Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI - Paper

Sep 24, 2026

Upload by nayy☆

Data Cards provide structured, human-centered documentation for machine learning datasets across their full lifecycle, supporting responsible AI development in research and industry. The paper addresses how transparent, comparable, and user-centric dataset explanations can reduce burdens on intelligibility, conciseness, and comprehensiveness. It proposes Data Cards as boundary objects, includes rationales beyond metadata (provenance, representation, usage, fairness-informed evaluations), and presents frameworks validated through case studies and lessons learned from deploying 20+ Data Cards.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages51

Reading Time18 min

AccessFree to download

1 INTRODUCTION

Transparency and explainability through datasets

The Data Cards framework

Data Cards as boundary objects

Summary contributions

arXiv : 2204 .0 1075v 1 [cs .HC] 3 Apr 2022 Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI MAHIMA PUSHKARNA, ANDREW ZALDIVAR, ODDUR KJARTANSSON, Google Research,Canada, United States, United Kingdom Fig. 1. A page from a static, PDF data card describing a text dataset. The completed Data Card does notinclude questions from the template, but preserves its thematically contained row-and-column structure.Blocks increase in detail from left to right, and authors have introduced links to elegantly expose readers toadditional documentation using context offered in the Data Card. As research and industry moves towards large-scale models capable of numerous downstream tasks, thecomplexity of understanding multi-modal datasets that give nuance to models rapidly increases. A clearand thorough understanding of a dataset’s origins, development, intent, ethical considerations and evolu -tion becomes a necessary step for the responsible and informed deployment of models, especially those inpeople-facing contexts and high-risk domains. However, the burden of this understanding often falls on theintelligibility, conciseness, and comprehensiveness of the documentation. It requires consistency and compa -rability across the documentation of all datasets involved, and as such documentation must be treated as auser-centric product in and of itself. In this paper, we propose Data Cards for fostering transparent, purposefuland human-centered documentation of datasets within the practical contexts of industry and research. Data 2022-04-05 01:22 . Page 1 of 1–51 . 1 Cards are structured summaries of essential facts about various aspects of ML datasets needed by stakeholdersacross a dataset’s lifecycle for responsible AI development. These summaries provide explanations of processesand rationales that shape the data and consequently the models—such as upstream sources, data collection andannotation methods; training and evaluation methods, intended use; or decisions affecting model performance.We also present frameworks that ground Data Cards in real-world utility and human-centricity. Using twocase studies, we report on desirable characteristics that support adoption across domains, organizationalstructures, and audience groups. Finally, we present lessons learned from deploying over 20 Data Cards.x CCS Concepts: • Social and professional topics → User characteristics; • General and reference →Evaluation; • Software and its engineering → Software creation and management; • Human-centeredcomputing; Additional KeyWordsand Phrases: data cards, dataset documentation, transparency, responsible AI, datasheets,model cards # 1 INTRODUCTION The challenge of transparency in machine learning (ML) models and datasets continues to receiveincreasing attention from academia and the industry [1, 2] . Often, the goal has been to attaingreater visibility into ML models and datasets by exposing source code [4], contribution trails [7], introducing ML-drive data analysis methods [18], and introducing diverse oversight [17] .Transparency and explainability of model outcomes throughthelens of datasets has become a hugeconcern in regulation from government bodies internationally. However, attempts to introducestandardized, practical and sustainable mechanisms for transparency that can be create value at scaleoften meet limited success in research and production contexts. This reflects real world constraintsof the diversity of goals, workflows, and backgrounds of individual stakeholders participating inthe life cycles of datasets and artificial intelligence (AI) systems [10,12,13] . As a step towards creating value that connects dataset success to research and productionexperiences, we propose a new framework for transparent and purposeful documentation ofdatasets, called Data Cards. A Data Card contains a structured collection of summaries gatheredover the life cycle of a dataset about observable (e.g., dataset attr

FAQs

What are Data Cards intended to achieve for responsible AI?

They foster purposeful, transparent, human-centered documentation of ML datasets so stakeholders can make informed decisions during the dataset lifecycle.

What information do Data Cards include beyond basic metadata?

They add explanations, rationales, and instructions about provenance, representation, usage, and fairness-informed evaluations that cannot be derived from the dataset alone.

How do the paper validate Data Cards in real settings?

It presents frameworks and uses two case studies to report desirable characteristics for adoption across domains, organizations, and audiences, along with lessons learned from deploying 20+ Data Cards.

Sep 24, 2026

Upload by nayy☆