Positioning Paradata: A Conceptual Frame for AI Processual Documentation in Archives and Recordkeeping Contexts

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

XAI literature assumes that particular user groups need explanations designed for their domain-specific needs.

“XAI literature operates on the premise that given user groups will need explanations that are designed to meet their domain-specific needs [ 5 , 61 ].”

• cites [5] 5. Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion 58 (2020), 82-115. DOI: https://doi.org/10.1016/j.inffus.2019.12.012 sourceSupported
• cites [61] 61. Giulia Vilone and Luca Longo. 2020. Explainable Artificial Intelligence: A systematic review. arXiv:2006.00093 (2020). DOI: https://doi. org/10.48550/arXiv.2006.00093 sourceSupported

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

A piece of information understood as paradata at one moment may serve as metadata, provenance, or contextual information from another perspective.

“As Huvila's middle perspective notes, a piece of information that may be understood as paradata at one moment may serve as metadata, provenance, or contextual information from another perspective.”

• cites [31] 31. Isto Huvila. 2022. Improving the usefulness of research data with better paradata. Open Information Science 6, 1 (2022), 28-48. DOI: https: //doi.org/10.1515/opis-2022-0129 sourceSupported

May combine two claims.

Black box tools can generate wanted outputs from inputs but cannot explain the reasoning behind their conclusions.

“Black box tools may produce the desired outputs from the provided inputs but are incapable of providing an explanation of the reasoning they employ in reaching their conclusion.”

• cites [61] 61. Giulia Vilone and Luca Longo. 2020. Explainable Artificial Intelligence: A systematic review. arXiv:2006.00093 (2020). DOI: https://doi. org/10.48550/arXiv.2006.00093 sourceSupported

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

Huvila views paradata as a subset of metadata but advocates a dynamic middle perspective in which a piece of information may appear as paradata, metadata, or provenance depending on the question.

“Although he does view paradata as a subset of metadata, Huvila advocates for a dynamic middle perspective (as illustrated in Section 2 of Figure 1 ) toward paradata, noting that when addressing a particular question, a given piece of information may appear as part of one or more of an object's paradata, metadata, or provenance.”

• cites [31] 31. Isto Huvila. 2022. Improving the usefulness of research data with better paradata. Open Information Science 6, 1 (2022), 28-48. DOI: https: //doi.org/10.1515/opis-2022-0129 sourcePartly supported

  Section 3, “What is paradata?” (page unavailable), proposes a middle-range perspective and says whether information functions as paradata depends on perspective; the same piece can function as both metadata and paradata. It calls paradata a subset of contextual information, not a subset of metadata, and does not explicitly say that a piece of information may function as provenance depending on the question. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

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

Huvila notes increasing use of paradata for research process data across fields and a lack of standardization for defining, collecting, maintaining, and providing access to a dataset's paradata.

“Huvila notes the increasing use of the term to refer to research process data across multiple fields, and the emerging problem of a lack of standardization for how to define, collect, maintain, and make accessible a dataset's paradata.”

• cites [31] 31. Isto Huvila. 2022. Improving the usefulness of research data with better paradata. Open Information Science 6, 1 (2022), 28-48. DOI: https: //doi.org/10.1515/opis-2022-0129 sourceSupported

May combine two claims.

Complex machine-learning techniques offer the greatest analytical sophistication and the least transparency, reflecting a trade-off between interpretability and performance.

“complex ML techniques offer the greatest sophistication of analysis and, conversely, the least transparency in the inevitable”

• cites [5] 5. Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion 58 (2020), 82-115. DOI: https://doi.org/10.1016/j.inffus.2019.12.012 sourcePartly supported

  The Introduction (p. 2) describes complex deep neural networks as opaque and identifies a trade-off between performance and transparency. But section 5.1 (p. 30) says more complex models are not necessarily more accurate and that explainability methods may improve the trade-off, so the claim's categorical ‘greatest’ and ‘inevitable’ framing goes beyond the source. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

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

The importance of context in explanation and the question of whom an explanation is for have become key themes in machine learning and XAI literature.

“The importance of context in any explanation-and the attendant question of”

• cites [10] 10. Umang Bhatt, Mckane Andrus, Adrian Weller, and Alice Xiang. 2020. Machine learning explainability for external stakeholders. arXiv:2007.05408 (2020). sourceSupported

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

Any explanation must address its own epistemic limits.

“Furthermore, any explanation must address its own epistemic limits.”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourcePartly supported

  Section 2.4 says explainable AI systems should identify and declare when an answer falls outside their design or is unreliable; Section 9 likewise says the system should express its knowledge limits. The paper supports acknowledging limits in this setting, but does not require every explanation, in general, to address its own epistemic limits. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

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

Self-interpretable models such as decision trees or rule lists let users trace the logical paths from input to output.

“AI models reliant on decision trees or rule lists allow users to walk through the logical paths that lead a given input through the process of it becoming an output.”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

Lynch notes that the archivist's job may become reconstructing algorithm logic through reverse-engineering, the same process NIST describes for reconstructing global models of AI tools.

“Perceptively, Lynch notes that the archivist's job may become an effort to reconstruct the logic of algorithms through reverse-engineering their logic, the exact process that NIST describes as being quite common in reconstructing global models of AI tools [ 48 ].”

• cites [40] 40. Clifford Lynch. 2017. Stewardship in the "Age of Algorithms." First Monday 22, 12 (2017). DOI: https://doi.org/10.5210/fm.v22i12.8097 sourceNot checked — no full text of the source
• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

Ashby theorized methods for reconstructing black box logic through experimentation as early as 1956.

“Ashby [ 3 ] theorizes similar methods for reconstructing the logic of black box systems through experimentation as early as 1956.”

• cites [3] 3. W. Ross Ashby. 1956. An Introduction to Cybernetics . Chapman & Hall Ltd., London, England. http://pcp.vub.ac.be/books/IntroCyb.pdf. sourceSupported

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

Tim Miller argues that AI explanations should be situated within social and human sciences literature on cognition and decision making because XAI problems are fundamentally about relationships between people rather than between people and technology.

“Tim Miller argues for the situation of AI explanations within social and human sciences literature on cognition and decision making, given that the problems XAI addresses are fundamentally those of the relationships between people rather than people and technology [ 43 ].”

• cites [43] 43. Tim Miller. 2019. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence 267 (2019), 1-38. DOI: https://doi.org/10.1016/j.artint.2018.07.007 sourcePartly supported

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

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

NIST argues that AI tools should succeed or fail on the same metrics by which humans performing the same tasks succeed or fail.

“NIST argues that AI tools should succeed or fail on the same metrics by which humans performing the same tasks succeed or fail [ 48 ].”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourcePartly supported

  The Introduction and Section 8 use human performance as a baseline for comparing explainable AI, and the Discussion applies the same explainability principles to humans and AI. But Section 8 explicitly leaves open whether machines should be held to a higher standard. The paper does not argue that AI tools should succeed or fail by the same metrics as humans across the tasks they perform. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

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

Arrieta et al. propose that the specific audience become part of the definition of XAI.

“Arrieta et al. propose that the specific audience become part of the definition they propose for XAI:”

• cites [5] 5. Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion 58 (2020), 82-115. DOI: https://doi.org/10.1016/j.inffus.2019.12.012 sourceSupported

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

The risk of emergent bias, by which a tool produces biased outputs through its own functioning, is documented in the literature.

“The risk of”

• cites [25] 25. Batya Friedman and Helen Nissenbaum. 1996. Bias in computer systems. ACM Transactions on Information Systems 14, 3 (1996), 330-347. DOI: https://doi.org/10.1145/230538.230561 sourceSupported

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

Many current local explanation models work by varying actual data to identify which aspects of a dataset or item were decisive in an AI tool's actions.

“many of the current models operate by providing variations of actual data submitted within a given sample to identify which aspects of a dataset or an item provided the decisive factors leading up to an AI tool's actions.”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

Because audiences vary endlessly, an explanation must be contextually situated to be meaningful, and no explanation is universal.

“With the endless possible variation of audiences, an explanation must be contextually situated to be meaningful.No explanation will be universal [ 5 ].”

• cites [5] 5. Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion 58 (2020), 82-115. DOI: https://doi.org/10.1016/j.inffus.2019.12.012 sourceSupported

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

Well-understood models may outperform opaque ones because users can better understand, assess, and fine-tune an application for its context.

