Whose AI transparency needs are being met? A 400-person survey found the widest gap for users
In an unreviewed survey, 28% of 250 end-users said current transparency practices reflected their needs very well, versus 53% of 94 builders and 49% of 56 policymakers. The study measures perceptions—not whether products objectively disclose enough.
The short version
- What happened
- The preregistered English-language survey included 400 adults: 250 generative-AI end-users, 94 builders and 56 policymakers. Recruitment used Prolific and professional-network or snowball outreach from October 2025 to April 2026.
- Who it affects
- People using AI for health, work or information may need sources and limits at the moment of use, not only a technical document elsewhere.
- What remains unproven
- Unreviewed cross-sectional preprint measuring self-reported importance and perceived availability, not objectively verified product disclosure.
Keep in mind
Unreviewed cross-sectional preprint measuring self-reported importance and perceived availability, not objectively verified product disclosure.
Non-probability recruitment differed by group: Prolific dominated end-user recruitment, while professional networks and snowballing were more important for policymakers.
The sample was English-speaking, concentrated in the United States and South Africa, and included only 12 people in the low- and lower-middle-income exploratory subgroup.
The gap score subtracts a binary availability percentage from a rescaled Likert importance mean and excludes unsure responses, complicating interpretation.
One appendix says four respondents were classified as other after listing groups that already total 400; the manuscript does not reconcile the denominator inconsistency.
The 20-person focus group was self-selected, online and illustrative, with qualitative coding led by one team member and no reported inter-coder agreement.
No behaviour, comprehension, compliance, redress, safety outcome, trust calibration or reduction in harm was measured.
Source analysis · 1 linked source. Primary analysed source date: 3 November 2026.
Read the full report · 10 min The evidence, methods, implications and limitations
The answer: users and builders wanted similar goals, but users perceived fewer useful tools
An unreviewed preprint finds that end-users, AI builders and policymakers broadly agreed on what transparency should help them achieve—especially accuracy, privacy, safety and trust—while reporting sharply different experiences of what is available. Only 28% of the 250 end-users said current transparency practices reflected their needs very well, compared with 53% of 94 builders and 49% of 56 policymakers. Across 13 features, end-users' average importance–availability gap was 35 points, versus 17 for builders and 22 for policymakers.
The result is not an objective score for any company or system. Participants rated how important features were and whether they perceived them in the latest or most-used generative-AI tool. A person can miss a feature that exists, find its language unusable or encounter a genuinely absent disclosure; the survey does not distinguish those explanations. Its contribution is to show whose experience can disappear when technical documentation is treated as sufficient. It does not prove that more disclosure alone produces accurate outputs, safer products or trust.[1]
Who took part, how they were recruited and why the groups are not interchangeable
The preregistered survey ran from October 2025 through April 2026. Participants were adults, English-speaking and located across approved countries in North America, South America, Europe, Africa, Asia and Oceania. End-users had to use generative AI in their main role; builders needed AI product-management or technical-development experience; policymakers needed experience informing or advising on AI regulation or governance. End-users and builders were recruited mainly through Prolific, while policymakers relied more heavily on the researchers' professional networks and snowball referrals. Each survey took about 20 minutes and paid US$6.25.
The main methods describe 250 end-users, 94 builders and 56 policymakers, which sum to 400. The demographic table reports 144 US residents and 126 South African residents; only 12 respondents fell in the combined low- and lower-middle-income group used for one exploratory comparison. Recruitment route and group differ together, so a role difference can also reflect sampling, occupation or network. One appendix confusingly adds four respondents classified as 'other' after listing the same three groups and total of 400. The preprint does not resolve that arithmetic inconsistency, so the report relies on the main analysis denominators and treats the appendix line as a limitation needing correction.[1]
The distinction that matters
What the survey compared
The figures are perceptions in a recruited English-speaking sample, not objective product-compliance scores.
