CoChat
Sign Up
Cited by 102Explainable Artificial Intelligence (XAI)Ethics and Social Impacts of AIData Visualization and Analytics

Interactive Model Cards: A Human-Centered Approach to Model Documentation

Open access
Crisan, Anamaria, Drouhard, Margaret, Vig, Jesse, Rajani, Nazneen·2022·2022 ACM Conference on Fairness Accountability and Transparency
DOI 10.1145/3531146.3533108
Open this paper in CoChat — free accountOpen PDF

Abstract

Deep learning models for natural language processing (NLP) are increasingly adopted and deployed by analysts without formal training in NLP or machine learning (ML). However, the documentation intended to convey the model’s details and appropriate use is tailored primarily to individuals with ML or NLP expertise. To address this gap, we conduct a design inquiry into interactive model cards, which augment traditionally static model cards with affordances for exploring model documentation and interacting with the models themselves. Our investigation consists of an initial conceptual study with experts in ML, NLP, and AI Ethics, followed by a separate evaluative study with non-expert analysts who use ML models in their work. Using a semi-structured interview format coupled with a think-aloud protocol, we collected feedback from a total of 30 participants who engaged with different versions of standard and interactive model cards. Through a thematic analysis of the collected data, we identified several conceptual dimensions that summarize the strengths and limitations of standard and interactive model cards, including: stakeholders; design; guidance; understandability & interpretability; sensemaking & skepticism; and trust & safety. Our findings demonstrate the importance of carefully considered design and interactivity for orienting and supporting non-expert analysts using deep learning models, along with a need for consideration of broader sociotechnical contexts and organizational dynamics. We have also identified design elements, such as language, visual cues, and warnings, among others, that support interactivity and make non-interactive content accessible. We summarize our findings as design guidelines and discuss their implications for a human-centered approach towards AI/ML documentation.

Verified 2026-10-06 · 2 references not checked yet

AI snapshot

Written from the paper’s full text.

TL;DR

The paper designs and evaluates interactive model cards (IMCs) that augment static model cards with interactivity and guidance for non-expert NLP/ML analysts. Across two qualitative studies with 30 participants, IMCs were generally favored over standard model cards for understandability and sharing, though ethics and safety content remained hard to make actionable.

Key findings

  • In Study 2, nearly all 20 non-expert analysts (n=18) found the IMC easier to read and understand; n=15 said they would use it in daily work, while n=2 said it was context-dependent, and the IMC was overall favored over the standard model card (SMC).
  • With the SMC alone, almost half of Study 2 participants could not correctly interpret that the model performed sentiment analysis: four thought it was a general text classification model, two thought DistilBERT and its fine-tuned version were both sentiment analysis models, one did not answer, and one thought the SST-2 dataset was the model; even three who correctly described it were not confident.
  • IMC design elements helped engagement: layout (n=13), language (n=6), ability to add examples (n=15), and visual/interactive elements (n=16); n=16 also suggested improvements, and n=4 reported information overload.
  • Study 1 experts surfaced conceptual dimensions including stakeholders, design, guidance, trust and safety, and sensemaking and skepticism, leading to design guidelines DG1–DG4 for interactive model cards.

Methods

  • Two-study design inquiry: Study 1 was a conceptual study with 10 experts in ML/AI, NLP, ethics, and auditability using paper-prototype probes of an SMC, Robustness Gym report (RGR), and IMC; Study 2 was an evaluative study with 20 non-expert analysts using a functional IMC prototype compared with a reference SMC.
  • Both studies used semi-structured interviews with a think-aloud protocol; Study 1 recorded approximately 11 hours 15 minutes of video, and Study 2 recorded approximately 23 hours 30 minutes; analysis used iterative thematic analysis with grounded theory, with Study 1 themes seeding Study 2 analysis.
  • Study 2 recruitment used the User Interviews platform: 176 responses, 117 excluded (89 never used ML/NLP models; 28 had graduate or undergraduate training in ML, statistics, or computer science), leaving 59 candidates split into Type I (n=35, no or infrequent programming) and Type II (n=24, daily or weekly programming), from which 20 participants were recruited evenly; the reference model was DistilBERT fine-tuned on SST-2 for sentiment analysis.

