By Hendrik Schäfer, Christoph Renschler
25 min read
Cite this report
This report is published by Musa Guide under a Creative Commons Attribution 4.0 (CC BY 4.0) licence. When citing or republishing, you must credit the authors and include a working hyperlink back to the canonical article at https://www.musa.guide/zh-HK/research/state-of-ai-in-museums-2026. Plain-text attribution alone is not sufficient.
AI is already inside the museum. In a population-adjusted estimate across the United Kingdom, United States, Germany, and France, roughly 1 in 3 adult museum visitors used a general-purpose AI assistant (ChatGPT, Gemini, Claude, Copilot, and the like) for museum content in the past 12 months. On the most recent museum visit, the figure was roughly 1 in 6. Among visitors who had used AI for museum content in the past year, AI was used more often than audio guides on the most recent visit: 54.7% used AI compared with 41.0% who used a free or paid audio guide.
The deeper finding is that museums have not lost visitor trust; they are losing the visitor interface. 61% of AI-using visitors say they trust the museum more than AI, but only 27% verified what AI told them. The unmet need is broader than AI: 50% of all visitors report at least occasional navigation confusion, and 35% had a question they had no easy way to ask. Stated willingness to pay for an audio guide on top of the ticket clusters at £3 to £5.50, with the optimal price around £4.55 in the Van Westendorp analysis.
Method: 400 adult museum visitors surveyed via Prolific in May 2026, pre-registered on OSF. The unweighted Prolific sample over-represents AI-fluent visitors (raw past-12-month rate 53.7%); after adjusting for country-level AI-use frequency, our population-adjusted estimate is 32%. Full methodology and limitations in Methodology.
An estimated 32% of adult museum visitors in the UK, US, Germany, and France used a general-purpose AI assistant (ChatGPT, Gemini, Claude, Copilot, etc.) in the past 12 months to ask about something they saw at a museum.
On the most recent museum visit specifically, 17% used AI. Past-12-month use skews younger — 38.5% among under-45s, 22% for 45+. The most-recent-visit cohort split was too noisy to report at sub-population level. Both reweighted estimates use mid-2025 country anchors, so the present-day population figures are likely higher (see reweighting section).
In the raw sample, the behavior is strongly concentrated among frequent AI users: among visitors who use AI at least weekly for any purpose, 35% used AI on their most recent museum visit, and 59% in the past 12 months. Among visitors who use AI daily, those rates rise to 44% and 67% respectively. Among visitors who use AI rarely or never, past-12-month museum AI use was zero in our sample, and most-recent-visit use was very rare.
Because Prolific over-represents digitally fluent users, we report both raw sample rates and population-adjusted estimates. The headline figures use the adjusted estimates; the full weighting procedure is in the reweighting section.1
The analytic sample was 54% UK, 20% US, 14% Germany, and 12% France; median age was 34. AI fluency and age splits are reported below.
Figure 1.1
UK / US / DE / FR pooled, reweighted to 2025-2026 country-level weekly+ AI-use shares. Whiskers show 95% Wilson confidence intervals on the raw sample proportions (n=395), which understate the additional uncertainty introduced by reweighting (see Limitations).
Past 12 months: respondents who used an AI assistant at least once in the past year to ask about something they saw at a museum. Most recent visit: respondents who used AI on the most recent museum trip. The under-45 estimate for the most-recent-visit metric is omitted because its reweighted bootstrap CI was dominated by a small occasional-AI-use cell (32 under-45s, 2 yeses).
Source: Musa Guide, State of AI in Museums 2026
Generative AI assistant uptake has grown quickly relative to prior consumer technologies, per the figures below.
The Reuters Institute Generative AI and News Report measured weekly use of AI assistants among adults across six Western markets going from 18% in mid-2024 to 34% in mid-2025, roughly doubling in twelve months.2 Single-country surveys fielded in early 2026 land in the same direction:
Indexed against the smartphone adoption curve since the iPhone launch in January 2007, AI assistant adoption among US adults since the ChatGPT launch in November 2022 is running steeper in its first three years than smartphones did in theirs (Figure 1.2).