“models that are well understood may perform better than those that are opaque, given users' increased ability to understand, assess, and fine-tune a given application for its context.”

• cites [5] 5. Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion 58 (2020), 82-115. DOI: https://doi.org/10.1016/j.inffus.2019.12.012 sourceSupported

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

According to NIST, explanations must consider their intended audience, and jargon-laden outputs incomprehensible to laypersons are inadequate when the process must be comprehensible to non-specialists.

“According to NIST, explanations must consider their intended audience.Jargon-laden or encoded outputs incomprehensible to laypersons will not do if the process must be comprehensible to non-specialists.”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

At a fine enough level, even riding a bicycle becomes a black box problem for Ashby because few people can explain the subatomic forces involved.

“At a fine enough level of granularity, a task as basic as riding a bicycle becomes a black box problem for Ashby, as very few of us could explain the operation of forces at a subatomic level propelling our motion forward [ 3 ].”

• cites [3] 3. W. Ross Ashby. 1956. An Introduction to Cybernetics . Chapman & Hall Ltd., London, England. http://pcp.vub.ac.be/books/IntroCyb.pdf. sourceSupported

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

Global models approximate the entirety of a model and show how input variables lead to different outputs, often reconstructed from multiple local explanations and visualizations.

“Global models approximate the entirety of a model and tend to illustrate how variables in the inputs lead a model to different outputs.Often this information has been reconstructed from multiple local explanations and may be represented through visualizations [ 48 ].”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

NIST published four principles for explainable AI and a list of strategies XAI models could use to open the black box.

“In 2021, the National Institute of Standards and Technology (NIST) published a summary of four principles that might characterize truly”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

Adler et al. suggested carefully testing input variables to trace the indirect influence of each as a source of unintended racial bias.

“Adler et al. [ 1 ] have suggested carefully testing input variables to trace the indirect influence of each as a source of unintended racial bias.”

• cites [1] 1. Philip Adler, Casey Falk, Sorelle A. Friedler, Gabriel Rybeck, Carlos Scheidegger, Brandon Smith, and Suresh Venkatasubramanian. 2018. Auditing black-box models for indirect influence. Knowledge and Information Systems 54, 1 (2018), 95-122. DOI: https://doi.org/ 10.48550/arXiv.1602.07043 sourceSupported

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

NIST categorizes XAI approaches into self-interpretable models and post hoc explanations, with post hoc explanations divided into local and global types.

“NIST goes on to categorize the approaches that XAI tools might take to produce explanations for AI systems' actions.Self-interpretable models are AI tools in which the tool itself may serve as an explanation of its own workings.”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

May combine two claims.

A medical machine-learning tool can recognize cancerous growth from images as verified by physicians, but it cannot identify the indicators that suggested abnormalities, leaving physicians unable to refine their techniques, refine the model, or understand its limits.

“one medical ML tool successfully recognizes the signs of cancerous growth from images of a patient's body, as verified by further examination by physicians; however, the same tool is unable to identify the indicators that suggested the presence of abnormalities within its samples.”

• cites [22] 22. Juan Manuel Durán and Karin Rolanda Jongsma. 2021. Who is afraid of black box algorithms? On the epistemological and ethical basis of trust in medical AI. Journal of Medical Ethics 47, 5 (2021), 329-335. DOI: https://doi.org/10.1136/medethics-2020-106820 sourcePartly supported

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

  Experimental — accuracy not yet measured

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

The black box problem is not new; Ashby's 1956 cybernetics treated it as a logical and ontological issue.

“An early approach to the black box is provided in Ross Ashby's 1956 Introduction to Cybernetics , which poses the problem as a logical and ontological one.”

• cites [3] 3. W. Ross Ashby. 1956. An Introduction to Cybernetics . Chapman & Hall Ltd., London, England. http://pcp.vub.ac.be/books/IntroCyb.pdf. sourceSupported

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

Drawing on data visualization methods is increasingly common for XAI practitioners to represent processes to diverse nonspecialized audiences, recognizing that explanations need not be only words.

“Drawing on data visualization methods is an increasingly common method by which XAI practitioners represent their processes to diverse nonspecialized audiences, recognizing that explanations need not be in words only [ 61 ].”

• cites [61] 61. Giulia Vilone and Luca Longo. 2020. Explainable Artificial Intelligence: A systematic review. arXiv:2006.00093 (2020). DOI: https://doi. org/10.48550/arXiv.2006.00093 sourceSupported

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

Bhatt et al. caution ML developers to consider who the stakeholders in a dataset might be and what questions they may have before devising an explanation model, because explainability tools cannot be developed without regard to deployment context.

“Bhatt et al. caution ML developers to consider who the stakeholders in a given dataset might be, and to ask what questions they may be interested in before devising an explanation model:”

• cites [10] 10. Umang Bhatt, Mckane Andrus, Adrian Weller, and Alice Xiang. 2020. Machine learning explainability for external stakeholders. arXiv:2007.05408 (2020). sourceSupported

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

A recent literature review divided AI applications in archives according to their role in the creation, capture, organize, and pluralize stages of a record's life under the records continuum model.

“A recent literature review has divided AI applications in archives according to their role in the creation, capture, organize, and pluralize stages of a record's life under the records continuum model [ 15 , 60 ].”

• cites [15] 15. Giovanni Colavizza, Tobias Blanke, Charles Jeurgens, and Julia Noordegraaf. 2021. Archives and AI: An overview of current debates and future perspectives. Journal on Computing and Cultural Heritage 15, 1 (2021), 1-15. DOI: https://doi.org/10.1145/3479010 sourceSupported
• cites [60] 60. Frank Upward. 2005. The records continuum. In Archives: Recordkeeping in Society , Sue McKemmish, Michael Piggott, Barbara Reed, and Frank Upward (Eds.). Chandos Publishing, Wagga Wagga, New South Wales, 197-222. DOI: https://doi.org/10.1016/B978-1-876938- 84-0.50008-1 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The black box problem occurs when AI systems have complex, non-linear structures that are extremely hard to interpret and explain to laypeople.

“the black box problem emerges when AI systems'”

• cites [61] 61. Giulia Vilone and Luca Longo. 2020. Explainable Artificial Intelligence: A systematic review. arXiv:2006.00093 (2020). DOI: https://doi. org/10.48550/arXiv.2006.00093 sourceSupported

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

Local explanations clarify one or more individual decisions, while global explanations recreate an approximate model of the non-interpretable tool.

“Within this category, local explanations elucidate one or more individual decisions, whereas global explanations re-create an approximate model of the non-interpretable tool.”

• cites [48] 48. P. Jonathon Phillips, Carina A. Hahn, Peter C. Fontana, Amy N. Yates, Kristen Greene, David A. Broniatowski, and Mark A. Przybocki. 2021. Four Principles of Explainable Artificial Intelligence . National Institute of Standards and Technology, Gaithersburg, MD. DOI: https: //doi.org/10.6028/NIST.IR.8312 sourceSupported

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

Jenkinson identified archival records' quality of not having been produced with posterity and future readers in mind as testimony to their authenticity and integrity.

“Archival eminence grise Hilary Jenkinson, for instance, identified archival records' quality of not having been produced with posterity and future readers in mind as a testimony to their authenticity and integrity.”

• cites [34] 34. Hilary Jenkinson. 1937. A Manual of Archive Administration . P. Lund, Humphries & Co., Ltd., London, England. http://archive.org/ details/manualofarchivea00iljenk . sourceNot checked — no full text of the source
• cites [56] 56. Richard Stapleton. 1983. Jenkinson and Schellenberg: A comparison. Archivaria 1983 (1983), 75-85.Partly supported

  In “A Comparison” (page unavailable), the paper says Jenkinson regarded records not made for later research purposes as impartial. It attributes their authenticity instead to responsible custody, which makes forgery or falsification exceptional. The claim captures his point about intended readers but conflates impartiality with authenticity and integrity. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

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

For Jenkinson, impartiality distinguished archival records from documents encountered outside archival contexts.

“For Jenkinson, this”

• cites [34] 34. Hilary Jenkinson. 1937. A Manual of Archive Administration . P. Lund, Humphries & Co., Ltd., London, England. http://archive.org/ details/manualofarchivea00iljenk . sourceNot checked — no full text of the source
• cites [56] 56. Richard Stapleton. 1983. Jenkinson and Schellenberg: A comparison. Archivaria 1983 (1983), 75-85.Supported

Classified as probably not a substantial claim.