Limitations

  • Stated: design choices and the choice of sentiment analysis task likely influenced findings; specific IMC design elements may be refined in future work, though the broader design guidelines and conceptual dimensions are expected to generalize; creating and learning to use IMCs is a heavier burden than text-based model cards.
  • Stated/future: future work should explore strategies for reducing and redistributing the work of model documentation.
  • Evident from the text: the qualitative samples are small (10 experts and 20 analysts) and not statistically representative; Study 2 found no large response differences between analyst groups; ethics and safety remained difficult, with only four participants saying ethical templates could stimulate discussion and two pushing back against more actionable ethics/safety content.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3531146.3533108

Cites

62 of 64 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 20 cited sources checked · 0 supported · 0 flagged

  • Findingown data
    § 4.3 Results · p. 12

    Almost half of participants were not able to provide the correct interpretation of what the model did using the SMC alone.

    “Overall, almost half of the participants were not able to provide the correct interpretation of what the model did using the SMC alone.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 4.3 Results · p. 10

    Non-expert analysts echoed concerns previously articulated by experts in Section 3.

    “with participants echoing the concerns that were previously articulated by the experts in Section 3.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 3.3 Results · p. 7

    Nearly all (n=7) participants indicated that examples could provide some useful guidance for supporting interpretability.

    “Supporting interpretability is complex but nearly all (n=7) participants indicated that examples could provide some useful guidance.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 4.3 Results · p. 11

    A few participants (n=4) expressed that text, visuals, and interaction options led to information overload.

    “A few participants (n=4) expressed that text, visuals, and interaction options led to information overload.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 4.3 Results · p. 10

    The reference model card and IMC each had features preferred by both types of analysts.

    “The reference model card (Section 3.1), here acting as the SMC, and the IMC each had features that were preferred by both types of analysts.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 24
  • Classified as probably not a substantial claim. 35

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 11

    The addition of guiding elements suggested in the previous study were well received and did help participants understand and interpret the model card information.

    “Notably, the addition of guiding elements, suggested in the previous study (Table 2), were well received and did help participants understand and interpret the model card information.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    The conceptual study emphasized the importance of information design within both standard and interactive model cards.

    “Our conceptual study emphasized the importance of information design within both standard and interactive model cards.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingown data
    § 3.3 Results · p. 7

    Study 1 participants indicated that greater attention to information, visual, and interaction design would improve trust and safety.

    “Participants indicated that for these issues and overall, greater attention to information, visual, and interaction design would improve trust and safety.”
    Quote in the paperSubstantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    Several (n=4) study 1 participants expressed that adding sensitive populations to disaggregated model performance could be beneficial, but cautioned that definitions of these groups are culturally and geographically contextual.

    “Several (n=4) participants expressed that it could be beneficial to add sensitive populations to the disaggregated model performance, but they suggested caution since the definition of these groups was culturally and geographically contextual.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 4.3 Results · p. 11

    Participants suggested that integrating the model's code and files, present in the SMC, would be valuable for the IMC.

    “Participants suggested that integrating the model's code and files, present in the SMC, would be valuable for the IMC.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    Warnings, prompts, and summaries are examples of guidance that participants suggested could help guide analysts in making sense of model cards.

    “Warnings, prompts, and summaries are all examples of guidance that participants suggested could help guide analysts in making sense of model cards.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 11

    When asked, nearly all participants (n=15) indicated that the IMC was something they would use in their daily work, while some (n=2) indicated it was context-dependent.

    “When asked, nearly all participants (n=15) indicated that the IMC was something they would use in their daily work, while some (n=2) indicated it was context-dependent.”
    Quote in the paperSubstantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    One study 1 participant initially thought the model performed well for a particular subpopulation, then revised their interpretation after seeing the small sample size.