Figure 1.2
Share of US adults who have ever used a smartphone (since iPhone launch, January 2007) versus who have ever used ChatGPT (since ChatGPT launch, November 2022), indexed to years since launch.
Both series use the closest ever-used or owns-one metric available from Pew. Both curves are anchored at 0% at launch (year 0). Smartphone ownership before 2011 is undermeasured because Pew began the series that year; pre-2011 points are best-effort estimates from contemporary Nielsen and comScore reports.
Source: Pew Research Center smartphone ownership surveys (2011, 2013, 2015, 2018, 2021, 2024); Pew Research Center ChatGPT use surveys (Jul 2023, Feb 2024, Feb-Mar 2025). Compiled by Musa Guide.
ChatGPT alone went from 200 million weekly active users in August 2024 to 900 million in February 2026, a 4.5x increase in 18 months.3 UK ChatGPT visits rose from 368 million in the first eight months of 2024 to 1.8 billion in the same period of 2025, roughly a 5x year-on-year increase, per Ofcom Online Nation 2025.4
Our May 2026 figure of 32% past-12-month museum-AI use among UK/US/DE/FR adult visitors sits within the weekly+ AI-use range used for the weighting anchors (28% to 36%) and below the highest early-2026 national estimates, consistent with museum-AI use tracking general AI fluency.
A second driver is the enterprise rollout: McKinsey measured organization-level generative AI use rising from 33% in late 2023 to 72% by November 2025.5 If daily AI familiarity spreads across age cohorts, museum-AI use is likely to rise with it: in our sample, museum-AI use tracks general AI fluency closely.
Weekly use of generative AI roughly doubled across six Western markets in twelve months. Museum-AI use appears to sit on the same adoption curve, tracking general AI fluency closely in our sample.
Past-12-month AI use rises sharply with general AI fluency (Q1): 67% of our sample's daily AI users have used AI for museum content in the past year, versus 42% of weekly users, 37% of monthly users, and 0% of rarely-or-never users. The same pattern holds for most-recent-visit use, with sharper segment differences: 44% of daily AI users used AI on their most recent visit versus 17% of weekly users and 4% of rarely-or-never users.
The age cut, after reweighting to country-level adult populations, shows under-45s carry most of the past-12-months gap (38% Q8 vs 22% in the 45+ cohort) — but on the most recent visit, the rate is essentially flat across the under-45 / 45+ split (16% vs 17%). The 35-to-44 cell carries a wide confidence interval because its non-heavy AI-use class is small (n=6); read it as suggestive, not as a level estimate.
Figure 1.3
By Q1 frequency: rates within each AI-use class (n=395; Wilson 95% intervals on the raw proportions)—these apply directly to the population assuming class-level rates are stable. By age: reweighted to country-level adult populations using the same heavy/occasional model as the headline; whiskers are 95% bootstrap intervals on the reweighted estimate.
Past year, by AI fluency
Past year, by age
Most recent visit, by AI fluency
Most recent visit, by age
Top two panels: past-12-month museum AI use (Q8 = Yes). Bottom two panels: AI use on the most recent museum visit (Q9 = AI). Reweighted age estimates are sensitive to small heavy/occasional cells in older buckets; treat 55+ rates as directional.
Source: Musa Guide, State of AI in Museums 2026
The age skew and the Q1 frequency skew both move in the same direction: the general adult population has more older respondents and fewer daily AI users than our Prolific panel does. Reweighting is intended to correct for both shifts, though residual bias from unobserved differences remains (see Limitations).
Our country sample sizes outside the UK are small (US 78, DE 56, FR 48), so country-level rates carry wide confidence intervals and should be read as directional. Two patterns are stable enough to flag.
Figure 1.4
Raw rates within our analytic sample, by country. Whiskers show 95% Wilson confidence intervals. Country sample sizes are small outside the UK; read these as directional, not as point estimates.
Source: Musa Guide, State of AI in Museums 2026
France leads on both Q8 (69%) and the most-recent-visit Q9 anchor (52%). The most likely driver is selection: our French respondents are bilingual (the fluent-English screener biases the French sample toward digitally fluent adults more sharply than the UK or US samples). With n=48, that skew alone could account for much of the gap. France also has the highest general weekly+ AI use of the four countries (IFOP-Converteo, March 2026: 41%).
