AI transparency serves builders better than end-users, survey suggests

Shared goals, unequal experiences

Generative AI transparency may look different depending on who needs to use it. In a November 3, 2026 arXiv preprint, Genevieve Smith, Sahiba Chopra, Natalia Luka and Min Kyung Lee report that end-users, AI builders and policymakers share broad priorities around accuracy, privacy and reliability. Yet end-users perceive substantially greater shortfalls in the transparency available to them.

The paper is a preprint; the supplied material does not establish peer-reviewed publication. Its findings map participants’ reported needs and experiences, rather than independently verifying which transparency features AI products actually provide.

What the researchers measured

The authors describe a preregistered survey conducted from October 2025 through April 2026, with 250 end-users, 94 builders and 56 policymakers. Recruitment combined Prolific with professional-network outreach and snowball sampling. A structured end-user focus group supplied qualitative context.

Participants rated the importance of 13 transparency features and reported whether those features were available in their latest or regularly used generative AI tool. Features included citations, plain-language explanations, confidence indicators, usage limits, training-data information, privacy policies, independent audits and ways to challenge harmful outputs.

The researchers converted importance ratings to a 0–100 scale and subtracted perceived availability, measured as the percentage answering yes among those answering yes or no. Unsure responses were excluded from that denominator.

These “gap” scores combine different kinds of measurement: graded importance ratings and binary availability reports. They are useful for mapping perceived unmet needs, but should not be read as precise product-performance scores.

End-users reported the largest shortfalls

Only 28% of end-users said current transparency practices reflected their needs very well, compared with 53% of builders and 49% of policymakers.

The main text reports average gaps of approximately 35 points for end-users, 17 for builders and 22 for policymakers. Figure 2 gives somewhat different averages—35.6, 19.4 and 23.0 respectively. Both presentations show the same ordering, but the discrepancy warrants caution about exact magnitudes.

Documentation illustrates the divide. Just 29% of end-users reported awareness or use of system cards, compared with 73% of builders and 68% of policymakers. Model-card awareness was similarly uneven: 32% among end-users, versus 76% among builders and 70% among policymakers. Among participants who knew about these resources, end-users also rated them as less useful.

The authors interpret these differences as consistent with transparency infrastructure being more accessible and legible to technical communities. That is an interpretation of the observed pattern, not a demonstrated causal explanation. The survey cannot establish whether a feature was absent, difficult to find, misunderstood or judged inadequate.

Context changed what people wanted

In a shared low-stakes information scenario, sources, explanations and confidence indicators were among the most-selected features across all three groups.

Priorities shifted in role-specific scenarios. End-users’ selection of usage limits rose from 12% in the neutral scenario to 44% in a health scenario. For builders, it rose from 15% to 38% when choosing a model for a product. Builders also became more likely to select at least one governance feature, increasing from 56% to 77%.

These comparisons do not isolate the effect of higher stakes. The scenarios also changed subject matter and task structure, and the builder scenario offered a different feature menu. The authors suggest that choosing a system can activate different concerns from checking an individual answer.

What the findings support—and what they do not

The authors recommend retaining documentation while making its information accessible in other forms, involving underserved users in design, and adapting transparency to context. They also identify confidence indicators as a widely unmet need. However, this study does not test whether such indicators are accurate or improve decisions; demand for a feature is not evidence of its effectiveness.

The sample is international but not globally representative. Participants had to speak English, recruitment was non-random, and only 12 respondents belonged to the paper’s lowest country-income grouping. The appendix also introduces an unresolved count inconsistency by listing four additional “other” respondents alongside role counts already totaling 400.

The strongest takeaway is therefore narrower than a verdict on the whole AI industry: in this sample, shared transparency goals coexisted with unequal reported access. The authors’ proposed response is not less disclosure, but information that different audiences can find, understand and use.