    “one participant initially thought the model performed well for a particular subpopulation, then saw the small sample size and revised their interpretation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.3 Results · p. 8

    Study 1 participants indicated that interactivity would enable people to elicit more questions, form and test mental models about the data, and better see what the model lacks.

    “the interactivity would enable people to 'elicit more questions' [E04] that allowed them to form and test their mental models about the data and better 'see the things that model lacks' [E05].”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 11

    Participants articulated that several components of the IMC helped them engage with the model and data more effectively by clarifying and prioritizing information, particularly layout (n=13), language (n=6), ability to add examples (n=15), and visual and interactive elements (n=16).

    “Participants articulated that several components of the IMC helped them engage with the model and data more effectively by clarifying and prioritizing information, particularly: layout of information (n=13); language (n=6); ability to add examples (n=15); and visual and interactive elements (n=16).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    Experts saw the opportunity to use interaction as a way to help non-expert analysts develop their conceptual understanding of the NLP model.

    “experts saw the opportunity to use interaction as a way to help non-expert analysts develop their conceptual understanding of the NLP model.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 11

    Three participants voiced the importance of being able to contest the model's results and wanted this ability expanded beyond adding sentences to the model's training and testing data or uploading their own dataset.

    “Three participants voiced the importance of being able to contest the model's results, and wanted to see this ability expanded beyond adding sentences (Section 4.1) to the model's training and testing data or uploading their own dataset.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    Some study 1 participants viewed ethical statements that only say there is some bias somewhere as unhelpful.

    “Increasingly, such statements are seen as a 'cop-out because it just kind of says 'there is some bias somewhere' which is not entirely helpful.' [E06]”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 11

    Nearly all (n=16) participants made suggestions for improvements to the IMC design, including changes to confusing interface elements (n=5), layout (n=6), and types of guidance (n=1).

    “Nearly all (n=16) participants also made suggestions for improvements to the IMC design. These included changes to elements of the interface that were confusing (n=5), the layout (n=6), and types of guidance (n=1) provided.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    Study 1 participants emphasized the need to prominently display the total size of subpopulations to support valid interpretation.

    “they emphasized the need to prominently display the total size of these subpopulations (and others) to support valid interpretation.”
    Quote in the paperNot a substantial claim

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

    Findingnone given
    § 2 BACKGROUND AND RELATED WORK · p. 3

    Our approach to guidance, shareability, and traceability begins to address the needs for safe and effective deployment of AI/ML systems, and we surfaced additional relevant implications for design.

    “We found that our approach to guidance, shareability, and traceability begins to address these needs, and we surfaced additional relevant implications for design.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The DistilBERT SST-2 sentiment analysis task is representative of NLP model development regimens for other tasks.

  • May combine two claims.

    Findingown data
    § 3.3 Results · p. 7

    All participants expressed doubt that a non-expert analyst would understand the model card or know how to act on the information it contained.

    “All participants expressed doubt that a non-expert analyst would understand the model card or know how to act on the information it contained.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.3 Results · p. 7

    A sense of focus/overwhelm varied among participants, and one participant articulated that the IMC contained too much information and is very cluttered.

    “However, a sense of focus/overwhelm varied among participants, and one participant articulated that the IMC contained 'too much information and is very cluttered' [E04].”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    Two study 1 participants suggested stating explicitly whether bias assessments had been conducted, and one suggested raising a warning if such bias analyses are not run.

    “Two participants suggested stating explicitly whether bias assessments had been conducted, and one suggested raising a warning if such bias analyses are not run.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    Four participants stated that a summary of the model's purpose was needed and that it would improve the readability of the model card.

    “Four participants stated that a summary of the model's purpose (i.e., 'this is a sentiment analysis model') was needed and that it would improve the readability of the model card.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 3.3 Results · p. 7

    Having an effective default example to prompt end-users was also important.

    “However, having an effective default example to prompt end-users was also important [E04].”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 14

    Although the IMC included templates to examine ethically challenging examples, only four participants expressed that these features could stimulate discussion.