The UK figure (n=213) has the narrowest confidence interval and lands lowest on Q8 (48%), consistent with Ofcom's reading of UK weekly+ AI use as below the other three markets. The raw country order is FR > DE > US > UK, broadly following general AI fluency, but non-UK cells are too small for firm country ranking.
The four countries in our sample contain roughly 436 million adults (UK ~53M, US ~260M, Germany ~70M, France ~53M; aged 18+, 2026 estimates).6
Approximately 35% of adults visit at least one museum or cultural site each year on average across the four markets, putting adult museum visitors at roughly 150 million per year.7
If 32% of those visitors have used an AI assistant for museum content in the past 12 months, the population total is on the order of 48 million adults. Q9 (17% used AI on the most recent visit) puts the most-recent-visit rate at roughly 1 in 6 among adult museum visitors in these markets.
AI use in museums is no longer an edge case. In these four markets, the population-adjusted estimate is roughly 1 in 3 adult museum visitors over the past year and roughly 1 in 6 on the most recent visit, against a general weekly+ AI use range of 28 to 36%. The next sections cover what visitors ask AI, what they do with the answer, and where museums fit in that behaviour.
Within the Q8=Yes cohort (n=212), we asked why they used AI rather than the museum's own information (Q13, multi-select). 69% say they wanted more depth or detail than the museum provided. The second-most-cited reason, at 37%, is that the museum's information did not answer the question the respondent had.
The 13% who said they did not see any museum-provided information likely combines two cases: museums with no content layer, and museums whose content layer the visitor did not notice or could not access.
Figure 2.1
Q13, multi-select. Asked of visitors who used AI in a museum in the past 12 months (Q8 = Yes, n=212).
Source: Musa Guide, State of AI in Museums 2026
Among visitors in our sample who used AI for museum content in the past year (Q8=Yes, n=212), AI assistants were used more often than audio guides on the most recent visit: in this subgroup, 54.7% used AI and 41.0% used an audio guide (free or paid), a gap of roughly 14 percentage points.
The AI figure comes from Q9 ("Asked an AI assistant about something I saw"). The audio guide figure comes from Q5, counting respondents who used an audio guide included in the ticket or paid extra for one.
Audio guide adoption is not a standard public reporting metric and is not uniformly available; vendors and operators tend to keep it private. The published references span a wide range: a 2015 study by the British Museum's own digital team reported roughly 3% rental on the museum's paid permanent-collection audio guide,8 while free-bundled handouts have been reported as high as roughly 80% in at least one operator interview, where the device is handed out at the desk.9 Audio guide adoption depends heavily on what each museum offers and how it is priced, so we compared AI and audio guide use within our own sample rather than against an external benchmark.
Open question we did not measure: whether AI on visitors' phones substitutes for paid audio guide rental. Audio guides are both a content layer and a paid revenue line, so if AI displaces rental, museum digital and revenue budgets would likely be affected. Our survey did not measure this directly.
Among visitors who used AI for museum content in the past year (Q8=Yes subgroup), AI was used more often than audio guides on the most recent visit.
Figure 2.2
Within visitors who used AI for museum content in the past year (Q8 = Yes, n=212). AI use from Q9; audio guide use from Q5, counting included-used and paid-used responses.
Source: Musa Guide, State of AI in Museums 2026
Among AI users in our sample (Q8=Yes, n=212), we asked when they typically use AI for museum questions (Q11, multi-select). Visit-related AI use peaks inside the museum, and visitors also use AI before the visit (to plan, 40%) and after the visit (to recall information about exhibits, whether right after leaving at 43% or days later at 42%). The 66% inside-museum share points to AI being used alongside or in place of museum-provided interpretation (labels, audio guides, docents) during the visit itself.
Figure 3.1
Q11, multi-select. Asked of visitors who used AI in a museum in the past 12 months (Q8 = Yes, n=212).
Source: Musa Guide, State of AI in Museums 2026
Among AI users in our sample (Q8=Yes, n=212), we asked how often they ask AI about museum content (Q10). Just over half (52%) use AI on most or every museum visit.
Figure 3.2
Q10, single-select. Asked of visitors who used AI in a museum in the past 12 months (Q8 = Yes, n=212).
