Experiences with Improving the Transparency of AI Models and Services

AI models and services are used in a growing number of highstakes areas, resulting in a need for increased transparency. Consistent with this, several proposals for higher quality and more consistent documentation of AI data, models, and systems have emerged. Little is known, however, about the needs of those who would produce or consume these new forms of documentation. Through semi-structured developer interviews, and two document creation exercises, we have assembled a clearer picture of these needs and the various challenges faced in creating accurate and useful AI documentation. Based on the observations from this work, supplemented by feedback received during multiple design explorations and stakeholder conversations, we make recommendations for easing the collection and flexible presentation of AI facts to promote transparency.

💡 Research Summary

The paper “Experiences with Improving the Transparency of AI Models and Services” investigates how the recently proposed FactSheet—a structured documentation format for AI models and services—can be adopted in practice and what challenges developers face when creating and using such documentation. The authors begin by noting the growing deployment of AI in high‑stakes domains (finance, healthcare, hiring, social services, policing, etc.) and the consequent demand for clear, consistent information about model purpose, data provenance, performance, fairness, and safety. While several documentation proposals (Data Sheets, Model Cards, FactSheets) have been put forward, little is known about the real‑world needs of those who must produce or consume them.

To fill this gap, the researchers conducted a two‑pronged formative study. First, they performed semi‑structured interviews with six data scientists and research scientists to elicit potential use cases for FactSheets. Participants highlighted four main scenarios: (1) understanding models inherited from external sources, (2) comparing multiple candidate models, (3) generating reports for non‑technical stakeholders such as business managers, and (4) integrating documentation directly into development environments (e.g., Jupyter notebooks, testing pipelines). Second, they organized two FactSheet creation sessions with AI developers. In the first session, nine developers were given a 39‑question FactSheet template (derived from Arnold et al., 2019) and asked to fill it out for a model they had built, with the intention of publishing it in an internal model marketplace. In the second session, six of those developers completed a streamlined 10‑question version during a one‑hour co‑development session, after which the researchers probed why each piece of information was included.

The findings reveal a broadly positive perception of FactSheets: all but the most senior interviewee and all participants in the creation sessions considered FactSheets valuable for a wide audience. Developers appreciated that FactSheets could capture essential “facts” about a model—training data, feature engineering, hyper‑parameter choices, intended use, and known limitations—in a single, reusable artifact. However, the study also uncovered concrete challenges. Items that can be derived automatically (e.g., accuracy, loss, dataset metadata) were easy to populate, whereas items requiring human judgment (e.g., ethical risks, bias mitigation strategies, appropriate use cases, safety considerations) were time‑consuming, often forgotten, and sometimes omitted entirely. Even the reduced 10‑question template proved insufficient for domain‑specific metrics; for example, a language‑translation model’s developers felt the need to include BLEU scores, a metric not captured by the generic template.

Based on these observations, the authors propose a design framework consisting of four interrelated components: (1) Automated collection pipelines that integrate with CI/CD to harvest reproducible metrics and metadata, reducing manual effort; (2) Human‑review checklists for the non‑automatable sections, ensuring that ethical, fairness, and safety considerations are explicitly addressed; (3) A flexible schema that defines a core set of mandatory fields (the 10‑question baseline) while allowing domain‑specific extensions via plug‑ins or custom fields; and (4) Web‑based registries that store FactSheets as searchable, filterable, and visualizable records rather than static PDFs, thereby supporting model marketplaces where users can compare candidates along multiple dimensions.

The paper situates FactSheets within the broader literature on software documentation, noting that traditional API docs focus on implementation details and usage examples, whereas AI documentation must also convey data provenance, model behavior under distribution shift, and societal impact. Consequently, FactSheets serve as a “model‑life‑cycle transparency record” rather than a mere technical manual. The authors argue that successful adoption hinges on balancing automation with expert input, embedding documentation tools into existing development workflows, and providing flexible presentation layers that meet the needs of regulators, auditors, and end‑users alike.

In conclusion, the study validates the practical utility of FactSheets while highlighting the need for systematic support mechanisms—automated metric extraction, structured human‑review processes, extensible schemas, and searchable repositories—to make FactSheet creation sustainable at scale. These insights offer a concrete roadmap for organizations seeking to implement AI governance frameworks, comply with emerging regulations (e.g., EU AI Act checklists), and foster trust in AI systems deployed across critical domains.