Paradata needs may include preservation of novel forms of information such as datasheets explaining training datasets or model cards documenting an AI tool, and of these documents' relation to archival collections processed using them.

“Paradata needs may include the preservation of novel forms of information, such as datasheets explaining training datasets or model cards documenting an AI tool, and of these documents' relation to archival collections processed using them [ 26 ].”

• cites [26] 26. Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2021. Datasheets for datasets. arXiv:1803.09010 (2021). sourcePartly supported

  The Introduction (p. 1) proposes datasheets documenting datasets, and Impact and Challenges (p. 10) mentions model cards documenting machine-learning models. The paper does not address preserving these documents as paradata or recording their relationship to archival collections processed using AI. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

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

Early archival theorists supposed that documents suitable to illustrate organizational practices could reasonably be expected to be produced in the ordinary course of business.

“In contrast to this application-oriented view of documentation, early archival theorists supposed that documents suitable to illustrate organizational practices could reasonably be expected to be produced in the ordinary course of business.”

• cites [34] 34. Hilary Jenkinson. 1937. A Manual of Archive Administration . P. Lund, Humphries & Co., Ltd., London, England. http://archive.org/ details/manualofarchivea00iljenk . sourceNot checked — no full text of the source
• cites [56] 56. Richard Stapleton. 1983. Jenkinson and Schellenberg: A comparison. Archivaria 1983 (1983), 75-85.Supported

Classified as probably not a substantial claim.

Jo and Gebru write that archives provide an example of a thoughtful approach to consent, power, inclusivity, transparency, and ethics and privacy in curating training datasets for machine learning.

“Jo and Gebru [ 35 ] write that archives provide an example of a thoughtful approach to fraught issues of”

• cites [35] 35. Eun Seo Jo and Timnit Gebru. 2020. Lessons from archives: Strategies for collecting sociocultural data in machine learning. In Proceed- ings of the 2020 Conference on Fairness, Accountability, and Transparency . ACM, New York, NY, 306-316. DOI: https://doi.org/10.1145/ 3351095.3372829 sourceSupported

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

In this context, paradata is a relational category falling within the broader category of metadata defined as data about data.

“In this specific context, paradata is a relational category that would fall into the broader category of metadata defined as data about data.”

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

XAI addresses the problem of AI tools' relative inscrutability to human oversight.

“XAI addresses the problem of AI tools' relative inscrutability to human oversight.”

• AssumptionRests on an assumption: Machine learning algorithms that learn from large datasets typically cannot articulate the basis of their decisions.

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

Under the SAGE definition, both formalized quantitative data collected by design and qualitative observations generated incidentally during research would count as paradata.

“The two extremes of formalized, quantitative information collected by design and qualitative observations generated incidentally throughout the research process would be considered paradata under this definition.”

• cites [46] 46. Henrietta O'Connor and John Goodwin. 2020. Paradata. SAGE Research Methods Foundations . SAGE Publications Ltd., London, England. DOI: https://doi.org/10.4135/9781526421036948039 sourceNot checked — no full text of the source

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

Although AI and XAI literature is closely related, it does not explicitly address the documentation problem in the way paradata does.

“Although literature on AI and XAI is thus closely related to the topic, it does not address explicitly the documentation problem as paradata does.”

Classified as probably not a substantial claim.

In heritage visualization, paradata intentionally describes the actions and agency exercised in a reconstructive activity, rather than documenting observations made during data collection.

“Here, rather than documenting observations made during the process of data collection, paradata provides an intentional description of the actions taken and agency exercised in the process of a reconstructive activity.”

• cites [20] 20. Hugh Denard, Richard Beacham, Franco Niccolucci, Sorin Hermon, and Anna Bentkowska-Kafel. 2009. The London charter for the computer-based visualization of cultural heritage. London Charter . Retrieved November 8, 2021 from http://w w w.londoncharter.org/ downloads.html .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), S4 (based).

Across the domains where paradata has entered technical and professional vocabularies, its role in explaining how information is created or processed recurs.

“The role of paradata in offering explanation of the process of creating or processing information recurs across the domains in which paradata has emerged in context-specific technical and professional lexicons.”

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

The NISS 2010 report defined metadata as formalized data about statistical data needed for searching, displaying, and analyzing those data.

“Metadata: Formalized data about statistical data needed to search for, display and analyze those data.”

• cites [36] 36. Alan Karr. 2010. Metadata and Paradata: Information Collection and Potential Initiatives . National Institute of Statistical Sciences, Washington, DC. https://w w w.niss.org/sites/default/files/research _ attachments/Metadata%20vs%20Paradata-FT.pdf.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.

XAI literature recognizes that explanations satisfactory to computer scientists may not satisfy the general public, and vice versa.

“At present, XAI literature recognizes that explanations satisfactory for computer scientists will not necessarily be satisfactory for the general public, or vice versa.”

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

AI tools may or may not be explainable and may or may not provide adequate paradata for an application; paradata is a property of records emerging from related documentation.

“AI tools may be explainable or not and may provide adequate paradata for an application or not; paradata is meanwhile a property of records emerging from related documentation.”

Classified as probably not a substantial claim.

Under Baker's definition, metadata and paradata do not overlap conceptually.

“By this definition, metadata and paradata do not overlap conceptually.”

• cites [4] 4. Drew Baker. 2012. Defining paradata in heritage visualization. In Paradata and Transparency in Virtual Heritage , Anna Bentkowska-Kafel and Hugh Denard (Eds.). Routledge, London, England, 163-175. https://doi.org/10.4324/9781315599366 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

From that perspective, paradata may change over time because it describes dynamic processes that influenced primary data throughout its life span.

“From this perspective, paradata, in contrast, may change over time, as it describes dynamic processes that have influenced the primary data throughout its life span.”

Classified as probably not a substantial claim.

Contextual metadata tends to focus on how an object has been interpreted, while paradata focuses on the processes of interpretation through which understanding of objects is communicated or sought.

“Paradata differ in emphasis from 'contextual metadata': whereas the latter tends to focus upon how an object has been interpreted, the central focus of paradata tends to be the processes of interpretation through which understanding of objects [is] communicated [or] sought [ 9 ].”

• cites [9] 9. Anna Bentkowska-Kafel and Hugh Denard (Eds.). 2012. Paradata and Transparency in Virtual Heritage . Routledge, London, England. DOI: https://doi.org/10.4324/9781315599366 sourceNot checked — no full text of the source

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

User data from interactions with apps or websites is fundamentally similar to paradata generated and preserved in research processes and recorded by computerized survey instruments, even if firms do not use the term paradata.

“While the term paradata may not be in vogue for the firms likely to collect, record, and analyze the micro-data generated in users' interactions with their apps or websites, this user data is fundamentally similar to paradata generated and preserved in a research process and recorded by computerized survey instruments.”

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

Explainability is an evaluable property of AI systems that can be assessed for suitability to a given application.

“Explainability is in this context an evaluable property of AI systems, one that can be assessed in relation to its suitability to a given application.”

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

Bentkowska-Kafel and Denard define paradata as information about human processes of understanding and interpreting data objects, constantly created whether or not it is recorded.

“The London Charter defines 'paradata' as information about human processes of understanding and interpretation of data objects.Paradata is thus constantly being created, irrespective of whether they are systematically recorded or disseminated.”

• cites [9] 9. Anna Bentkowska-Kafel and Hugh Denard (Eds.). 2012. Paradata and Transparency in Virtual Heritage . Routledge, London, England. DOI: https://doi.org/10.4324/9781315599366 sourceNot checked — no full text of the source

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

The SAGE Research Methods Foundations definition treats paradata as data outside intentionally or purposefully collected data, captured beyond methodological specification or alongside primary research concerns.

“'[P]ara' data are those data which fall outside the intentionally or purposefully collected data.They are data captured beyond the confines of any methodological specification or data that are collected alongside and in addition to the primary research concerns [ 46 ].”

• cites [46] 46. Henrietta O'Connor and John Goodwin. 2020. Paradata. SAGE Research Methods Foundations . SAGE Publications Ltd., London, England. DOI: https://doi.org/10.4135/9781526421036948039 sourceNot checked — no full text of the source

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

Existing literature in heritage visualization and statistical sciences tends to treat metadata as limited to static descriptive fields such as creator, date, and place.