Source: Musa Guide, State of AI in Museums 2026
Q12 was an open-text question: "Please give an example of a question you've asked an AI about something you saw at a museum." 212 visitors answered. Each response was coded into 1 to 3 themes (multi-label).
Half of all questions are about historical context. The next most common is general curiosity at 30%, followed by identity questions ("who/what is this") at 20%.
Figure 3.3
Q12, open-text. 212 responses, each multi-coded into 1 to 3 themes. Categories are not mutually exclusive, so shares add to more than 100%.
Source: Musa Guide, State of AI in Museums 2026
Among AI users in our sample (Q8=Yes, n=212), we asked two questions: who they trust more in general (Q14b) and what they actually do after getting an AI answer (Q14).
61% of AI-using visitors say they trust the museum more than AI (Q14b: 32% "much more" plus 30% "somewhat more"; components are rounded independently from underlying sub-percent values, so the combined share is 61% rather than 62%). 27% verified AI's answer against another source or the museum (Q14: 14% checked another source online, 13% checked the museum's wall text or audio guide).
The gap is 34 percentage points (bootstrap 95% CI: +26 to +43). Of the 130 AI-using visitors who said they trust the museum more, 65% (85 of 130) did not check what AI told them.
61% of AI-using visitors say they trust the museum more than AI. Only 27% actually checked AI's answer.
Figure 4.1
Within visitors who used AI in a museum in the past 12 months (Q8 = Yes, n=212). Q14b combines 'much more' and 'somewhat more' museum-trust responses. Q14 combines all three verify-against-source responses.
Gap: 34 points (95% CI +26 to +43)
Source: Musa Guide, State of AI in Museums 2026
Among AI users in our sample, we asked what they noticed comparing AI's answer to the museum (Q15, Q8=Yes only). When the museum had comparable information, AI agreed 94% of the time (79 of 84 cases where the visitor reported either agreement or contradiction). For 45% of AI use, no museum comparison was possible: 11% where the museum had no information at all, and 34% where AI went beyond what the museum offered. The five reported contradictions are 2.4% of all AI users and 6.0% of the 84 overlap cases. The small base means this is suggestive, not precise.
These are visitor-reported comparisons, so contradictions visitors did not notice are not counted. Specialist queries are particularly exposed: museum collection data is unevenly digitised and often not openly accessible online, and a substantial share of scholarly publications sit behind paywalls,10 so AI does not always draw on the same sources curators can cite. In a 2025 experiment at the Mariners' Museum, a curator testing ChatGPT, Grok, and Perplexity on visual analysis of an unsigned 18th-century maritime painting found that all three made specialist errors (misidentifying flags, ship counts, and the engagement's location), though through guided dialogue they helped surface a plausible artist attribution she had not considered.11 Her conclusion was that AI is a useful brainstorming aid for curators, not a replacement for curatorial expertise.
External benchmarks bracket the picture, though they measure different tasks than visitor-reported agreement. On Vectara's HHEM summarisation leaderboard, the best model's hallucination rate fell from 3.0% in late 2023 (GPT-4) to under 1% by early 2025; Vectara has since moved to a harder dataset where the current floor sits at 1.8%.12 On open-ended factual QA with web access (the default for consumer ChatGPT), GPT-5 hallucinates on 9.6% of questions on OpenAI SimpleQA.13 In specialist contexts the picture is worse: Stanford HAI's AI Index 2026 reports hallucination rates ranging from 22% to 94% across 26 frontier models on a new accuracy benchmark, with much higher rates on domain-specific tasks (e.g., medical case-summarisation up to 64%).14 These benchmarks measure different things (grounded summarisation, open-domain factuality, specialist-domain factuality) and are not directly comparable to visitor-reported Q15 agreement.
Figure 4.2
Q15, single-select. Asked of visitors who used AI in a museum in the past 12 months (Q8 = Yes, n=212).
Source: Musa Guide, State of AI in Museums 2026
The issue is not that visitors always receive visibly wrong answers. The issue is that most AI use happens outside a reliable museum-controlled comparison layer: 45% of reported AI uses had no comparable museum content (11% the museum had none at all, 34% AI went beyond what the museum offered), and even where comparison was possible, only a small subset of visitors actively cross-checked (see Trust without verification).