    “Although the IMC included templates to examine ethically challenging examples (Section 4.1), only four participants expressed that these features could stimulate discussion.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 14

    While the IMC provided template sentences to promote ethical thinking, participants largely favored their own examples.

    “While the IMC also provided template sentences to promote ethical thinking, participants largely favored their own examples.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 4.3 Results · p. 13

    Three participants expressed that IMCs, if they could be easy to produce, would lower their existing model documentation burden.

    “Moreover, three participants expressed that IMCs, if they could be easy to produce, would lower their existing model documentation burden.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5 DISCUSSION · p. 15

    We identified some cynicism toward ethics statements, as well as assertions that ethical questions require contextual, situated examination.

    “However, our research supports prior findings that these concepts are challenging to integrate into decision-making [48, 60], and we identified some cynicism toward ethics statements, as well as assertions that ethical questions require contextual, situated examination.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Model cards are intended to surface ethical and safety implications, a topic especially pressing for deep learning models.

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

    Findingown data
    § 4.3 Results · p. 14

    Others felt that topics of safety and ethics were too abstract.

    “Others felt that topics of safety and ethics were too abstract, with one participant stating that they 'wouldn't worry too much about that to be honest, I guess it depends on the type of analysis you're doing.' [A2-05].”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 4.3 Results · p. 13

    The reference model card contained information and actions to obtain the model code and launch it, which was absent in the IMC.

    “Here, the reference model card contained information and actions to obtain the model code and launch it - something that was absent in the IMC.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Results · p. 14

    Unexpectedly, two participants pushed back against the inclusion of more actionable ethical and safety content, articulating concerns that too much guidance could be detrimental and provide a false sense of security.

    “Unexpectedly, two participants pushed back against the inclusion of more actionable ethical and safety content, articulating concerns that too much guidance could be detrimental and provide a false sense of security.”
    Quote in the paperNot a substantial claim

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

    Findingown data
    § 4.3 Results · p. 14

    One participant expressed doubt that any model documentation could really give an adequate picture of ethical and safety risks.

    “One participant expressed doubt that any model documentation could really give an adequate picture of ethical and safety risks, emphasizing that ' 'this [model card] would not be my only data point or else, you know, that's a very quick way to lose your job' [A1-04].”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    The hierarchy and design of model card content is equally important as the categories and depth of information to ensure content is effectively understood by a wide audience.

    “However, to be sure that this content is effectively understood by a wide audience, the hierarchy and design of this content is equally important.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Prior work related to SMCs has focused primarily on the categories and depth of information that model documentation should include.

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

    Study design and conductnone given
    § 3.2 Study Procedures · p. 5

    Video transcripts were automatically generated, including speaker identification, and were verified for accuracy by the authors.

    “Video transcripts were automatically generated, including speaker identification, and were verified for accuracy by the authors.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationown data
    § 4.3 Results · p. 10

    This depth allowed the authors to expand on subthemes and broaden conceptions of how model cards might be used.

    “This depth allowed us to expand on the subthemes from Table 2 and broaden our conceptions of how model cards might be used.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data
    § 3.2 Study Procedures · p. 5

    All participants had exposure to documentation for ML/AI models in some form, but only six were familiar with model cards and two with RGRs.

    “All participants had exposure to documentation for ML/AI models in some form, but only six were familiar with model cards and two with RGRs.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation
    § 2 BACKGROUND AND RELATED WORK · p. 3

    Robustness Gym helps model developers report on model behavior through the generation of robustness reports (RGRs) that summarize model performance on various data slices.

    “Robustness Gym helps model developers report on model behavior through the generation of robustness reports (RGRs) that summarize model performance on various data slices [23].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23]

      Not checked in the free run.

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

    Interpretationreasoning
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    Clear guidance can support non-experts in understanding and knowing how to act on the information.

    “Clear guidance can support non-experts in understanding and knowing how to act on the information.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    Additional guidance will be necessary to support interpretability of the outputs.

