring feature importance in a model [9] (see Table E.2).”

• cites [9] [9] Scott M. Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17). Curran Associates Inc., Red Hook, NY, USA, (Dec. 2017), 4768-4777. isbn: 978-1-5108-6096-4. Retrieved Feb. 12, 2025 from. -Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Of the 347 organizations in the dataset, Crunchbase records were accessible for 202 (accounting for 270 tools); 173 of these records include estimated employee counts.

“Of the 347 organizations in our dataset, we were able to access Crunchbase records for 202 (accounting for 270 tools); 173 of these records include estimated employee counts.”

• cites [44] [44] 2023. Crunchbase. (2023). Retrieved Aug. 24, 2023 from https://www.crunchb ase.com. -Not checked yet

  Not checked in the free run.

May combine two claims.

Performance Analysis tools (N=129) are designed to help auditors evaluate and explain model behavior, usually through calculation of quantitative metrics related to accuracy/safety (N=71), explainability (N=36), or fairness (N=30).

“Tools for Performance Analysis ( 𝑁 = 129) are designed to help auditors evaluate and explain model behavior, usually through the calculation of quantiative metrics related to accuracy/safety ( 𝑁 = 71), explainability ( 𝑁 = 36), or fairness ( 𝑁 = 30).”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

Classified as probably not a substantial claim.

Participants' aspirations for tooling envision a path towards lasting public infrastructure that gives auditors additional levers to hold model operators accountable.

“participants' aspirations for tooling envision another path-a path towards lasting public infrastructure that gives auditors additional levers to hold model operators accountable [96].”

• cites [96] [96] Nik Marda, Jasmine Sun, and Mark Surman. 2024. Public AI: Making AI Work for Everyone, by Everyone. Tech. rep. Mozilla Foundation, (Sept. 2024). -Not checked yet

  Not checked in the free run.

May combine two claims.

The stages of harms discovery, audit communication, and advocacy are the stages participants described as most requiring contextual awareness and typically under-studied participatory and community engagement methods.

“Perhaps unsurprisingly, these are also the stages of the audit process that participants described as most requiring contextual awareness and typically under-studied participatory and community engagement methods.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Currently, the burden of building and maintaining tools to address technical debt falls on the auditors, raising concerns about the sustainability of auditing efforts.

“Currently, the burden of building and maintaining tools to address technical debt falls on the auditors, raising concerns about the sustainability of auditing efforts.”

May combine two claims.

The external auditors interviewed were especially skeptical of the data provided from these tools and aware of their unreliability.

“The external auditors we interviewed were especially skeptical of the data provided from these tools and aware of their unreliability”

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

OAT team members presented these findings to regulators at the U.K.'s OfCom and the U.S. Federal Trade Commission.

“OAT team members presented these findings to regulators at the U.K.'s OfCom and the U.S. Federal Trade Commission.”

May combine two claims.

Performance Analysis tools found, most popular of which were built and maintained by disproportionately large for-profit firms, often focused on a narrow set of technical fairness definitions and explainability methods popularized in academic literature.

“The Performance Analysis tools we foundthe most popular of which were built and maintained by disproportionately large, for-profit firms (Figures E.6)-often focused on a narrow set of technical fairness definitions and explainability methods popularized in academic literature.”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

May combine two claims.

Performance Analysis tools found (48.1% built by for-profit organizations) were disproportionately not open source compared to other tools in the dataset, especially tools for Explainability.

“However, the Performance Analysis tools we found (48.1% of which were built by for-profit organizations) were disproportionately not open source compared to other tools in our dataset, especially tools for Explainability (Fig. E.3).”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Auditors envisioned audit report databases as accountability tools to facilitate amalgamation and communication of audit findings to key stakeholders.

“Likewise, auditors envisioned audit report databases as accountability tools to facilitate the amalgamation and communication of audit findings to key stakeholders, mirroring interventions such as the U.S. Security and Exchange Commission (SEC) EDGAR database for financial accounting audits [88].”