Among all adult museum visitors in our sample (n=395, not just AI users), 35% said they had wanted to ask a question about something they saw at a museum in the past year but had no easy way to do it (Q17). 54% said no, and 11% were not sure.
Among adults who visited a museum in the past year, 35% recall at least one moment when they wanted to ask a question and had no easy way to do it. This is a lower bound on unanswered curiosity; visitors who didn't try to ask aren't counted.
Figure 5.1
Q17, single-select. Full analytic sample (n=395).
Source: Musa Guide, State of AI in Museums 2026
Q18 asked the full sample (n=395) whether they had been confused about directions or navigation inside a museum in the past 12 months. 6% of visitors report being confused about navigation often, and another 44% occasionally; half of visitors report some level of navigation confusion in a 12-month window.
Figure 5.2
Q18, single-select. Full analytic sample (n=395).
Source: Musa Guide, State of AI in Museums 2026
Q19 asked the full sample (n=395) whether they had felt they did not get the most out of a visit because they did not know about events, discounts, or things they could have done. 10% of visitors report often feeling they missed events, discounts, or activities they could have taken part in; another 44% report this occasionally.
Figure 5.3
Q19, single-select. Full analytic sample (n=395).
Source: Musa Guide, State of AI in Museums 2026
Q20 to Q23 used the Van Westendorp Price Sensitivity Meter to ask at what price an AI audio guide on top of the museum ticket would feel too cheap, a bargain, expensive, or too expensive. After excluding incoherent orderings (n=359 of 395), the PSM intersection points are:
| PSM point | Price | Meaning |
|---|---|---|
| Point of Marginal Cheapness | £2.95 | Below this, "too cheap" share exceeds "expensive" share |
| Optimal Price Point | £4.55 | "Too cheap" share equals "too expensive" share |
| Indifference Price Point | £4.99 | "Cheap" share equals "expensive" share |
| Point of Marginal Expensiveness | £5.45 | Above this, "too expensive" share exceeds "cheap" share |
Stated price tolerance clusters at £3 to £5.50, with the Van Westendorp optimum at £4.55.
Figure 6.1
Cumulative shares for an AI audio guide on top of the museum ticket. Q20 to Q23, restricted to respondents with a coherent ordering (n=359 of 395).
The four marked points are the analytic intersections of the cumulative curves. Q21 ('too cheap') likely read as 'cheapest acceptable price' for many respondents, so the Point of Marginal Cheapness at £2.95 is a soft lower bound; the Optimal Price Point of £4.55 and Point of Marginal Expensiveness of £5.45 are firmer.
Source: Musa Guide, State of AI in Museums 2026
Methodology note: the "too cheap" question (Q21, at what price would the guide feel so cheap that you'd worry about its quality) returned a median of £1 (p25 £0, p75 £2). Most respondents appear to have interpreted Q21 as "the cheapest acceptable price." The PMC of £2.95 is therefore the softer estimate. The PME of £5.45 and the optimal of £4.55 rest on Q20, Q22 and Q23, which were not affected by that wording, and are firm. The lower bound is a soft floor.
Disclosure: Musa Guide, the publisher of this report, builds software in the museum-AI category discussed below. Full disclosure in Disclosure.
The opportunity for museums is not to replace museum authority with AI. It is to put museum authority back into the interface visitors are already using.
The findings above point to three options museums can take in response.
Accept the default. Visitors will use general AI assistants in the gallery whether the museum acts or not. The AI will draw on whatever public sources it can reach. For 45% of reported AI uses, visitors said there was no comparable museum content or that AI went beyond what the museum offered, so the AI's answer may become the visitor's takeaway with no museum reference. This option also means the museum has no visibility into what visitors are asking AI about: which exhibits prompt curiosity, which questions go unanswered by current signage, which pieces drive the strongest engagement. It requires no museum action and gives the museum no influence over the AI layer.
Make museum content available to AI. Open collection data, digitised wall text, and scholarship that is currently paywalled or unindexed change what general AI surfaces when a visitor asks about a piece in the museum (see When comparison was possible, AI usually agreed). This can happen through the open web (open licensing of collection data and academic publications) or through direct content licensing deals with AI providers. Either route lets the museum keep its voice in the answer without owning the AI layer.