“Existing literature in heritage visualization and statistical sciences tends to view metadata as limited to static descriptive fields: who created an object, when, where, and so on.”

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

Existing metadata schemas in institutions may contain some aspects of AI paradata, but not necessarily all.

“Specific schema for metadata existing in current institutions may contain some aspects of AI paradata, but not necessarily all.”

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

Solving the technical challenge of understanding what an algorithm does and why is only part of the puzzle; the availability of XAI tools will not instantly make AI tools appropriate for archives and their parent organizations' needs.

“Although solving the technical challenge of understanding what an algorithm is doing and why is certainly part of the puzzle when considering the use of AI in archival contexts, the availability of XAI tools will not instantly render AI tools appropriate for archives and their parent organizations' needs.”

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

The London Charter describes paradata as documentation of evaluative, analytical, deductive, interpretative, and creative decisions in computer-based visualisation, disseminated so relationships among research sources, implicit knowledge, explicit reasoning, and visualisation outcomes can be understood.

“Documentation of the evaluative, analytical, deductive, interpretative and creative decisions made in the course of computer-based visualisation should be disseminated in such a way that the relationship between research sources, implicit knowledge, explicit reasoning, and visualisation-based outcomes can be understood [ 20 ].”

• cites [20] 20. Hugh Denard, Richard Beacham, Franco Niccolucci, Sorin Hermon, and Anna Bentkowska-Kafel. 2009. The London charter for the computer-based visualization of cultural heritage. London Charter . Retrieved November 8, 2021 from http://w w w.londoncharter.org/ downloads.html .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), S4 (based).

Paradata is well established in the statistical social sciences as a term for data documenting the process of collecting primary data.

“As a term referring to data that documents the process of collection of primary data, paradata is well established within the statistical social sciences.”

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

By illuminating the context of creation and processing, paradata may allow readers to understand information objects in greater depth.

“By illuminating the context of their creation and processing, paradata may allow readers to understand the information objects in greater depth.”

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

To records professionals, metadata consists of both static and dynamic information points related to other information objects and must be maintained and updated as needed to manage a collection responsibly.

“To records professionals, metadata consists of both static and dynamic information points related to other information objects and must be maintained and updated as needed to responsibly manage a collection.”

• cites [66] 66. ITRUSTAI. 2018. Terminology database: "Metadata." InterPARES Trust AI . Retrieved November 12, 2022 from https://interparestrustai. org/terminology/term/metadata .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), S3 (load-bearing), S4 (based).

The specific forms of AI paradata documentation will vary by institutional, social, technical, and legal context.

“The specific forms of this documentation will vary based on institutional, social, technical, and legal contexts-a subject to which Section 3 of this article will return.”

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

Couper saw that data on survey respondents' keystrokes, cursor movement, and page scrolling could both deepen primary survey data and help evaluate survey instruments.

“Couper realized the potential value of data recording keystroke entries, cursor movement, and page scrolling behavior of survey respondents, both as a means of adding depth to the primary survey data and as a means of evaluating survey instruments [ 18 ].”

• cites [18] 18. Mick P. Couper. 1998. Measuring survey quality in a CASIC environment. In Proceedings of the Section on Survey Research Methods of the American Statistical Association . American Statistical Association, Dallas, TX, 41-49. http://w w w.asasrms.org/Proceedings/papers/ 1998 _ 006.pdf.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). May combine two claims.

AI tools fundamentally change archival processing by delegating agency to non-human actors, which necessitates new approaches and lenses for documentation.

“AI tools fundamentally change the nature of archival processing by delegating agency to non-human actors, a phenomenon that, argued here, necessitates new approaches and lenses for documentation.”

• AssumptionRests on an assumption: Archivists and records professionals are accustomed to understanding and documenting human agents rather than opaque AI processes.

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

From an information science standpoint, the NISS metadata definition is narrow because it leaves out administrative and use metadata that most repositories collect.

“From an information science perspective, NISS's definition of metadata is rather limited.It describes perhaps the metadata a typical researcher might encounter, but it excludes the administrative and use metadata that most repositories would collect.”

• cites [36] 36. Alan Karr. 2010. Metadata and Paradata: Information Collection and Potential Initiatives . National Institute of Statistical Sciences, Washington, DC. https://w w w.niss.org/sites/default/files/research _ attachments/Metadata%20vs%20Paradata-FT.pdf.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.

Although paradata may require more documentation effort than researchers are used to, it may reduce duplication of effort and ease comparison of datasets and information objects.

“Although paradata may demand that greater effort be dedicated to the documentation of research processes than researchers are currently accustomed to, it may also reduce duplication of effort and ease the process of comparing datasets and information objects.”

May combine two claims.

Mick Couper first proposed the term paradata, which came from his discussions about possible uses of keystroke data from a computer survey tool.

“Sociologist Mick Couper first suggested the term, which emerged from his discussions of the potential uses of keystroke data gathered by a computer survey tool [ 18 , 39 ].”

• cites [18] 18. Mick P. Couper. 1998. Measuring survey quality in a CASIC environment. In Proceedings of the Section on Survey Research Methods of the American Statistical Association . American Statistical Association, Dallas, TX, 41-49. http://w w w.asasrms.org/Proceedings/papers/ 1998 _ 006.pdf.Not matched

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

• cites [39] 39. Lars Lyberg, Frauke Kreuter, and Mick Couper. 2010. The use of paradata to monitor and manage survey data collection. In Proceedings of Joint Statistical Meetings 2010 . 282-296. http://sampieuchair.ec.unipi.it/wp-content/uploads/2018/10/Couper-et-al.pdf.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).

In this project, paradata is proposed as an analytical category that lets archivists interrogate AI or XAI tools and record actions and decisions they make or delegate to AI concerning records in their care.

“Paradata is proposed in this project as an analytical category that allows archivists to interrogate both the AI or XAI tools that they may apply in archives and to record the actions and decisions that they make themselves or delegate to AI that concern the records under their care.”

• AssumptionRests on an assumption: Archivists and records professionals are accustomed to understanding and documenting human agents rather than opaque AI processes.

Classified as probably not a substantial claim.

The NISS 2010 report defined paradata as formalized data on methodologies, processes, and quality tied to producing and assembling statistical data.

“Paradata: Formalized data on methodologies, processes and quality associated with the production and assembly of statistical data [ 36 ].”

• cites [36] 36. Alan Karr. 2010. Metadata and Paradata: Information Collection and Potential Initiatives . National Institute of Statistical Sciences, Washington, DC. https://w w w.niss.org/sites/default/files/research _ attachments/Metadata%20vs%20Paradata-FT.pdf.Not matched

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

Classified as probably not a substantial claim.

Baker distinguishes metadata, which describes properties of data objects for management and dissemination, from paradata, which describes processes of interpreting and creating data objects for understanding and evaluation.

“[W]hereas metadata, including contextual metadata, describe the properties of data objects in order to facilitate the management and dissemination of collections, paradata, as the term is used here, describe the processes of interpreting and creating data objects in order to enable understanding and evaluation [ 4 ].”

• cites [4] 4. Drew Baker. 2012. Defining paradata in heritage visualization. In Paradata and Transparency in Virtual Heritage , Anna Bentkowska-Kafel and Hugh Denard (Eds.). Routledge, London, England, 163-175. https://doi.org/10.4324/9781315599366 sourceNot checked — no full text of the source

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

The term paradata arose in the last quarter-century to name information that documents research processes beyond the usual boundaries of primary data.

“The term paradata has emerged in the past 25 years to describe information documenting research processes that falls outside of the traditional limits of the primary data collected.”

• cites [18] 18. Mick P. Couper. 1998. Measuring survey quality in a CASIC environment. In Proceedings of the Section on Survey Research Methods of the American Statistical Association . American Statistical Association, Dallas, TX, 41-49. http://w w w.asasrms.org/Proceedings/papers/ 1998 _ 006.pdf.Not matched

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

• cites [39] 39. Lars Lyberg, Frauke Kreuter, and Mick Couper. 2010. The use of paradata to monitor and manage survey data collection. In Proceedings of Joint Statistical Meetings 2010 . 282-296. http://sampieuchair.ec.unipi.it/wp-content/uploads/2018/10/Couper-et-al.pdf.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).