End-users
250 surveyed
Mean gap 35 points; 28% said needs reflected very well
Builders
94 surveyed
Mean gap 17 points; 53% said needs reflected very well
Policymakers
56 surveyed
Mean gap 22 points; 49% said needs reflected very well
Focus group
20 end-users
Illustrated why sources, limits and privacy mattered
How the gap score was constructed—and what it cannot say
Participants rated the importance of six user-facing features—AI labels, plain-language explanations, confidence indicators, usage limits, citations and system prompts—and seven governance features including training-data information, documentation, impact reporting, independent audits, benchmarks, privacy policies and recourse. Importance used a five-point Likert scale rescaled to zero–100. Perceived availability was the percentage answering yes among people who answered yes or no. Subtracting availability from importance produced a gap for each feature and group.
That subtraction is intuitive but methodologically imperfect. It combines a graded rating with a binary perception and excludes 'unsure' responses from the availability denominator. Non-response was concentrated in governance items: 121 respondents could not say whether independent audits were available, 112 for training data, 97 for documentation and 96 for social and ethical impact reporting. A high gap may indicate genuine absence, low discoverability, inaccessible wording or uncertainty. It should guide questions and design work, not rank companies or certify compliance. The authors say the measure maps stakeholder experience rather than auditing the transparency landscape.[1]
The largest difference appeared in whether people could use established documentation
Every feature was rated important on average, with group means ranging from 3.97 to 4.75 on the five-point scale. End-users rated citations and sources more highly than builders and policymakers, while policymakers placed more weight on training-data information, social and ethical impact reporting and independent audits. Yet only 29% of end-users said they were aware of or had used system cards, compared with 73% of builders and 68% of policymakers. Model-card awareness was 32% among end-users, 76% among builders and 70% among policymakers.
Among people who knew the tools, end-users also rated them as less useful: system cards averaged 69 out of 100 for aware end-users and 80 for builders. This does not mean documentation should be removed. It suggests that publishing a technical artefact is different from making its content findable and usable at the moment someone needs to judge an output, privacy risk or limit. The paper argues for plural forms of transparency, including concise user-facing information alongside governance documentation. Whether those forms change behaviour or reduce harm remains untested.[1]
Needs changed when the task became higher stakes
The survey first showed everyone the same low-stakes information task and then a higher-stakes scenario matched to role: a health query for end-users, policy research for policymakers and model procurement for builders. Participants selected three useful transparency features. End-users chose usage limits 12% of the time in the neutral scenario and 44% in the health scenario. Builders' selection rose from 15% to 38% in procurement. In the role-specific comparison, end-users selected sources more often than policymakers or builders—72% versus 55% and 48%—although role and scenario changed together.
That design supports within-person observations about context but cannot cleanly attribute cross-group differences to profession. The policymaker scenario still involved evaluating a chatbot response, while the builder scenario involved a procurement decision; the tasks were not equivalent. The useful practical finding is that a static disclosure bundle is unlikely to serve every moment. A health query may need prominent usage limits and sources near the answer; procurement needs model, benchmark, data and oversight information. The study did not test which interface communicates these facts accurately or whether users act on them.[1]
What the focus group added—and why it remains illustrative
Survey end-users could volunteer for participatory design. The researchers randomly selected 20 from eligible volunteers while limiting US participation and seeking gender and regional representation. Participants were paid US$40 per session for three two-hour online sessions in February and March 2026; this paper draws on the first structured focus group. About one-third were outside the United States, with participants from Mexico, India, South Africa, Canada and the United Kingdom. They brainstormed transparency needs, clustered ideas, reviewed three misleading-output vignettes and voted on features.
The researchers coded workshop artefacts and transcripts in NVivo. One team member completed initial coding, the full team refined the codebook and a final round followed. Participants repeatedly connected sources and confidence information with checking accuracy, and privacy information with organisational accountability. These accounts help interpret survey numbers, but 20 self-selected people with reliable internet and English fluency are not a representative global public. The paper reports no independent second-coder agreement for the qualitative material and no field test of the proposed transparency designs.[1]
What product teams and policymakers can do without overstating the evidence
Product teams can test whether ordinary users can locate and understand sources, usage limits, confidence information, privacy terms and recourse at the point of decision. They can recruit people outside technical networks, separate discovery from comprehension and measure whether the information changes appropriate reliance. System and model cards should remain available for experts and oversight, while their most consequential facts can also appear in plain language inside the product. A disclosure that is technically public but practically invisible has limited value for the person facing an immediate decision.