Provide a museum-controlled AI layer. A museum-hosted AI guide that grounds answers in the museum's own catalog gives the museum control over the sources AI uses. It can answer questions that current interpretation does not cover, route visitors through the building, and surface events or discounts that visitors otherwise miss. It can also show museums, in aggregate, where visitor questions cluster and where current signage leaves gaps. The trade-off is cost and operational complexity: the museum must build, buy, or partner for the layer and maintain the underlying content.
Within the third option, museum-controlled AI can take several forms for the visitor:
Visitor-side AI is becoming a new interface between visitor and museum content, and that interface is still scaling: ChatGPT alone went from 200M weekly active users in August 2024 to 900M in February 2026 (see How fast AI is spreading), and museum-AI use moves with general AI fluency.
The data do not prove that general AI is replacing audio guides, nor that visitors will automatically pay for museum-controlled AI when free AI tools are already on their phones. They do show that visitor behaviour has moved ahead of institutional provision. The next question is whether museums let general AI absorb more of the visitor relationship, or whether they make their own content, voice, and navigation available through the interface visitors are already adopting.
Open questions for the next survey wave:
This appendix documents the survey instrument, fielding, exclusions, weighting, and analysis steps. The full pre-registration is on OSF: osf.io. Aggregated counts and analysis code are available on request (see Statistical methods).
Respondents were recruited via Prolific Standard Sample. Screeners required residence in the United Kingdom, United States, Germany, or France; age 18 or older; fluent English; and at least one museum or cultural site visit in the past 12 months. Compensation was £0.50 per completer. Median completion time was approximately 3.5 minutes.
A pilot of 30 completers ran earlier on the same instrument with a different reward. Pilot data is excluded from the analytic sample.
The main study was fielded on May 5, 2026. The survey was hosted on Typeform (form id iQMapsPo). Responses were exported via the Typeform API and merged with the Prolific demographic CSV export on participant id.
| Step | n |
|---|---|
| Prolific approvals on main study, May 5, 2026 | 401 |
| Excluded: self-disqualified on screener ("I haven't visited any") | -2 |
Eligible respondents (Typeform thanks_complete ending) | 399 |
| Pre-registered exclusion: completion time under 90 seconds | -4 |
| Pre-registered exclusion: Q12 under 10 characters on the Q8 = Yes path | -0 |
| Pre-registered exclusion: incoherent PSM ordering (see Statistical methods) | applied only to PSM analysis |
| Final analytic n | 395 |
Sub-90-second completion was treated as evidence of non-careful response. The Q12 length rule applied only to AI-using respondents who were asked the open-text follow-up; no responses fell below the threshold.
Final country composition after exclusions:
| Country | n | Share |
|---|---|---|
| United Kingdom | 213 | 53.9% |
| United States | 78 | 19.7% |
| Germany | 56 | 14.2% |
| France | 48 | 12.2% |
| Total | 395 | 100% |
Reweighting was added post-hoc and is therefore not pre-registered. Both raw and reweighted figures are reported throughout the article.
The procedure has three steps:
Country-level weekly+ AI-use anchors:
| Country | Weekly+ | Source |
|---|---|---|
| United States | 36% | Reuters Institute, Generative AI and News Report 2025 |
| United Kingdom | 28% | Reuters Institute, Generative AI and News Report 2025 (UK n = 2,041, June 2025) |
| Germany | 34% | Bitkom KI-Studie 2026 (n = 1,003, telephone CATI, KW 8 to 11 / 2026) |
| France | 34% | Reuters Institute, Generative AI and News Report 2025 |
The procedure brings the sample's AI-fluency profile (86% weekly+ AI users) closer to the four countries' general adult profile (28% to 36% weekly+ in 2025 country data). The reweighting code, country source tables, and per-country / per-age cross-tabs are available on request and tracked in the repository.