For records managers documenting AI actions, limiting potential liability for harms caused by AI systems will be a priority, especially in high-risk cases involving bias or tort.

“For records managers tasked with documenting the actions of an AI tool, limiting potential”

• cites [44] 44. Norman Mooradian. 2019. AI, records, and accountability. ARMA Magazine 2019 (2019), 9-13.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).

Polls suggest that 60% of Americans anticipate AI tools will reproduce the biases of their creators.

“polls suggest that 60% of Americans anticipate AI tools will reproduce the biases of their creators.”

• cites [44] 44. Norman Mooradian. 2019. AI, records, and accountability. ARMA Magazine 2019 (2019), 9-13.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 black box problem appeared in medicine as early as 1975, when computerized diagnosis tools were effective but could not explain their results.

“The black box problem emerged in the medical field as early as 1975, as computerized diagnosis tools proved effective but incapable of providing explanations for their results [ 54 ].”

• cites [54] 54. Edward H. Shortliffe, Randall Davis, Stanton G. Axline, Bruce G. Buchanan, C. Cordell Green, and Stanley N. Cohen. 1975. Computer- based consultations in clinical therapeutics: Explanation and rule acquisition capabilities of the MYCIN system. Computers and Biomed- ical Research 8, 4 (1975), 303-320. DOI: https://doi.org/10.1016/0010-4809(75)90009-9 sourceNot checked — no full text of the source

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

Mooradian recommends developing a strategy to document big data-level informational inputs, in full or part, before AI takes on a greater role within organizations.

“For each site, Mooradian [ 44 ] recommends developing a strategy to document big data-level informational inputs-whether that be documentation in full or part-to develop a strategy for AI documentation before AI takes on a greater role within organizations.”

• cites [44] 44. Norman Mooradian. 2019. AI, records, and accountability. ARMA Magazine 2019 (2019), 9-13.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).

Lynch identified the chief problem in documenting algorithm behavior as a shift from artifacts requiring mediation and curation to experiences.

“Focusing on social media platforms, Clifford Lynch has identified the chief problem in documenting the behavior of algorithms as being a”

• cites [40] 40. Clifford Lynch. 2017. Stewardship in the "Age of Algorithms." First Monday 22, 12 (2017). DOI: https://doi.org/10.5210/fm.v22i12.8097 sourceNot checked — no full text of the source

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

In the archival realm, AI tools should be as capable of explaining their own actions as archivists are.

“Translating this into the archival realm, it follows that AI tools should be as capable of explaining their own actions as archivists are themselves”

Classified as probably not a substantial claim.

The AI record Mooradian anticipates would document actions, transactions, and events carried out fully or partly by AI algorithms.

“The AI record he anticipates documenting”

• cites [44] 44. Norman Mooradian. 2019. AI, records, and accountability. ARMA Magazine 2019 (2019), 9-13.Not matched

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

Classified as probably not a substantial claim.

Lynch writes that basic documentation of algorithmic systems and their actions is not currently produced in society as presently organized.

“Lynch writes that basic documentation of algorithmic systems and their actions is not currently produced in society as presently organized.”

• cites [40] 40. Clifford Lynch. 2017. Stewardship in the "Age of Algorithms." First Monday 22, 12 (2017). DOI: https://doi.org/10.5210/fm.v22i12.8097 sourceNot checked — no full text of the source

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

All this information would comprise paradata because it records AI processes to which given records are individually or collectively subjected.

“All of this information would comprise paradata, given that it records AI processes to which given records are individually or collectively subjected.”

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

An AI record would need to offer a preliminary assessment of risks and benefits of a proposed AI solution, technical documentation of system design and testing, and documentation of specific decision paths.

“At a basic level, an AI record would have to offer first a preliminary assessment from the firm of the risks and benefits of a proposed AI solution; second, technical documentation of the AI's system design and testing; and third, documentation of the specific decision paths taken leading toward a decision.”

• cites [44] 44. Norman Mooradian. 2019. AI, records, and accountability. ARMA Magazine 2019 (2019), 9-13.Not matched

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

Classified as probably not a substantial claim.

The AI record, or paradata, is a preservation target in its own right because it documents the archivist's AI processing in relation to records at group or item level.

“The AI record, or here, paradata, comprises a preservation target in its own right given its documentation of the archivist's AI processing in relation to records at the group or item level [ 23 ].”

• cites [23] 23. Luciana Duranti, Adam Jansen, Giovanni Michetti, Mumma Courtney, Daryll Prescott, Corinne Rogers, and Thibodeau Kenneth. 2016. Preservation as a service for trust. In Security in the Private Cloud, John R. Vacca (Ed.). CRC Press, Boca Raton, FL, 47-72. DOI: https:// doi.org/10.1201/9781315372211-5 sourceNot checked — no full text of the source

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

Wadden defines the black box problem in healthcare as occurring when the reasons for an AI decision are not understandable to the patient or care providers because the system itself is not understandable to them.

“Wadden defines”

• cites [62] 62. Jordan Joseph Wadden. 2021. Defining the undefinable: The black box problem in healthcare artificial intelligence. Journal of Medical Ethics 48, 10 (2021), 764-768. DOI: https://doi.org/10.1136/medethics-2021-107529 sourceNot checked — no full text of the source

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

In creating AI records, situating decision-making loci within an organization and identifying AI roles at those sites will become key tasks for archivists of algorithmic agency.

“In creating these AI records, situating the loci of decision-making capabilities within an organization and identifying the roles of AI within those sites will become key tasks for the archivist of algorithmic agency.”

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

Guidelines developed based on stakeholder engagement were widely applicable across different contexts.

“guidelines developed based on stakeholder engagement were widely applicable across different contexts [ 7 ].”

• cites [7] 7. Jean-Christophe Bélisle-Pipon, Erica Monteferrante, Marie-Christine Roy, and Vincent Couture. 2022. Artificial Intelligence Ethics has a Black Box Problem . AI & Society. Early access, January 4, 2022. DOI: https://doi.org/10.1007/s00146-021-01380-0 sourceNot checked — no full text of the source

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

Explanations must satisfy not only the healthcare provider but also the patient.

“Explanations must satisfy not just the healthcare provider but the patient as well.”

Classified as probably not a substantial claim.

Mooradian argues that archivists need to develop an appropriate AI record to illustrate the decision processes behind an AI tool's implementation.

“As such, Mooradian [ 44 ] argues that archivists need to develop an appropriate”

• cites [44] 44. Norman Mooradian. 2019. AI, records, and accountability. ARMA Magazine 2019 (2019), 9-13.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.

The authors of that review do not discuss papers addressing documentation of AI tools themselves, suggesting key work remains on this topic.

“However, the authors do not discuss papers addressing the documentation of AI tools themselves, suggesting that key work remains to be done on this topic.”

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

The scope of the AI record will grow along with the expansion of AI tools' domains of action.

“This scope will grow along with the expansion of AI tools' domains of action.”

May combine two claims.

Bélisle-Pipon et al. reviewed 47 major AI ethics guidelines from private and public stakeholders; only 38% reported any stakeholder engagement, and only 9% reported engagement with citizens at large.

“Bélisle-Pipon et al. [ 7 ] reviewed 47 major AI ethics guidelines from private and public stakeholders.Their findings show that only 38% reported any kind of stakeholder engagement in their processes; only 9% reported any engagement with citizens at large.”

• cites [7] 7. Jean-Christophe Bélisle-Pipon, Erica Monteferrante, Marie-Christine Roy, and Vincent Couture. 2022. Artificial Intelligence Ethics has a Black Box Problem . AI & Society. Early access, January 4, 2022. DOI: https://doi.org/10.1007/s00146-021-01380-0 sourceNot checked — no full text of the source

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

Approaches to ethical AI show a division parallel to that in archives: ethics as a professional and social obligation on one side, and accountability embodied in workflows or reporting relationships and enforced by legislation on the other.

“Approaches to ethical AI display a similar social vs. institutional division: on the one hand, ethics as a professional and social obligation, and on the other, accountability as embodied in workflows or reporting relationships and enforced by legislation.”

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

Existing metadata fields and institutional documentation may serve to record archival paradata, although new means of capturing this information may also need to be devised.