    “Additional guidance will be necessary to support interpretability of the outputs.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: SMCs have been primarily tailored toward developers and machine learning experts.

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

    Interpretationnone given
    § 3.4 Summarizing Design Guidance for an Interactive Model Card · p. 8

    Productive skepticism would make ideas of intended and unintended uses more intuitive than simply being told what they are.

    “We argue that this productive skepticism would make ideas of intended and unintended uses more intuitive than simply being told what they are.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Not all readers of a model card will want to interact with it.

  • Classified as probably not a substantial claim.

    Premisecitation
    § 1 INTRODUCTION · p. 2

    Non-expert analysts benefit from interacting with machine learning models and their data.

    “Prior work has shown that non-expert analysts benefit from interacting with machine learning models and their data [1, 19, 59].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [1]

      Not checked in the free run.

    • cites [19] source

      Not checked in the free run.

    • cites [59]

      Not checked in the free run.

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

    Premisecitation
    § 2 BACKGROUND AND RELATED WORK · p. 3

    Safe and effective deployment of AI/ML systems in practice requires a robust understanding of different stakeholders' needs and goals for explainability; support for traceability and auditing; and strategies to address concerns within existing organizational structures and dynamics.

    “As many researchers have articulated, safe and effective deployment of AI/ML systems in practice requires a robust understanding of different stakeholders' needs and goals for explainability [8, 32]; support for traceability and auditing [71]; and strategies to address concerns within existing organizational structures and dynamics [8, 31, 32, 57, 71].”
    Quote in the paperNot a substantial claim0 of 5 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [32] source

      Not checked in the free run.

    • cites [71] source

      Not checked in the free run.

    • cites [31] source

      Not checked in the free run.

    • cites [57]

      Not checked in the free run.

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

    Interpretationreasoning
    § 5 DISCUSSION · p. 15

    Scaffolding these interactions to support sensemaking requires clear and actionable guidance.

    “Importantly, we found that scaffolding these interactions to support sensemaking requires clear and actionable guidance.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning
    § 5 DISCUSSION · p. 14

    Interactive model cards are a bridge to further analysis, not a substitute.

    “1) Interactive model cards are a bridge to further analysis, not a substitute.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given
    § 5 DISCUSSION · p. 15

    Creating and learning how to use IMCs is a heavier burden than for text-based model cards.

    “We also recognize that creating and learning how to use IMCs is a heavier burden than for text-based model cards.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The recruited participants are representative of the future data workforce.

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

    Interpretationreasoning
    § 5 DISCUSSION · p. 15

    For model cards to be adopted and impact organizational processes, they must support knowledge sharing and negotiation across stakeholders with diverse backgrounds and perspectives.

    “For model cards to be adopted and impact organizational processes, they must support knowledge sharing and negotiation across stakeholders with diverse backgrounds and perspectives.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning
    § 5 DISCUSSION · p. 14

    With thoughtfully designed interaction, model cards can be a first step for data workers to explore how trustworthy a model is and whether it might meet their needs, serving as a guide to deeper interrogation when desired.

    “We argue that with thoughtfully designed interaction, model cards can be a first step for data workers to explore how trustworthy a model is and whether it might meet their needs, serving as a guide to deeper interrogation when desired.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 5 DISCUSSION · p. 15

    Model documentation can spark ethical thinking, but will never substitute for it.

    “Our findings indicate, unsurprisingly, that model documentation can spark ethical thinking, but will never substitute for it.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Model cards are intended to surface ethical and safety implications, a topic especially pressing for deep learning models.

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

    Study design and conductnone given
    § 5 DISCUSSION · p. 15

    Design choices and the choice of sentiment analysis tasks likely influenced the study's findings.

    “Design choices and the choice of sentiment analysis tasks (Section 3.1) likely influenced our findings.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given
    § 5 DISCUSSION · p. 14

    For individual stakeholders, the most effective form for interactivity in model cards will be situated and contextual.

    “For individual stakeholders, the most effective form for interactivity in model cards will be situated and contextual.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 5 DISCUSSION · p. 15

    Considerations of ethics are personal and contextual, so they are unlikely to be addressed with a single approach.

