AI Transparency Isn't One Document — It's Several, for Different Readers

In a global survey of 400 end-users, AI builders, and policymakers, 28% of end-users said current transparency practices reflected their needs very well, compared with 53% of builders and 49% of policymakers. ¹ The 25-point difference between end-users and builders is worth examining: a disclosure designed around one group may not give another group what it needs.

The answer isn't necessarily more documentation. It's asking who needs what information, for which task, and at what moment. The study argues that transparency should move beyond one-size-fits-all approaches, including an emphasis on AI disclosure and static documentation, toward infrastructure that reflects diverse stakeholder needs. ¹

What breaks when everyone gets the same AI disclosure

End-users, builders, and policymakers share similar understandings of transparency, centered on accuracy and privacy. ¹ Shared priorities don't mean every group needs the same information. As an implementation example, someone deciding whether to rely on an answer might need information about its limitations. A builder investigating a failure might need operational details. A policymaker reviewing accountability might need evidence that supports oversight. These are proposed examples, not findings about specific tasks measured by the study.

A single notice may address accuracy and privacy in general terms without answering the questions a particular person has for a particular task. That is a design risk to test, not a result the study directly measured.

The figures make the difference in reported needs concrete: 28% of end-users, 53% of builders, and 49% of policymakers said current practices reflected their needs very well. ¹ The authors also report that the gap between the importance of transparency features and their perceived availability is substantially larger for end-users than for builders. ¹

Analysis: Read those findings as a prompt to test transparency practices by audience, not as an external scorecard. The survey covered end-users, builders, and policymakers; the source summary does not establish a complete disclosure standard for every workplace or oversight setting. ¹ A useful question for organizations is: what information lets this person do their part?

What the study on AI stakeholder transparency actually measured

The paper, “Whose Transparency is it Anyway? Understanding Generative AI Transparency Needs and Gaps Across Stakeholders,” was published on arXiv on November 5, 2026. ¹ It draws on a global survey of 400 end-users, builders, and policymakers, complemented by a structured focus group with end-users. ¹

One finding relevant to process design is that transparency is situational: priorities shift with context and task structure, and role-specific frames become more or less relevant as context changes. ¹ That suggests a static checklist may be a weak substitute for asking what someone is trying to do. For example, reviewing a routine draft and deciding whether to act on a consequential recommendation could call for different information. This is an implementation illustration, not a comparison tested in the study.

The source summary does not specify respondents' geographic distribution, recruitment method, demographics, focus-group size, or analysis method. ¹ It describes the surveyed groups as end-users, builders, and policymakers, and does not provide separate findings for employees, customers, or auditors. ¹ The summary supports adapting transparency to context, but does not specify which fields or controls an organization should implement.

Matching AI transparency to each audience

The study does not provide a universal feature list. ¹ The table below is an implementation proposal based on its finding that stakeholder needs differ; it is not a set of features the researchers tested.

Audience	Information to consider	Governance purpose
End-users	What the system is used for, relevant limitations, and how to raise a concern	Judge whether and how to rely on an output
Builders	Which model or service is involved, what changed, and which failure modes matter	Diagnose and maintain the system
Policymakers and auditors	Evidence on accuracy, privacy, accountability, and whether controls operate as described	Support review and oversight
Internal employees	Permitted uses, review expectations, sensitive-data handling, and escalation routes	Connect transparency to work practices
Customers	Concise information at the point of use, with a path to more detail	Explain the interaction without overwhelming people with technical detail

This is not a claim that every group needs its own document. It is a way to start from the task instead of the artifact. For example, an organization might provide a concise customer-facing explanation and maintain a more detailed internal record for investigation. That proposal follows the study's emphasis on context, but customer preferences were not separately reported in the source summary. ¹

Why generative AI disclosure alone doesn't scale

The authors argue against treating disclosure statements and static documents as the main transparency mechanism, and call for infrastructure that accounts for different users. ¹

Analysis: One way to operationalize that argument is to connect three layers rather than rely on one document:

1. Point-of-use information. A short explanation available when someone interacts with an AI feature or acts on its output.
2. Task-specific guidance. More detail for questions that arise during use, such as how to handle uncertainty or report a concern.
3. Durable records. Information for authorized people who may need to investigate or review the system later. Controls such as audit logs and usage analytics in an enterprise AI gateway are one place such records can come from.