“Existing metadata fields and institutional documentation may serve the purpose of recording archival paradata, although new means of capturing this information may also need to be devised.”

May combine two claims.

The regulatory shift toward explanation rather than institutionally limited documentation encourages active oversight by records subjects and aligns with recent calls for explainability in AI from the European Commission High Level.

“the regulatory shift toward explanation rather than institutionally limited documentation encourages active oversight by records subjects, and lines up with recent calls for explainability in AI from the European Commission High Level [ 11 , 28 ].”

• cites [11] 11. Jenny Bunn. 2020. Working in contexts for which transparency is important: A recordkeeping view of Explainable Artificial Intelligence (XAI). Records Management Journal 30, 2 (2020), 143-153. DOI: https://doi.org/10.1108/RMJ-08-2019-0038 sourceNot checked — no full text of the source
• cites [28] 28. High-Level Expert Group on Artificial Intelligence. 2018. Ethics Guidelines for Trustworthy AI . European Commission, Brussels, Belgium. https://ec.europa.eu/newsroom/dae/document.cfm?doc _ id=60419 .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), S4 (based). May combine two claims.

The right to an explanation is bolstered by robust new European data protection authorities and is likely to encourage data protection by design and wide-scale algorithmic auditing as recommended by the European Commission to prevent bias or unjust dealing in algorithmic data processing involving human subjects.

“The”

• cites [14] 14. Bryan Casey, Ashkon Farhangi, and Roland Vogl. 2019. Rethinking explainable machines: The GDPR's "right to explanation" debate and the rise of algorithmic audits in enterprise. Berkeley Technology Law Journal 34 (2019), 145-189. sourceNot checked — no full text of the source
• cites [65] 65. European Commission. 2017. Guidelines on Data Protection Impact Assessment (DPIA) . European Commission, Brussels, Belgium. https://ec.europa.eu/newsroom/article29/items/611236 .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).

Accountability does not create itself and must be situated within specific institutional, technical, and social contexts.

“Beyond the archivist's ethical imperative, accountability does not create itself and must instead be situated within specific institutional, technical, and social contexts.”

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

This is arguably a complication of records continuum perspectives that require beginning at the instantiation of a record, given the obscurity AI agency introduces.

“This is arguably a complication of records continuum perspectives that pose a need to begin at the instantiation of a record, given the obscurity that AI agency introduces into the process.”

May combine two claims.

The majority of paradata concerning AI records processing would fall into the category metadata after record capture, and paradata could be argued to be only a subset of that category.

“Into the latter category would fall the majority of paradata concerning AI records processing; here could be made the argument that paradata would only be a subset of the category”

May combine two claims.

ISO 23081 includes provisions for documentation of records' processing after creation, but it is designed for contexts where human agency, rather than AI tools, carries out the processing.

“although the ISO 23081 Metadata for Records standard includes provisions for documentation of records' processing after creation, it is designed to address contexts in which human agency, rather than AI tools, carry out the processing.”

• cites [64] 64. International Organization for Standardization. 2017. Information and Documentation-Records Management Processes-Metadata for Records-Part 1: Principles . International Organization for Standardization, Geneva, Switzerland.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).

Broad, socially oriented views of the value of archives for ensuring accountability have proliferated in archival literature since the 1970s.

“Broad, socially oriented views of the value of archives in ensuring accountability have proliferated in archival literature since the 1970s [ 6 , 47 ].”

• cites [6] 6. David Bearman. 1993. The implications of "Armstrong v. Executive of the President" for the archival management of electronic records. American Archivist 56, 4 (1993), 674-689. sourceNot checked — no full text of the source
• cites [47] 47. Jane Parkinson. 1993. Accountability in Archival Science . University of British Columbia. DOI: https://doi.org/10.14288/1.0086151 sourceNot checked — no full text of the source

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

ARMA's Generally Accepted Recordkeeping Principles place accountability and transparency as the first and second priorities for records managers, though their definition of accountability is bureaucratically situated rather than expansive.

“The list of Generally Accepted Recordkeeping Principles by the Association of Records Managers and Administrators (ARMA), for instance, offers accountability and transparency as the first and second priorities for records managers, although their articulated definition of accountability is bureaucratically situated rather than expansive [ 69 ].”

• cites [69] 69. ARMA International. 2022. The Principles: Generally accepted recordkeeping principles. ARMA International . Retrieved October 11, 2022 from https://w w w.arma.org/page/principles .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). May combine two claims.

AI tools demand that archivists take future archival readers into account at the instant of their application and devise documentation practices accordingly.

“AI tools demand archivists take future archival readers into account at the instant of their application and devise documentation practices accordingly.”

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

In complicated AI contexts, elaborate ML models may intersect with complex organizational structures, merging traditional recordkeeping problems of identifying responsible agents and documenting their activity with the new technical context.

“In more complicated AI contexts, elaborate ML models may intersect with complex organizational structures, merging the traditional recordkeeping problems of identifying responsible agents and documenting their activity with the new technical context.”

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

Addressing paradata and AI documentation requires archivists to have both an ethical mandate toward transparency in AI applications and the development of professional standards, policy, and legislation where needed for responsible and accountable AI applications.

“For archivists, the problem of paradata and AI documentation will necessitate both an ethical mandate obliging them toward transparency in AI applications and additionally the development of professional standards, policy, and legislation where necessary to ensure responsible and accountable AI applications.”

• AssumptionRests on an assumption: AI tools are in fact being implemented in archival and recordkeeping processes, so that documenting their activity is a live problem.

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

The inadequacy of existing archival categories emerges because the fundamentally different character of AI necessitates novel approaches if archival accountability and transparency are to be maintained.

“The inadequacy of existing archival categories emerges as the fundamentally different character of AI necessitates novel approaches if archival accountability and transparency are to be maintained.”

• AssumptionRests on an assumption: Existing documentation practices are premised on the presumption of human agency.
• AssumptionRests on an assumption: AI is not necessarily fully understandable by humans.

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

General policy documents will presumably still be developed by human actors, while operational records pertaining to specific instances of bureaucratic activity may increasingly be produced by AI.

“Whereas general policy documents will presumably still be developed by human actors, operational records pertaining to specific instances of bureaucratic activity may increasingly be produced by AI.”

• AssumptionThe authors state this as an assumption: General policy documents will still be developed by human actors.
• AssumptionRests on an assumption: Operational records for specific bureaucratic activities may increasingly be produced by AI.

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

Fritz Marx saw archives as ensuring transparency, following from the liberal-democratic insistence that public business be conducted according to public preference and under public scrutiny.

“Similarly, German archivist Fritz Marx saw archives as ensuring transparency, following from the liberal-democratic “insistence that public business be conducted along the lines of public preference and under the eyes of the public” [ 41 ].”

• cites [41] 41. Fritz Marx. 1947. The role of records in administration. American Archivist 10, 3 (1947), 241-248. DOI: https://doi.org/10.17723/aarc.10. 3.f354v0x358486416 sourceNot checked — no full text of the source

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

The increasing social power and technical complexity of AI is followed by emerging legislative definitions of the responsibilities of agencies that use AI.

“It is thus evident that the increasing social power and technical complexity of AI is followed by emerging legislative definitions of the responsibilities of agencies that make use of AI.”

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

Archivists must consider present, potential, and future readers within the content of accountability when considering AI documentation practices.

“Rather than seeing records as an inherent and trustworthy by-product of actions or processes without respect to their future audiences, archivists must consider present, potential, and future readers within the content of accountability when considering AI documentation practices.”

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

Andresen points out that the distinction between general policy records and specific operational records will remain the relevant materials to consult when an explanation is necessary.

“In many cases, Andresen points out that the distinction between general policy records and specific operational records will remain the relevant materials to consult when an explanation is necessary.”

• cites [2] 2. Herbjørn Andresen. 2019. A discussion frame for explaining records that are based on algorithmic output. Records Management Journal 30, 2 (2019), 129-141. DOI: https://doi.org/10.1108/RMJ-04-2019-0019 sourceNot checked — no full text of the source

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

Archivists and recordkeepers of all kinds are reasonably likely to understand their role as enabling accountability within their own institutions and in society more broadly.

“At present, archivists and recordkeepers of all stripes are reasonably likely to understand their role as enabling accountability in their specific institutional contexts and in society more broadly.”

