From Reflection to Repair: A Scoping Review of Dataset Documentation Tools

AI summary

Overview

Research area: Human-Computer Interaction (HCI) and Responsible AI (RAI), specifically the design, adoption, and standardization of dataset documentation tools (e.g., datasheets, data cards, documentation frameworks and applications).

Technical level: Intermediate. Readers will benefit from familiarity with Responsible AI and HCI concepts, but the paper is a qualitative and descriptive review rather than a technical or algorithmic contribution.

Scope (1 sentence): A PRISMA-guided scoping review and reflexive thematic analysis of 59 dataset documentation publications, examining the goals behind documentation tools, how their creators conceptualize documentation, and how the tools connect (or fail to connect) to systems, regulation, and cultural norms.

What This Paper Is About

Dataset documentation is widely promoted as essential to responsible development of automated systems, yet adoption and standardization remain stagnant despite a rapid growth of documentation frameworks, toolkits, and applications. The authors argue that while existing research has studied practitioners' perspectives on documentation, no one has systematically examined the role that documentation tools themselves—and the assumptions of their creators—play in creating adoption barriers. The paper's goal is to review all available dataset documentation tools, assess the values they implicitly and explicitly advance, and synthesize findings into recommendations for designing and integrating these tools into widespread community practice.

Key Contributions

1. A scoping review of the dataset documentation landscape. The authors analyze a corpus of 59 academic contributions across computing, machine learning, and Responsible AI venues to identify gaps and common threads that can inform adoption efforts.

2. A mixed-method characterization of fragmentation. They show how diversity in goals, conceptualizations of documentation, and integration efforts fragments the perceived value of documenting datasets, which in turn undermines adoption and standardization.

3. An account of systemic misalignment between RAI research and practice. The authors argue that documentation fragmentation stems from a deeper gap between Responsible AI research—specifically tool-building—and real-world practice.

4. A research agenda for institutional rather than individual design. They propose shifting RAI tool design away from solutions that place responsibility on individuals and toward institutional solutions, and outline actions the HCI community can take to enable sustainable documentation practices.

Main Findings

• Four persistent patterns impede adoption. The analysis identifies four recurring patterns in how dataset documentation is conceptualized: unclear operationalizations of documentation's value, decontextualized designs, unaddressed labor demands, and a tendency to treat integration as future work.

• Tool production is growing and diversifying. The total number of documentation tools has increased over time, with a concentration of new tools between 2022 and 2024 (31 out of 51 proposals). The authors identified seven tool types: applications, datasheets, frameworks, markup formats, schemas, studies, and toolkits. In the text, frameworks (n=16), applications (n=13), and datasheets (n=13) are described as the most popular types, amounting to 82% of the sample (note: the accompanying table lists 15 frameworks, 14 datasheets, and 13 applications).

• Automation is rising. Manual-input tools are the most prevalent (n=25), but recent years show an increase in hybrid tools that use natural language to automatically produce documentation (n=17) and automated tools that use machine-readable information (n=9), a trend the authors align with advances in large language models.

• Three competing views of documentation's value. Creators operationalize documentation as (1) personal reflection for dataset creators, (2) a means for dataset consumers to scrutinize datasets at scale, and (3) a way to repair broken dataset infrastructure. These views produce very different tool attributes, audiences, and intended impacts, even among tools built by similar groups of researchers.

• Evidence of impact is largely missing. Across all three viewpoints, the authors found that tools rarely provided evidence that their design choices achieved their stated aims—for example, that reflection-oriented designs increased users' reflexivity, that consumer-facing designs made dataset analysis faster or more effective, or that repair strategies actually repaired data ecosystems.

• Stakeholders are largely absent from design. More than half of the tools (n=30, reported as 58.8%) did not mention engaging relevant stakeholders during design. Of the remainder, most only elicited views without integrating them, and only two items in the sample described both eliciting and integrating stakeholder input.