• cites [88] [88] Inioluwa Deborah Raji, Peggy Xu, Colleen Honigsberg, and Daniel Ho. 2022. Outsider Oversight: Designing a Third Party Audit Ecosystem for AI Governance. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society (AIES '22). Association for Computing Machinery, New York, NY, USA, (July 2022), 557-571. isbn: 978-1-4503-9247-1. doi: 10.1145/3514094.3534181. - sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Auditors especially wished for resources that would bring together the diverse, interdisciplinary groups involved in auditing.

“Auditors especially wished for resources that would bring together the diverse, interdisciplinary groups involved in auditing, similar to the few tools for Community Spaces we found in our tool survey.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

Classified as probably not a substantial claim.

Participants noted a lack of long-term investment.

“participants noted a lack of long-term investment”

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

OAT team members participated in advising the selection of two rounds of Mozilla Technology Fund awardees, including 5 teams in an inaugural cohort (2022), and 8 teams in a follow-up cohort (2023) focused on AI audit tooling.

“OAT team members also participated in advising the selection of two rounds of Mozilla Technology Fund (MTF) awardees, including 5 teams in an inaugural cohort (2022), and 8 teams in a follow-up cohort (2023) focused on AI audit tooling.”

Classified as probably not a substantial claim.

Participants agreed on the need for shared infrastructure that supports the auditing process.

“Participants agreed on the need for shared infrastructure that supports the auditing process.”

Classified as probably not a substantial claim.

Some participants were concerned about the efficacy of many AI audit tools, particularly tools whose methods were not made available for public scrutiny.

“Some participants were concerned about the efficacy of many AI audit tools-particularly tools whose methods were not made available for public scrutiny.”

May combine two claims.

The authors found more tools for Advocacy (N=14) than for Audit Communication but still fewer than in other stages of the taxonomy.

“We found more tools for Advocacy ( 𝑁 = 14)-including Community Spaces (e.g., the Benefits Tech Advocacy Hub [32], which facilitates collaboration between advocates who oppose algorithmbased cuts to public benefits), tools for Organizing/Resistance (e.g., the Algorithmic Ecology framework [112], a tool for mapping the non-technical dimensions of algorithmic impact), and tools for Legal Search (e.g.”

• cites [32] [32] 2023. Benefits Tech Advocacy Hub. (Dec. 2023). Retrieved Dec. 12, 2023 from https://btah.org/. -Not checked yet

  Not checked in the free run.

• cites [112] [112] Stop LAPD Spying Coalition and Free Radicals. 2020. The Algorithmic Ecology: An Abolitionist Tool for Organizing Against Algorithms. (Mar. 2020). Retrieved Dec. 8, 2023 from https://freerads.org/2020/03/02/the-algorithmicecology-an-abolitionist-tool-for-organizing-against-algorithms/. -Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Several participants emphasized the importance of open-sourcing tools for others in the community to inspect.

“To allay methodological concerns, several participants emphasized the importance of open-sourcing tools for others in the community to inspect.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.
• AssumptionRests on an assumption: Open-sourcing audit tools improves their inspectability and reproducibility.

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

Total venture funding (adjusted to U.S. dollars) was collected for 48 firms out of the 112 firms that are still private (i.e., have not undertaken an initial public offering).

“We also collected total venture funding (adjusted to U.S. dollars) for 48 firms out of the 112 firms that are still private (i.e., have not undertaken an initial public offering).”

• cites [44] [44] 2023. Crunchbase. (2023). Retrieved Aug. 24, 2023 from https://www.crunchb ase.com. -Not checked yet

  Not checked in the free run.

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

Some participants attributed methodological deficiencies to differences in rigor employed by the various disciplines involved with auditing.

“Some participants put methodological deficiencies down to differences in the rigor employed by the various disciplines involved with auditing.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

Tools specifically devoted to qualitative, as opposed to quantitative, analysis were much harder to find (N=3) and rarely mentioned by participants.