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

Hurley gives three principles for ensuring accountability in archives: preserving evidence that protects records subjects' rights; documenting similar functional processes in similar ways; and ensuring transparency through consistent, pre-determined retention schedules.

“Hurley provides three main principles to ensuring accountability in archives: first, ensuring the preservation of evidence that preserves records' subjects' rights; second, ensuring consistency in that similar functional processes are documented in similar ways; and third, ensuring transparency in that retention schedules are consistent and determined in advance [ 29 ].”

• cites [29] 29. Chris Hurley. 2005. Recordkeeping and accountability. In Archives: Recordkeeping in Society . Chandos Publishing, Wagga Wagga, New South Wales, 223-252. DOI: https://doi.org/10.1016/B978-1-876938-84-0.50009-3 sourceNot checked — no full text of the source

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

Archivists might approach the XAI problem as situated within an institutional documentation context, in contrast with the technology/user focus of the XAI literature.

“As such, archivists might approach the XAI problem as situated within an institutional documentation context, in contrast with the technology/user focus of the XAI literature.”

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

As machine learning developments may draw on archival approaches, archivists are well positioned to consider how to push ML tools to abide by ethics compatible with archival standards and emerging legislation.

“As ML developments may draw on archival approaches, archivists are well positioned to consider how they can push ML tools to abide by ethics compatible with archival standards and emerging legislation.”

May combine two claims.

Legislative contexts spur not only development of XAI techniques but also documentation and records management practices for AI tools within firms using them.

“Such legislative contexts spur not only the development of XAI techniques but also of documentation and records management practices for AI tools within firms making use of them.”

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

Andresen explains that tools predicting likely reasons for an output given an input, much like XAI reconstructive models, might be the most certain forms of explanation available.

“Andresen [ 2 ] explains that tools which predict the likely reasons for an output given a certain input-much like XAI reconstructive models-might be the most certain forms of explanation available to those seeking one.”

• cites [2] 2. Herbjørn Andresen. 2019. A discussion frame for explaining records that are based on algorithmic output. Records Management Journal 30, 2 (2019), 129-141. DOI: https://doi.org/10.1108/RMJ-04-2019-0019 sourceNot checked — no full text of the source

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

For archive and recordkeeping contexts, this will necessitate development of vocabulary and capacity in a multidisciplinary approach to AI recordkeeping, as Andresen argues.

“For archive and recordkeeping contexts, this will necessitate the development of vocabulary and capacity in a multidisciplinary approach to AI recordkeeping, as Herbjørn Andresen argues in providing his”

• cites [2] 2. Herbjørn Andresen. 2019. A discussion frame for explaining records that are based on algorithmic output. Records Management Journal 30, 2 (2019), 129-141. DOI: https://doi.org/10.1108/RMJ-04-2019-0019 sourceNot checked — no full text of the source

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

Livia Iacovino described archives as arsenals of democracy, making a broad case for the liberal democratic value of archival practice in tracing actions and responsibility to their authors within complex social organizations.

“Representing the expansive social view of accountability, Livia Iacovino has described “Archives as arsenals of democracy” in an encompassing case for the liberal democratic value of archival practices as ensuring that actions and responsibility are traceable to their authors within complex social organizations [ 32 ].”

• cites [32] 32. Livia Iacovino. 2012. Archives as arsenals of accountability. In Currents of Archival Thinking , Terry Eastwood and Heather MacNeil (Eds.). Libraries Unlimited , Santa Barbara, CA, 181-212.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).

Bunn proposes several questions recordkeepers might pose when challenged to document AI or mixed human/AI decision making.

“Assessing the extent of necessary AI documentation, Jenny Bunn proposes several questions that recordkeepers might pose when challenged to document the AI or mixed human/AI decision making:”

• cites [11] 11. Jenny Bunn. 2020. Working in contexts for which transparency is important: A recordkeeping view of Explainable Artificial Intelligence (XAI). Records Management Journal 30, 2 (2020), 143-153. DOI: https://doi.org/10.1108/RMJ-08-2019-0038 sourceNot checked — no full text of the source

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

Paradata has emerged as a term for process documentation in several research fields, although simplified distinctions between metadata as static and paradata as dynamic are too common.

“To summarize the article's content to this point, paradata has emerged as a term for process documentation in several research fields, although too common are simplified distinctions between metadata as a static descriptive field and paradata as a dynamic field.”

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

The challenge posed by AI requires an approach to accountability that encompasses both the institutional and the broader social conceptions.

“The emergent challenge of AI calls for an approach to accountability encompassing both.”

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

An archivist's consideration of the paradata needed for a given context must be grounded in a clear assessment of the context-specific accountability relations.

“A clear assessment of the context-specific accountability relations must be at the basis of an archivist's task in considering the paradata necessary for a given context.”

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

Sophisticated AI tools can function as agents in their own right and may perform business and records management processes with a degree of agency comparable to that of human actors.

“Sophisticated AI tools can act as an agent in their own right and may carry out business and records management processes with agency comparable to human actors.”

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

AI introduces new problems in archives; it is not necessarily explainable as human or bureaucratic activity usually may be, and its explanations depend on technical documentation unfamiliar to many archivists and those to whom they are answerable.

“AI introduces new problems in archives; it is not necessarily explainable as human or bureaucratic activity usually may be, and to the extent that it is, its explanations are dependent on a bevy of technical documentation unfamiliar to many archivists and to those to whom archivists are answerable.”

Classified as probably not a substantial claim.

These documents would become part of any explanation necessary to uncover a record's processing context.

“These documents would become part of any explanation necessary to uncover a record's processing context [ 2 ].”

• cites [2] 2. Herbjørn Andresen. 2019. A discussion frame for explaining records that are based on algorithmic output. Records Management Journal 30, 2 (2019), 129-141. DOI: https://doi.org/10.1108/RMJ-04-2019-0019 sourceNot checked — no full text of the source

May combine two claims.

The European Union's AI Act proposal sets out a multilayered risk assessment approach that prohibits AI applications in unacceptably risky contexts and specifies minimum requirements for recordkeeping, transparency, and oversight of high-risk applications.

“For instance, the European Union's AI Act has recently proposed a multilayered risk assessment approach to AI applications, which proscribes AI applications in unacceptably risky contexts, and which outlines minimum requirements for recordkeeping, transparency, and oversight of implementations in high-risk applications.”

• cites [67] 67. European Commission. 2021. Laying down harmonized rules on artificial intelligence (Artificial Intelligence Act) and amending certain Union legislative acts. European Commission, Brussels, Belgium. https://digital-strategy.ec.europa.eu/en/library/proposal-regulation- laying-down-harmonised-rules-artificial-intelligence .Not matched

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

Classified as probably not a substantial claim.

The problem for archivists is the emergence of AI agency.

“The problem for archivists is the emergence of AI agency.”

• AssumptionRests on an assumption: AI can be a potential agent in its own right.

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

The White House Blueprint demands that users receive notice when AI processing impacts their treatment by an organization and that explanations be available for the AI's actions.

“The White House's Blueprint, for instance, demands that users receive notice of when AI processing is impacting their treatment by an organization, and that explanations be available for the AI's actions.”

• cites [68] 68. Whitehouse.gov. 2022. Blueprint for an AI bill of rights. The White House . Retrieved October 10, 2022 from https://w w w.whitehouse. gov/ostp/ai-bill-of-rights/ .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).

The introduction of useful but inscrutable black box AI tools demands deliberate creation of documentation strategies that may frequently entail additional AI tools to reconstruct the logic of a primary AI instrument.

“The introduction of useful but entirely inscrutable black box AI tools demands the deliberate creation of documentation strategies that may frequently, especially in the case of XAI, entail additional AI tools used to reconstruct the logic of a primary AI instrument.”

May combine two claims.

The archival tradition accommodates both a formal, institutionally situated accountability and a broader conception of documentation's role in democratic society, rather than treating accountability as a narrowly bureaucratic goal.

“Rather than understanding accountability as a narrowly defined bureaucratic goal, the archival tradition holds space for both a formal, institutionally situated accountability and a broader conception of documentation's role in democratic society.”

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

XAI as a field is dedicated to providing explanations of AI activity within a given context.

“XAI as a field is dedicated to the provision of explanations of AI activity within a given context”

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

Schellenberg's concept of evidential value is a conceptual precursor to later, more fully developed ideas about accountability.