• Academia and industry dominate authorship. Half the papers were authored solely by researchers in academia (31 papers, 52.54%), and a majority included at least one academic author (46 papers, 77.97%). Industry contributed to 27 papers (45.76%), and participated in 12 papers (20.34%) alongside other groups. Other participants included government (7 papers, 11.86%), non-profits (6 papers, 10.17%), and other (1 paper, 1.69%). By author count: academia (180 authors, 51%), industry (139 authors, 39.38%), government (20 authors, 5.67%), non-profits (13 authors, 3.68%), and others (1 author, 0.28%).

• Tool terminology and target users vary. Creators used diverse terms for their target users (e.g., dataset creators, dataset experts, data curators), with "dataset creators" the most common. Datasheets tended to name a wide variety of audiences, while most other tools failed to engage users in design.

• The regulatory picture is fragmented. In the US, the now-repealed Executive Order 14110 called for authenticating content and tracking provenance, and America's AI Action Plan emphasizes data quality standards, but there are no concrete federal guidelines on which frameworks to follow and no oversight mechanisms for compliance. States have begun legislating independently: California's Generative AI training data transparency act requires documentation of training data composition, provenance, and processing upon model release, and Colorado's AI Act identifies Data Cards specifically, requiring them alongside high-risk models. NIST's AI risk management guide highlights dataset documentation as a key strategic element. Independent efforts include Stanford's Foundation Model Transparency index, Behavioral Use Licensing, NeurIPS' ethics requirements, and Hugging Face's Data Cards.

• Documentation is hard in practice. Datasets are difficult to retract once published (one example describes falsified clinical trial data still cited positively and uncritically eleven years after retraction), post-release documentation struggles with competing versions and unrecorded procurement, and web-scale datasets like Common Crawl—with text from over 1 billion pages—make per-item documentation infeasible. Prior quantitative work found documentation comprehensiveness varied wildly, with dataset cards generally populated for popular datasets but absent for the vast majority overall, and fields on data structure and function far more detailed than social, ethical, or contextual fields.

Methodology in Plain English

The authors conducted a scoping review, a type of literature review suited to exploratory questions rather than quality evaluation or pre-registered hypothesis testing. They followed PRISMA guidelines for scoping reviews and reported findings using reflexive thematic analysis.

Search. They queried three databases (ACM Digital Library, IEEE Xplore, Science Direct), two domain repositories (ArXiv and ACL Anthology), and two conferences (AAAI and NeurIPS), using terms such as "dataset," "documentation," "transparency," "accountability," and "provenance." They also explored grey literature via Google's advanced search, using a "five-page-noise rule" to declare saturation after five pages of irrelevant results. This yielded 1,491 results across databases, 118 results across Google services, all collected in February 2025; after snowballing and removing duplicates, the total was 1,582 materials. Only English materials were included.

Screening. Inclusion criteria covered items that proposed or advanced a documentation tool, proposed a framework with at least one data documentation dimension, extended an existing tool, or empirically evaluated a tool in practice. Excluded were items offering only design recommendations, and case studies without evaluation or reflexive analysis. One author desk-rejected 1,408 items; 115 remained for calibration, in which three authors reviewed random samples of 20 items across three screening rounds with increasing Cohen's kappa agreement (κ = 0.29, κ = 0.54, κ = 1.0). The final corpus was 59 items.

Extraction. For each item, the authors recorded whether it was based on needs assessment or user research and what type; stated audiences; whether stakeholder involvement was mentioned versus actually performed; intended use; whether an evaluation was included; the degree of automation (manual, hybrid, or fully automated); integration with ecosystems, tools, or regulation; the terminology authors used; and author affiliations by sector.

Analysis. Three authors each took roughly 20 items, coded independently in ATLAS.ti, and developed 133 codes grouped into 26 clusters that were mapped onto the research questions, then refined iteratively into themes with memos and representative quotes.