“For example, tools specifically devoted to qualitative-as opposed to quantitative-analysis were much harder to find ( 𝑁 = 3) and rarely mentioned by our participants.”

• AssumptionRests on an assumption: The authors' tool dataset is comprehensive enough to support claims about gaps in AI audit tooling.
• AssumptionRests on an assumption: The taxonomy's stages and categories are valid for classifying AI audit tools.

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

Participants had concerns about maintenance, effectiveness of automated monitoring processes, and the efficacy of synthetic data for representing real users.

“Participants had concerns about maintenance, effectiveness of automated monitoring processes, and the efficacy of synthetic data for representing real users instead of functioning as ' an academic exercise ' (P4):”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

Participants had audit results blocked from publication (P1) or unduly restricted in scope (P13) due to interventions from audit targets.

“Participants had audit results blocked from publication (P1) or unduly restricted in scope (P13) due to interventions from audit targets.”

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

In a survey of 152 audit practitioners, 62% used existing tools like AI Fairness 360, Scikit Fairness, or Parity, though only 7% used a standardized framework for their overall audit protocol.

“And in a survey of 152 audit practitioners, Costanza-Chock et al. [136] found that 62% of practitioners used existing tools like AI Fairness 360, Scikit Fairness, or Parity, though only 7% of respondents used a standardized framework for their overall audit protocol.”

• cites [136] [136] Sasha Costanza-Chock, Inioluwa Deborah Raji, and Joy Buolamwini. 2022. Who Audits the Auditors? Recommendations from a field scan of the algorithmic auditing ecosystem. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22). Association for Computing Machinery, New York, NY, USA, (June 2022), 1571-1583. isbn: 978-1-4503-9352-2. doi: 10.1145/3531146.3533213. - sourceNot checked yet

  Not checked in the free run.

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

Gathering empirical evidence is a key step in AI auditing but often poses the most significant challenge in practice.

“Gathering empirical evidence is a key step in AI auditing, but often poses the most significant challenge in practice.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

May combine two claims.

Across various industries, the common accountability practice of auditing has long relied on rituals, organizational processes and tools in order to make consistent, inter-operable and reliable judgments, as well as provide broader access to a larger range of stakeholder participants in the audit process.

“Across various industries, the common accountability practice of auditing has long relied on rituals, organizational processes and tools in order to make consistent, inter-operable and reliable judgments [113], as well as provide broader access to a larger range of stakeholder participants in the audit process [92].”

• cites [113] [113] Michael Power. 1999. The Audit Society . Oxford University Press, (Aug. 1999). isbn: 978-0-19-829603-4. doi: 10.1093/acprof:oso/9780198296034.001.0001. - sourceNot checked — no full text of the source
• cites [92] [92] AbebaBirhane, William Isaac, Vinodkumar Prabhakaran, Mark Diaz, Madeleine Clare Elish, Iason Gabriel, and Shakir Mohamed. 2022. Power to the People? Opportunities and Challenges for Participatory AI. In Proceedings of the 2nd - ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '22). Association for Computing Machinery, New York, NY, USA, (Oct. 2022), 1-8. isbn: 978-1-4503-9477-2. doi: 10.1145/3551624.3555290. -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).

In the E.U., enforcement reports for the Digital Services Act emphasize the importance of AI audit tooling for effective oversight enforcement.

“Similarly, in the E.U., enforcement reports for the Digital Services Act emphasize the importance of AI audit tooling for effective oversight enforcement [45].”

• cites [45] [45] Ulrike Klinger and Jakob Ohme. 2023. Delegated Regulation on Data Access Provided for the Digital Services Act: Response to the Call for Evidence DG CNECT-CNECT F2 by the European Commission. Tech. rep. 7. Weizenbaum Institute, (May 2023). https://www.weizenbaum-library.de/handle/id/380. -Not checked yet

  Not checked in the free run.

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

This task can be especially difficult for external auditors who may not know where AI systems are in use or what their impacts might be.