“This is not to say that the threads of transparency or accountability are entirely novel; the “evidential value” of Schellenberg [ 51 ] is certainly a conceptual precursor to later further fleshed ideas about accountability.”

• cites [51] 51. Theodore R. Schellenberg. 1984. The appraisal of modern public records. In A Modern Archives Reader . National Archives Trust Board, Washington, DC, 57-70. http://w w w.archivists.org/prof-education/pre-readings/FAA/Schellenberg _ Article.pdf.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 Blueprint draws heavily on XAI literature by emphasizing that explanations be clear and tailored to a given audience's comprehension.

“The Blueprint draws heavily on XAI literature by emphasizing that explanations be clear and tailored to the comprehension of a given audience [ 68 ].”

• cites [68] 68. Whitehouse.gov. 2022. Blueprint for an AI bill of rights. The White House . Retrieved October 10, 2022 from https://w w w.whitehouse. gov/ostp/ai-bill-of-rights/ .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), S3 (load-bearing), S4 (based). May combine two claims.

Although these pieces of legislation are in process, it is reasonable to anticipate eventual passage of similar laws and imperative to develop professional practices that meet or exceed their standards.

“Although these pieces of legislation are currently in process, it is reasonable to anticipate the eventual passage of similar laws and imperative to develop professional practices that will meet or exceed their standards.”

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

The ISO standard proposes the collection of five types of metadata, to be gathered both at the point of capture and afterward.

“The standard proposes the collection of five types of metadata, to be gathered both at the point of capture and afterward:”

• cites [64] 64. International Organization for Standardization. 2017. Information and Documentation-Records Management Processes-Metadata for Records-Part 1: Principles . International Organization for Standardization, Geneva, Switzerland.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).

A National Archives report notes that successful explanations must consider the individuals and the environment in which a tool will be used.

“a report on the National Archives' experiences with AI notes that successful explanations must”

• cites [33] 33. Lise Jaillant, Katherine Aske, and Annalina Caputo. 2021. The National Archives (UK): Case Study . AEOLIAN Network: Artificial Intel- ligence for Cultural Institutions. https://w w w.aeolian-network.net/case-study-1-the-national-archives-uk/ .Not matched

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

Classified as probably not a substantial claim.

Bunn suggests reframing the record as a document that provides an explanation of the actions taken by an organization.

“Bunn ultimately suggests reframing the record as being a document that provides an explanation of the actions taken by an organization.”

• cites [11] 11. Jenny Bunn. 2020. Working in contexts for which transparency is important: A recordkeeping view of Explainable Artificial Intelligence (XAI). Records Management Journal 30, 2 (2020), 143-153. DOI: https://doi.org/10.1108/RMJ-08-2019-0038 sourceNot checked — no full text of the source

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

Existing proposed regulations and frameworks leave responsible institutions to devise their own implementation-specific recordkeeping practices, because prescribing detailed requirements would be futile in such a dynamic technological context.

“These proposed regulations and frameworks leave responsible institutions to devise their own implementation-specific recordkeeping practices, since to prescribe detailed requirements would be a fruitless task in such a dynamic technological context.”

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

Paradata is proposed as a category for the documentation of AI processes that may have shaped the form in which readers encounter an archival record or records group.

“Paradata is proposed as a category to describe the necessary documentation of AI processes that may have influenced the form in which readers encounter the archival record or records group.”

• AssumptionRests on an assumption: AI tools are in fact being implemented in archival and recordkeeping processes, so that documenting their activity is a live problem.

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

Making paradata an integral part of the record will require a deliberate approach to prevent records from becoming dissociated from their paradata.

“The implementation of paradata as an integral part of the record will demand a deliberate approach to prevent dissociation of records from their paradata.”

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

Under the EU AI Act proposal, unacceptable risks include applications that may directly violate individuals' rights and to which individuals cannot give informed consent, while high-risk contexts may still risk rights violations but can be regulated to an acceptable level of risk for the specific application.

“Unacceptable risks include applications that may directly violate individuals' rights and that individuals are not able to grant informed consent to, whereas high-risk contexts may still risk violating individuals' rights but may be regulated to an acceptable level of risk within the specific context of the application.”

• cites [67] 67. European Commission. 2021. Laying down harmonized rules on artificial intelligence (Artificial Intelligence Act) and amending certain Union legislative acts. European Commission, Brussels, Belgium. https://digital-strategy.ec.europa.eu/en/library/proposal-regulation- laying-down-harmonised-rules-artificial-intelligence .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).

Title IV of the EU AI Act proposal calls for transparency from certain AI systems to address the specific risks of manipulation they pose.

“Additionally, Title IV of the proposal also calls for transparency on the part of “certain AI systems to take account of the specific risks of manipulation they pose” by, among other things, offering imitations of their real-world equivalents [ 67 ].”

• cites [67] 67. European Commission. 2021. Laying down harmonized rules on artificial intelligence (Artificial Intelligence Act) and amending certain Union legislative acts. European Commission, Brussels, Belgium. https://digital-strategy.ec.europa.eu/en/library/proposal-regulation- laying-down-harmonised-rules-artificial-intelligence .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).

Because paradata overlaps with existing categories of contextual and administrative metadata and is conceptually similar to the broader category of data about data, it may lend itself to being a component within a record's metadata.

“Paradata's overlap with existing categories of contextual and administrative metadata, along with its conceptual similarity to the broader category of “data about data” may lend itself to this approach.”

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

Recent trends toward multifarious conceptions of provenance remain relevant to the question of how AI processing affects provenance.

“Further theoretical and practical research on this matter would clarify the matter, although recent trends toward multifarious conceptions of provenance remain relevant here.”

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

As more human agency is delegated to AI tools, archivists must find new ways to document activity carried out by AI beyond their direct control, and paradata is proposed as a new category of analysis for this context.

“As increasing levels of human agency are delegated to AI tools, archivists must find new solutions to document activity carried out by AI beyond their direct control-and it is in this context that paradata is proposed as a new category of analysis.”

• AssumptionRests on an assumption: AI tools are in fact being implemented in archival and recordkeeping processes, so that documenting their activity is a live problem.

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

Every field of research captures and communicates some information about the process by which its research data were collected.

“Every field of research captures and communicates some information about the process of research data collection.”

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

Archivists need to devise practices that describe the AI tools used, their effects, the mandate and purpose of their application, and the individuals who select and apply them.

“Archivists must work from this imperative to devise adequate practices to describe the AI tools used, their effects, the mandate and purpose of their application, and the individuals selecting and applying these tools.”

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

In some fields the term paradata refers to collecting and disseminating quantitative or qualitative data that reveal the context of data creation, in order to provide transparency into the research process, expose researchers' choices and methods, and enable informed use of the dataset by later researchers.

“In some fields, this is referred to as “paradata” and entails the collection and dissemination of quantitative or qualitative data revealing the context of data creation to provide transparency into the research process, reveal the choices made and methods used by researchers, and enable informed usage of the dataset by future researchers.”

Classified as probably not a substantial claim.

ISO 23081 classifies recordkeeping metadata into five types: metadata about the record itself, about business rules or policies and mandates, about agents, about business activities or processes, and about records management processes.

“(3) metadata about agents; (4) metadata about business activities or processes; (5) metadata about records management processes” [ 64 ].”

• cites [64] 64. International Organization for Standardization. 2017. Information and Documentation-Records Management Processes-Metadata for Records-Part 1: Principles . International Organization for Standardization, Geneva, Switzerland.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), S3 (load-bearing), S4 (based).

Practical implementations of paradata design may face further difficulties arising from the commercialized basis of many AI tools.

“Implementations of paradata design in practice may encounter further difficulties in relation to the commercialized basis of many AI tools.”

• AssumptionRests on an assumption: AI tools are in fact being implemented in archival and recordkeeping processes, so that documenting their activity is a live problem.

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

AI processing may subtly alter a record or record group's provenance; the authors hope thorough paradata collection might document this, but it still must be minimized.

“For instance, AI processing may subtly change a record or record group's provenance; this the current research group hopes might be documented by the collection of thorough paradata but still must be minimized.”

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

AI in archives may require revisiting some fundamental archival principles or devising ways of using paradata to address the challenges AI poses.

“On a theoretical level, AI in archives may create a need to revisit some fundamental archival principles, or devise ways of using paradata to address the challenges AI poses.”