Positionality. All four authors have formal training in HCI. They explicitly frame the review not as a critique of individual researchers but as a collective reflection, and acknowledge that the vast majority of reviewed work represents a Western approach to dataset documentation, which limits what the findings can represent about other epistemological positions and governing processes.

Why This Matters

Impact on research. The paper reframes stalled adoption of dataset documentation as a design and infrastructure problem rather than a practitioner-motivation problem. It shows that the same stated values—transparency and accountability—produce divergent tool designs, and that this divergence itself fragments the value of documentation. It also opens a specific empirical agenda: testing whether documentation designs actually produce reflection, scrutiny, or repair.

Real-world applications:

• AI developers and model release teams can use the four patterns as a checklist when choosing or building documentation tooling, particularly the warning that automated quick fixes may strip out reflection-oriented content (as the paper notes for a compacted version of Datasheets for Datasets).
• Regulators and standards bodies working on transparency requirements, such as those behind California's training data transparency act or Colorado's AI Act, can see why naming a framework like Data Cards without addressing organizational support and integration may not translate into compliance.
• Dataset creators and data curators can recognize the labor burdens that documentation tools impose, especially where tools claim that automation removes effort.
• Open dataset platforms and repositories (for example, sites that host dataset cards) can consider the paper's point that integration should be designed in from the start rather than treated as future work.

Industry relevance. Industry authors contributed to 27 of the 59 papers (45.76%) and participated in 12 papers (20.34%) alongside other groups, by author count accounting for 139 authors (39.38%). Documentation requirements increasingly appear in model release processes, ethics review, and licensing, so the finding that most tools did not engage stakeholders in design—and only two both elicited and integrated stakeholder input—has direct implications for whether corporate documentation practices will be sustained.

Future Directions

1. Build the missing evidence base. The authors repeatedly note that tools lack evidence that their designs produce the intended outcomes. A direct next step is evaluating whether reflection-oriented, scrutiny-oriented, and repair-oriented tools actually improve transparency, accountability, reliability, or dataset quality.

2. Design institutionally rather than individually. The paper calls for a shift in Responsible AI tool design toward institutional solutions, addressing the organizational, collective, and infrastructural support needed to sustain documentation. What those institutional designs look like in practice is left open.

3. Treat integration as a design requirement, not future work. The paper identifies integration as aspirational across the corpus. Concrete guidance on how tools connect to existing systems, regulatory frameworks, and cultural norms—and evidence that such integration works—remains an open problem.

4. Address labor and stakeholder engagement explicitly. With more than half of tools not mentioning stakeholder engagement and most offering general-purpose tools derived from creators' own needs rather than users', a next step is co-designing documentation tools with the data workers and communities who bear the labor.

Target Audience

HCI and Responsible AI researchers studying documentation, tooling, and data work practices; ML practitioners and data curators who select or implement documentation frameworks; policy and standards professionals working on AI data transparency; and dataset platform or infrastructure teams deciding how to embed documentation into their systems. The paper is most useful to readers who want a critical, evidence-oriented view of why dataset documentation tools have not achieved the adoption their advocates expected.

Authors’ abstract

Dataset documentation is widely recognized as essential for the responsible development of automated systems. Despite growing efforts to support documentation through different kinds of artifacts, little is known about the motivations shaping documentation tool design or the factors hindering their adoption. We present a systematic review supported by mixed-methods analysis of 59 dataset documentation publications to examine the motivations behind building documentation tools, how authors conceptualize documentation practices, and how these tools connect to existing systems, regulations, and cultural norms. Our analysis shows four persistent patterns in dataset documentation conceptualization that potentially impede adoption and standardization: unclear operationalizations of documentation's value, decontextualized designs, unaddressed labor demands, and a tendency to treat integration as future work. Building on these findings, we propose a shift in Responsible AI tool design toward institutional rather than individual solutions, and outline actions the HCI community can take to enable sustainable documentation practices.