“This task can be especially difficult for external auditors who may not know where AI systems are in use or what their impacts might be.”

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

Accountability requires infrastructure.

“Accountability also requires infrastructure.”

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

Fairness toolkits do not typically support other necessary steps of an audit such as data collection.

“And fairness toolkits do not typically support other necessary steps of an audit such as data collection.”

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

Our theoretical sampling strategy deliberately over-represents some less common kinds of tooling (such as tools for advocacy).

“In particular, our theoretical sampling strategy deliberately over-represents some less common kinds of tooling (such as tools for advocacy).”

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

Our dataset should not be viewed as a comprehensive or perfectly representative sample.

“As a result, while our dataset reflects a structured and iterative search, it should not be viewed as a comprehensive or perfectly representative sample.”

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

Audit tooling represents not only a mechanism for maintenance and consistency of audit integrity but also a key capacity-building intervention to lower the barriers for broader participation.

“Audit tooling, then, represents not only a mechanism for maintenance and consistency of audit integrity but also a key capacity-building intervention to lower the barriers for broader participation.”

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

In the U.S., several recently proposed 'AI innovation' bills are explicitly geared towards investing in tooling and resource development for AI auditing.

“In the U.S. alone, several recently proposed 'AI innovation' bills are explicitly geared towards investing in tooling and resource development for AI auditing.”

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

The vetting process for API use often requires the auditor to disclose their intent for the evaluation in advance, which may compromise the integrity of the study.

“And currently, the vetting process for API use often requires the auditor to disclose their intent for the evaluation in advance, which may compromise the integrity of the study.”

• AssumptionRests on an assumption: The interview participants' statements reflect broader AI auditing practice.

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

Auditing an AI system first requires identifying the system that should be subject to scrutiny and identifying its potential harms.

“Auditing an AI system first requires identifying the system that should be subject to scrutiny and identifying its potential harms.”

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

Governments across the globe are developing their own tools and making use of existing resources as part of their enforcement regimes for AI governance.

“Governments across the globe are developing their own tools and making use of existing resources as part of their enforcement regimes for AI governance [101].”

• cites [101] [101] Kate Kaye and Pam Dixon. 2023. Risky Analysis: Assessing and Improving AI Governance Tools. Tech. rep. World Privacy Forum, (Nov. 2023). Retrieved Dec. 18, 2023 from https://www.worldprivacyforum.org/wp-content/uploads /2023/12/WPF\_Risky\_Analysis\_December\_2023\_fs.pdf. -Not checked yet

  Not checked in the free run.

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

Our search for tools was systematic but not exhaustive.

“So while our search for tools was systematic, it was not exhaustive.”

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

One author created the labels and at least one other author reviewed each label for agreement.

“One author created the labels and at least one other author reviewed each label for agreement.”

• Also stated in: § E Additional Landscape Analysis · p. 19 · plain

  “One author created the labels and at least one other author reviewed each label for agreement.”

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

A comprehensive survey of auditors' practical needs relative to the landscape of available tools is lacking.

“As yet, however, a comprehensive survey of auditors' practical needs relative to the landscape of available tools is lacking.”

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

Our protocol was approved by three university IRBs.

“Our protocol was approved by three university IRBs.”

Classified as probably not a substantial claim.

Practitioners often rely on tools-software, frameworks, and other resources-to support their AI audit work.

“In the face of these challenges, practitioners often rely on toolssoftware, frameworks, and other resources-to support their AI audit work.”

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

Embracing public evaluation results helps audit practitioners to learn from each other's experiences.

“Embracing public evaluation results also helps audit practitioners to learn from each other's experiences.”

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

AI audit studies translating into tangible outcomes of accountability is far from certain.

“Ideally, AI audit studies will translate into tangible outcomes of accountability, but this outcome is far from certain.”

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

There are substitutes and competitors for many of the tools in the dataset, but they were not included unless observed in the sources or subsequent searches.

“Note that there are substitutes and competitors for many of the tools in our dataset (such as qualitative coding software), but we did not include them unless we observed them in one of the sources above or our subsequent searches (§ C.2).”