Our reweighted figures should therefore be read as conservative. A 2026 set of country anchors would have been substantially higher than the mid-2025 anchors used here. Independent country tracking shows France's weekly+ AI use rose roughly 20% relative between mid-2025 and March 2026 alone (IFOP-Converteo France series: 35% in January 2026, 41% by March 2026), and likely further by our May 2026 fielding. Comparable drift almost certainly applies to the US and UK, whose anchors are also from mid-2025; only the German anchor (Bitkom, early 2026) is fresh. The reweighted 32% past-12-months and 17% most-recent-visit headlines are therefore most likely underestimates of population AI use in our four markets at the time of the survey. Other sources of bias listed in Limitations may add residual noise in either direction, but they are smaller than the anchor-staleness drift.
Analysis was done in Python 3.x with pandas, numpy, and scipy. Responses were exported via the Typeform API and merged with the Prolific demographic CSV on participant id. Raw response data is held privately for participant protection; aggregated counts and the analysis code can be shared on request.
This section lists the weaknesses we are aware of in the data and the analysis. Most are about our own sample and design, so external citations are sparse. Readers should treat the figures in the body of the report with these caveats in mind.
Online panel skew. Prolific recruits skew younger, more educated, and more digitally and AI fluent than the general adult museum-going population. Our raw sample reports 86% weekly-or-more AI use, against 28 to 36% in the four target countries (2025 country data). We reweighted using a two-class heavy/occasional reweighting on country weekly+ AI-use anchors to bring the AI-fluency profile closer to population. Reweighting is a model and corrects for what is observed; residual error from unobserved differences (income, education, urban vs rural, museum-visit frequency by sub-segment) is unmeasured. Both raw and reweighted figures appear throughout for transparency.
Country sample sizes. UK n=213 is reliable for sub-group splits. US n=78, Germany n=56, France n=48 carry wide 95% confidence intervals at the country level. Country-level rates are directional, not point estimates. The DE / US ordering in our raw sample is within sampling noise.
Q14b coverage gap (deviation from pre-registered plan). The trust-attitude question Q14b was supposed to be asked of all 395 respondents. Survey logic skipped it for non-AI-users (Q8 not equal to Yes), so the 61% "trust the museum more" figure in Trust without verification is within-AI-users only, not population-level. The trust gap (61% trust the museum more vs 27% verify) is a within-AI-users comparison. A population-level reading of trust attitudes is not available from this wave.
Q21 wording (Van Westendorp "too cheap"). The Q21 prompt likely read as "the cheapest acceptable price" rather than "the price below which you would doubt quality." The Q21 median of £1 fits that reading. The Point of Marginal Cheapness of £2.95 is a soft lower bound. The Optimal Price Point of £4.55 and the Point of Marginal Expensiveness of £5.45 rest on Q20, Q22 and Q23, which were not affected, and are firmer.
No general-population audio guide benchmark. Audio guide rental is not a standard public reporting metric, vendors keep operator data private, and published references (see the AI is overtaking the audio guide) span 3% to 80% depending entirely on how the museum offers the device. We compared AI use vs audio guide use within our Q8=Yes subgroup only and did not extrapolate to general museum visitors.
No data on cannibalization. Q9 measured AI use on the most recent visit; Q5 measured whether respondents used a free or paid audio guide on that visit. The survey did not ask whether AI use displaced an audio guide rental that would otherwise have happened on the same visit. The implication for audio guide rental revenue at the population level is therefore not measured by this study.
Self-report throughout. Q14, Q14b, Q15, and Q17 to Q19 are self-report. We cannot rule out over-reporting in Q14 (recall asymmetry on self-reported "good" behavior is documented in survey methodology) or under-reporting of AI trust in Q14b (saying you trust the museum may be more socially acceptable than saying you trust the chatbot). Both directions are plausible. We have no behavioral anchor (for example, AI session counts in the gallery) to discipline the self-report.
Twelve-month recall and survey-demand effect. Q8 asks about behavior in the past 12 months. Long recall and topical framing may push the Q8=Yes share upward relative to a more neutrally framed instrument (telescoping and survey-demand effects are documented in the survey-methods literature).
Fluent English screener. The survey required fluent English. This biases the German and French samples toward bilingual, more digitally fluent respondents more sharply than the UK and US samples. It is one plausible reason France leads on Q8 and a reason the country-level rates for DE and FR may sit above the level a native-language survey would return.