Policymakers should resist turning the percentages into a universal checklist. The sample is weighted toward the United States and South Africa, the policymaker group is small and recruited differently, and only 12 participants supported the low-income comparison. Confidence would rise with peer review, corrected denominators, probability-based multilingual sampling, preregistered replication and objective audits of the products participants used. Stronger evidence would connect particular disclosures to calibrated trust, safer decisions, successful redress or reduced harm. The present study identifies a design and power problem; it does not demonstrate that transparency by itself fixes model performance or accountability.[1]
Funding and disclosure boundaries
The preprint lists authors at Stanford University, the University of California Berkeley and the University of Texas at Austin. It reports institutional ethics approval in anonymised form and says participants gave informed consent. The extracted manuscript does not present a dedicated funding or competing-interest statement, so no claim of independent funding can be made. It discloses that Gemini was used for grammar and writing fluency. That is an authorship-process disclosure, not evidence that Gemini collected or analysed the study data.
The most defensible conclusion is that the surveyed end-users perceived a wider mismatch between important transparency information and what they could find than the builders and policymakers in this recruited sample. The study cannot tell whether a particular provider complies with law, whether every reported feature was truly absent or whether the percentages generalise to all countries and users. Its value is to redirect evaluation toward the people expected to use disclosures. The assessment should change when replicated work audits actual interfaces and shows which forms of transparency help people make better consequential decisions.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring product design: compare the finding and its evidence limits.
What is changing in ChatGPT for teens?
OpenAI says a US College Planner for grades 10–12 is coming, alongside flashcards, easier quizzes and multi-photo note capture. The company also released large usage counts, but no study protocol, denominators for several comparisons or evidence that the tools improve learning or admissions outcomes.
Read the next reportExplore the full source and evidence record
What this means for people
- People using AI for health, work or information may need sources and limits at the moment of use, not only a technical document elsewhere.
- Builders can mistake publication for accessibility unless they test whether non-technical users can find, understand and act on the information.
- Policymakers need evidence from more representative multilingual samples before turning the reported gaps into universal requirements.
Global context
The protocol sought participants across six continents, but residence was concentrated in the United States and South Africa and every participant had to speak English. Transparency rules and cultural expectations differ, as do internet access and familiarity with technical documentation. The results justify multilingual local research and user testing; they do not establish a worldwide ranking of needs or products.
What the evidence does not yet show
- Unreviewed cross-sectional preprint measuring self-reported importance and perceived availability, not objectively verified product disclosure.
- Non-probability recruitment differed by group: Prolific dominated end-user recruitment, while professional networks and snowballing were more important for policymakers.
- The sample was English-speaking, concentrated in the United States and South Africa, and included only 12 people in the low- and lower-middle-income exploratory subgroup.
- The gap score subtracts a binary availability percentage from a rescaled Likert importance mean and excludes unsure responses, complicating interpretation.
- One appendix says four respondents were classified as other after listing groups that already total 400; the manuscript does not reconcile the denominator inconsistency.
- The 20-person focus group was self-selected, online and illustrative, with qualitative coding led by one team member and no reported inter-coder agreement.
- No behaviour, comprehension, compliance, redress, safety outcome, trust calibration or reduction in harm was measured.
What to watch next
- Peer review and correction of the appendix denominator inconsistency.
- Multilingual probability-based replication with larger policymaker and lower-income-country samples.
- Objective audits matching reported perceptions with the features actually available in named products.
- Controlled tests of whether sources, limits, confidence information and privacy disclosures improve calibrated reliance.
- Evidence that participatory design changes product decisions and accountability rather than serving as consultation alone.
Living evidence record
Impact record IAI-1HODD2I
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
How end-users, AI builders and policymakers value and perceive generative-AI transparency features across ordinary and higher-stakes contexts
Topics: AI transparency · End-users · Policy · Product design · Trust · Participatory research
Regions: Global · United States · South Africa
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.