This three-layer split is an article-level framework, not one named in the paper. ¹ It is intended to reduce the risk that a brief notice leaves builders without information to investigate, or that a technical record gives ordinary users little they can act on. Those are design concerns, not study findings.

More information is not automatically more useful. Providing every audience with the same technical record could make it harder for some people to find what they need. The goal is to match the depth of information to the audience while making more detail available to people who need to investigate.

Building a stakeholder-task matrix

Analysis: Start with one AI use case and a stakeholder-task matrix. For each audience, ask:

• What decision or action depends on this system?
• What could go wrong in this specific task?
• What would let this person judge accuracy, privacy, or accountability?
• Where and when would that information reach them?

This matrix is a proposed practice, based on the study's finding that needs shift with context and task structure; it is not a method specified in the paper. ¹

From there, set a minimum point-of-use disclosure for the use case. Add detail where the task or audience calls for it rather than assuming everyone needs the same information. This is an application of the study's argument against one-size-fits-all transparency. ¹

Then build a path from user-facing information to internal review: a way to report a problem and a way for an authorized reviewer to connect the report to the relevant system and decision context. This is a design recommendation, not something the study measured. If the system involves agents with external access, the governance questions multiply; see how external agent access becomes a governance problem.

Finally, test the information with the people meant to use it. The reported difference between end-users and builders is a reason to check whether a disclosure answers real questions, rather than assuming that its existence makes it useful. ¹ Revisit it when the audience, task, or system changes; the paper reports that transparency priorities shift with context. ¹

What to measure before calling transparency effective

Measure whether people can find and use information relevant to their task, not just whether a policy document exists. The study distinguishes the importance of transparency features from their perceived availability and reports a substantially larger gap for end-users than for builders. ¹

Analysis: Track feedback by audience instead of relying only on an aggregate measure. A single average could obscure differences between groups; the study's end-user and builder results illustrate why group-level results can matter. ¹ For each use case, consider checking whether people can:

• Spot relevant limitations and uncertainty.
• Understand how their data or privacy is handled.
• Decide when to question or escalate an output.
• Find a route to raise a concern.

These are proposed operational checks, not metrics reported by the study. They translate its focus on accuracy, privacy, and differences among stakeholders into questions an organization can test. ¹ Use the results to simplify or expand each piece of information as needed. An aggregate result alone may not show whether a particular stakeholder group is being served.

AI governance practices: what leaders should do next

Pick a consequential generative-AI use case and map its audiences before writing new disclosure language. Ask the people who use, build, oversee, or are affected by the system what they need, then compare those needs with what the current process provides. The gap between end-users' and builders' reported experiences is a reason to examine whether existing notices serve both groups. ¹

Move away from the idea of one all-purpose transparency document. Consider linked, audience-specific information: a short notice at the point of use, operational detail for builders, and review evidence for oversight. This is an organizational recommendation drawn from the study, not a design the authors prescribe. ¹ For multi-agent systems, the same logic applies to internal boundaries; role contracts need a gate if they are to mean anything.

Write down what remains unknown. The research supports adapting transparency to context, but does not provide a universal feature list or separate findings for employees, customers, or auditors in the source summary. ¹ Make design choices explicit rather than letting a checklist imply the work is complete.

FAQ

What does AI transparency mean in a governance context?
Providing information that helps stakeholders evaluate an AI system in their context. The study reports shared emphasis on accuracy and privacy, alongside differences in how well current practices meet each group's needs. ¹

Why isn't disclosure enough on its own?
A static notice may not provide the information a particular audience needs for its task. The authors argue for moving beyond an emphasis on disclosure and static documentation toward infrastructure that reflects diverse stakeholders' needs. ¹

What information do AI users specifically need?
The source summary does not give a universal list. It reports that 28% of end-users said current practices reflected their needs very well, which supports testing disclosures against the questions users actually have. ¹

Does one transparency checklist cover every stakeholder?
The paper argues against a one-size-fits-all solution. Its findings show shared high-level concerns but differences in how well current practices meet each group's needs; end-users reported the lowest figure among the three groups. ¹

---

If you run AI governance, take a use case with recurring complaints or support requests, apply the four-question matrix above, and check whether current information serves audiences beyond the people who build the system. This is a suggested exercise, not a method tested by the study.