Classified as probably not a substantial claim.

The analysis is primarily scoped to the U.S. and the E.U.

“We thus recognize that our analysis is primarily scoped to the U.S. and the E.U.”

May combine two claims.

Open source tools (e.g., for data scraping) can sometimes fill the gap, but do not always cover the scope and complexity of practical audit work.

“Open source tools (e.g., for data scraping) can sometimes fill the gap, but do not always cover the scope and complexity of practical audit work.”

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

Communicating audit findings, lessons, and insights learned can help build trust and validate audit findings, recommendations, and subsequent interventions.

“Communicating audit findings, lessons, and insights learned can help build trust and validate audit findings, recommendations, and subsequent interventions.”

May combine two claims.

Making AI audit tools publicly available enables both external collaboration and third-party validation.

“Making AI audit tools publicly available enables both external collaboration and third-party validation.”

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

Audit tools like these can hold great power over the audit process.

“Audit tools like these can hold great power over the audit process:”

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

The dataset of AI audit tools is not an exhaustive list of all tools that have been or could be used in AI audit practice.

“Our dataset is not an exhaustive list of all tools that have been or could be used in AI audit practice.”

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

Audit report repositories could expand the forum holding model operators accountable, allowing policymakers, journalists, and other public stakeholders to engage with evaluations more easily.

“Audit report repositories could expand the forum holding model operators accountable, allowing policymakers, journalists, and other public stakeholders to engage with evaluations more easily.”

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

Categories with a more saturated set of examples were not given an equal amount of additional sourcing.

“Categories with a more saturated set of examples were not given an equal amount of additional sourcing.”

Classified as probably not a substantial claim.

A point of saturation for each search was determined by examining each Google search page of 30 results at a time until two pages in a row contained no references to specific audit tools.

“In order to determine a point of saturation for each search, we examined each Google search page of 30 results at a time until two pages in a row contained no references to specific audit tools.”

May combine two claims.

Unreliable performance or accuracy analysis tools that fail to meaningfully assess the audit target operate as misleading 'rubber stamps' for vendors and lead to 'audit washing,' posing a serious challenge to the legitimacy of audit results.

“Unreliable performance or accuracy analysis tools that fail to meaningfully assess the audit target operate as misleading 'rubber stamps' for vendors and lead to 'audit washing' [13], posing a serious challenge to the legitimacy of audit results.”

• cites [13] [13] Ellen P Goodman and Julia Tréhu. 2022. AI Audit-Washing and Accountability. Policy Paper. German Marshall Fund of the United States, (Nov. 2022). https: //www.gmfus.org/sites/default/files/2022-11/Goodman%20%26%20Trehu%2 0-%20Algorithmic%20Auditing%20-%20paper.pdf. - sourceNot checked — no full text of the source

May combine two claims.

Facebook initially insinuated that disabling the researchers' accounts was required by an FTC consent decree, backtracking only after the FTC called the statement inaccurate.

“Facebook initially insinuated that disabling the researchers' accounts was required by an Federal Trade Commission (FTC) consent decree, backtracking only after the FTC called the statement inaccurate [76].”

• cites [76] [76] Samuel Levine. 2021. Letter from Acting Director of the Bureau of Consumer Protection Samuel Levine to Facebook. (Aug. 2021). Retrieved Dec. 7, 2023 from https://www.ftc.gov/blog-posts/2021/08/letter-acting-director-bureau -consumer-protection-samuel-levine-facebook. -Not checked yet

  Not checked in the free run.

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

The analysis may not be representative of the global AI auditing landscape or appropriate for or informed by other contexts.

“and may not be representative of the global AI auditing landscape or appropriate for or informed by other contexts.”

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

Open tooling practices may contribute to knowledge-sharing, standards-setting, transparency, accessibility, and trust.

“Open tooling practices may also contribute to knowledge-sharing, standards-setting, transparency, accessibility, and trust”