    “Importantly, our analysis reinforces that considerations of ethics are personal and contextual, so they are unlikely to be addressed with a single approach.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Model cards are intended to surface ethical and safety implications, a topic especially pressing for deep learning models.

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

    Interpretationown data
    § 4.4 Summary · p. 14

    Their feedback validates many of the authors' design choices, but also offers avenues for further improvements.

    “Their feedback validates many of our design choices, but also offers avenues for further improvements.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given
    § 5 DISCUSSION · p. 15

    The broader design guidelines and conceptual dimensions surfaced by the study are robust and will generalize to other tasks and contexts.

    “However, we believe that the broader design guidelines (Section 3.4) and conceptual dimensions that we surfaced (Tables 2, 4, and 5) are robust and will generalize to other tasks and contexts.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The recruited participants are representative of the future data workforce.
50 of 50

  • The paper relies on this without support.

    Model cards and similar documentation rely on the developer to accurately and clearly report on the model and its performance.

    “However, these model cards and similar forms of documentation rely on the developer to accurately and clearly report on the model and its performance.”

  • The paper relies on this without support.

    Many important details, such as unintended uses or disaggregated model performance, are omitted from model cards.

    “Often, this process is labor-intensive and many important details, such as unintended uses or disaggregated model performance, are omitted.”

  • The paper relies on this without support.

    The DistilBERT SST-2 sentiment analysis task is representative of NLP model development regimens for other tasks.

    “We chose to focus on a sentiment analysis task because it is widely known and easily understood, while still being representative of NLP model development regimens for other tasks.”

  • The paper relies on this without support.

    The design probes (SMC, RGR, IMC) adequately represent the respective documentation types.

    “We developed design probes [70] of an SMC, RGR, and IMC. For the SMC, we used the all the text from the reference model card, with some exceptions, laid out according to the specifications in Mitchell et al. [44].”

  • The paper relies on this without support.

    The model card format is text-based and concise, making it both broadly accessible and applicable across model types.

    “The format they propose is text-based and concise, making it both broadly accessible and applicable across model types.”

  • The paper relies on this without support.

    Prior work related to SMCs has focused primarily on the categories and depth of information that model documentation should include.

    “Prior work related to SMCs has focused primarily on the categories and depth of information that model documentation should include.”

  • The paper relies on this without support.

    SMCs have been primarily tailored toward developers and machine learning experts.

    “SMCs have been primarily tailored toward developers and machine learning experts”

  • The paper relies on this without support.

    Not all readers of a model card will want to interact with it.

    “Not all readers of a model card will want to interact with it.”

  • The paper relies on this without support.

    Technology that enables low or no code data analysis is lowering barriers to developing and deploying deep learning models.

    “Technology that enables low or no code data analysis is lowering barriers to developing and deploying deep learning models.”

  • The paper relies on this without support.

    The data workers taking on AI/ML development do not have formal education in the theoretical and technical underpinnings of these systems.

    “Increasingly, the data workers taking on AI/ML development do not have formal education in the theoretical and technical underpinnings of these systems.”

  • The paper relies on this without support.

    The recruited participants are representative of the future data workforce.

    “We recruited and explored the perspectives of people who we believe are representative of the future data workforce.”

  • The paper relies on this without support.

    Model cards are intended to surface ethical and safety implications, a topic especially pressing for deep learning models.

    “Model cards are intended to surface, amongother things, ethical and safety implications - a topic that is especially pressing for deep learning models.”
12 of 12

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Model cards represent the field's standard for disclosing and documenting key facts about the architecture, training and evaluation of the model.

    “Model cards represent the field's standard for disclosing and documenting key facts about the architecture, training and evaluation of the model [14, 40].”

