TL;DR
This paper reveals a significant "transparency gap" in generative AI, where current tools primarily serve technical builders while leaving end-users underserved. By surveying 400 stakeholders, the authors demonstrate that transparency is not a static property but a situationally constructed need, requiring a shift from one-size-fits-all documentation to pluralistic, context-aware infrastructures.
What Problem Does This Paper Address?
While "transparency" is the most cited principle in AI ethics, there is a massive disconnect between the principle and its practice. Most current transparency mechanisms—like model cards or technical datasheets—are designed by engineers, for engineers.
The authors argue that transparency is "interpretively flexible," meaning its value and meaning change depending on who is asking and why. An end-user trying to decide if they can trust a medical chatbot has fundamentally different needs than a developer debugging a foundation model or a policymaker assessing regulatory compliance. The paper seeks to map these divergent needs and identify where current industry defaults are failing specific stakeholder groups.
Methods
The researchers employed a mixed-methods approach grounded in the Social Construction of Technology (SCOT) framework:
- Global Survey: 400 participants across three groups: end-users (n=250), AI builders (n=94), and policymakers (n=56).
- Feature Assessment: Participants rated 13 transparency features (e.g., confidence indicators, source citations, independent audits) on two scales: importance and perceived availability.
- Scenario Vignettes: To test how needs shift with context, participants were given low-stakes (e.g., a factual query about the North Pole) and high-stakes (e.g., a health query or model procurement) scenarios.
- Structured Focus Groups: Qualitative sessions with end-users to ground the quantitative survey data in lived experience.
Datasets
The study utilized a primary dataset generated from a pre-registered survey of 400 participants across 27 countries.
- Demographics: 60% from high-income countries, 37% from upper-middle-income.
- Recruitment: Conducted via the Prolific platform and professional networks.
- Benchmarks: The paper references the Stanford Foundation Model Openness Index to contextualize the current lack of industry transparency.
Key Insights and Ideas
- Shared Normative Foundation: Despite their different roles, all three groups agree on the "what" and "why." Protecting privacy and ensuring accuracy are the top priorities across the board.
- The Availability Gap is a Power Signal: While importance ratings were high for everyone, builders perceived transparency features as much more available than end-users did. This suggests that transparency tools are currently built to satisfy those in power (the developers) rather than those most impacted by the technology.
- Transparency is Situationally Constructed: Priorities are not fixed. In low-stakes scenarios, everyone wants simple explanations and sources. In high-stakes health scenarios, end-users’ demand for "usage limits" (knowing when not to use the tool) skyrocketed.
- The Failure of Static Documentation: Popular tools like "model cards" are largely invisible to end-users. Only 29% of end-users were even aware of system cards, and those who were found them significantly less useful than builders did.

Key Results
- Subjective Needs Alignment: Only 28% of end-users report that current transparency practices reflect their needs "very well," compared to 53% of builders and 49% of policymakers.
- Quantifying the Gap: The "gap score" (importance minus availability) was nearly double for end-users (35 points) compared to builders (17 points).
- Universal Unmet Needs: Certain features, such as confidence tagging (how certain the AI is about a specific output), showed massive gaps across all groups, indicating a major technical and design opportunity for the industry.
- Contextual Shifts: In the health scenario, end-users' selection of "usage limits" as a top-three feature rose from 12% to 44%.
Limitations and Future Work
The authors acknowledge several limitations:
- Perceived vs. Objective Availability: The study measures whether users think a feature is there, not if it actually exists in the code.
- Binary Availability: The survey used a yes/no/unsure measure for availability, which may not capture the nuance of "partial" transparency.
- Sample Bias: The focus groups were limited to English speakers with reliable internet access.
- Snapshot in Time: As a cross-sectional study, it doesn't capture how needs might evolve as users become more "AI literate."
Future Directions: The authors call for longitudinal studies to track how transparency needs change as regulations (like the EU AI Act) take effect and suggest that "participatory design" with underserved end-users is the only way to close the current power gap.
Summary updated