    From Rethinking open source generative AI: open washing and the EU AI Act (2024) · cites this paper as [14]

Cites

Flagged first, then by citations

Research works printed but not linked: 14

  1. Saleema Amershi, Maya Cakmak, W. Bradley Knox, and Todd Kulesza. 2014. Power to the People: The Role of Humans in Interactive Machine Learning. AI Mag. 35(2014), 105–120.
  2. Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. ” O’Reilly Media, Inc.”.
  3. Berkeley J. Dietvorst, Joseph P. Simmons, and Cade Massey. 2015. Algorithm aversion: people erroneously avoid algorithms after seeing them err.Journal of experimental psychology. General 144 1 (2015), 114–26.
  4. Alexandra Giannopoulou. 2020. Algorithmic Systems: the Consent is in the Detail?Internet Policy Review 9, 1 (2020).
  5. Karan Goel, Nazneen Rajani, Jesse Vig, Samson Tan, Jason Wu, Stephan Zheng, Caiming Xiong, Mohit Bansal, and Christopher Ré. 2021. Robustness Gym: Unifying the NLP Evaluation Landscape. arXiv preprint arXiv:2101.04840(2021). · arXiv
  6. Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. 2018. A Survey of Methods for Explaining Black Box Models. ACM computing surveys (CSUR) 51, 5 (2018), 1–42.
  7. Min Kyung Lee, Daniel Kusbit, Anson Kahng, Ji Tae Kim, Xinran Yuan, Allissa Chan, Daniel See, Ritesh Noothigattu, Siheon Lee, Alexandros Psomas, 2019. WeBuildAI: Participatory Framework for Algorithmic Governance. Proceedings of the ACM on Human-Computer Interaction 3, CSCW(2019), 1–35.
  8. Andreas Madsen Siva Reddy and Sarath Chandar. 2021. Post-hoc Interpretability for Neural NLP: A Survey. arXiv preprint arXiv:2108.04840(2021). · arXiv
  9. Christoph Molnar, Giuseppe Casalicchio, and Bernd Bischl. 2020. Interpretable Machine Learning–A Brief History, State-of-the-Art and Challenges. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 417–431.
  10. Michael Muller. 2014. Curiosity, Creativity, and Surprise as Analytic Tools: Grounded Theory Method. In Ways of Knowing in HCI. Springer, 25–48.
  11. Radim Rehurek and Petr Sojka. 2011. Gensim–python framework for vector space modelling. NLP Centre, Faculty of Informatics, Masaryk University, Brno, Czech Republic 3, 2(2011).
  12. Salome Viljoen. 2021. A Relational Theory of Data Governance. The Yale Law Journal 131, 573 (2021).
  13. Meredith Whittaker, Kate Crawford, Roel Dobbe, Genevieve Fried, Elizabeth Kaziunas, Varoon Mathur, Sarah Mysers West, Rashida Richardson, Jason Schultz, and Oscar Schwartz. 2018. AI Now Report 2018. AI Now Institute at New York University New York.
  14. Jilei Yang, Diana Negoescu, and Parvez Ahammad. 2021. Intellige: A User-Facing Model Explainer for Narrative Explanations. arXiv preprint arXiv:2105.12941(2021). · arXiv

Cited by

Sorted by citations

Related

Authors
Crisan, A., Drouhard, M., Vig, J., & Rajani, N.
Affiliations
  • Tableau Software (United States) (US)
  • Salesforce (United States) (US)
ORCID
Venue
2022 ACM Conference on Fairness Accountability and Transparency
Publisher
ACM
Published
2022-06-20
Pages
427-439
DOI
10.1145/3531146.3533108
OpenAlex
W4229447062
Cited by
101
References
56
Citations by year
2026: 24 · 2025: 31 · 2024: 22 · 2023: 23 · 2022: 1
FWCI
10.64
Type
conference paper
Language
en
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor)
Locations
  • ACM Conference on Fairness, Accountability, and Transparency (FAccT) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Data Visualization and Analytics
Keywords
model documentation, human-centered design, non-expert users, deep learning models, natural language processing, AI ethics, model interpretability, trust and safety, design guidelines, stakeholder perspectives, organizational dynamics, socio-technical context

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-06.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review