Reweighting was post-hoc. The two-class heavy/occasional reweighting scheme was added after data collection and was not in the OSF pre-registration. Both raw and reweighted figures are reported. Readers who prefer the raw sample figures (53.7% Q8=Yes, 31.4% Q9=AI) should use those; readers who want a population-adjusted estimate should use the reweighted figures (32% and 17%).
Hallucination floor in the AI-vs-museum comparison. Q15 records five reported contradictions: 2.4% of all AI users in the sample and 6.0% of the 84 cases where the visitor reported either agreement or contradiction against museum information. The rate is therefore a lower bound on the actual error rate visitors accept on beyond-museum questions, not a precise estimate. The external benchmark figures cited next to it (Vectara HHEM, OpenAI SimpleQA, Stanford HAI AI Index 2026) measure different tasks than visitor-reported agreement and are not directly comparable.
Anchor freshness. The reweighting uses country-level weekly+ AI-use anchors fielded between mid-2025 and early 2026, while our survey was fielded in May 2026. Because general AI adoption continued to rise over that period, the anchors likely understate AI fluency at the time of our survey. This means the population-adjusted estimates are likely conservative with respect to AI adoption growth. Other biases described above may still affect the final level in either direction, but the anchor staleness itself points downward, not upward.
Population scale numbers. The "roughly 48 million adults" figure (see Population scale) multiplies the reweighted 32% rate by an estimated adult-museum-visitor base of 150 million across the four markets. The 35% annual museum-visit rate is averaged across publicly available cultural-participation surveys with different definitions of "museum visit." The figure is best read as an order-of-magnitude estimate (tens of millions). The specific number depends on the 35% visit-rate assumption, which varies by survey definition.
This report is published by Musa Guide, which builds software for museum audio guides and AI-assisted visitor experiences. The survey was designed and analysed by Musa Guide; raw response data is held privately for participant protection.
External sources cited in the report, numbered in order of first citation. Each entry is tagged by role: [Reweighting source] sources are used to derive the population weights applied to our sample; [Hallucination benchmark] sources underpin the accuracy claims about AI agreement with museum information; [Audio guide] sources cover audio guide adoption context; [Industry growth] and [Trend context] sources are atmospheric, not load-bearing for headline numbers.
[Reweighting source] Simon, F. M., Nielsen, R. K., & Fletcher, R. (2025). Generative AI and News Report 2025: How people think about AI's role in journalism and society. Reuters Institute for the Study of Journalism, University of Oxford. Used for the US, UK, and France weekly+ AI-use anchors in Reweighting. reutersinstitute.politics.ox.ac.uk
[Trend context] Pew Research Center. (2026, March 12). Key findings about how Americans view artificial intelligence. pewresearch.org
[Trend context] Pew Research Center. (2025, June 25). 34% of U.S. adults have used ChatGPT, about double the share in 2023. pewresearch.org
[Reweighting source] Bitkom e. V. (2026, April). Künstliche Intelligenz in Deutschland: Studienbericht 2026. bitkom.org
[Trend context] Bitkom e. V. (2026). Ein Drittel nutzt KI mindestens einmal pro Woche, von Code bis Küche [Press release]. bitkom.org
[Trend context] Converteo & Ifop. (2026, March). Étude IA agentique: les chiffres clés de l'IA en 2026. Converteo. converteo.com
[Trend context] Ofcom. (2025, December 10). Online Nation 2025. ofcom.org.uk
[Trend context] Ofcom. (2025, December 10). From apps to AI search: how the UK goes online in 2025. ofcom.org.uk
[Trend context] Ofcom. (2026, April 2). Adults' Media Use and Attitudes 2026. ofcom.org.uk
[Industry growth] Singla, A., Sukharevsky, A., Yee, L., Chui, M., & Hall, B. (2025, March). The state of AI: How organizations are rewiring to capture value. McKinsey & Company. mckinsey.com
[Hallucination benchmark] Stanford Institute for Human-Centered AI. (2026). The 2026 AI Index Report: Responsible AI. Stanford University. hai.stanford.edu
[Trend context] OpenAI. (2026, February 27). ChatGPT reaches 900 million weekly active users [Announcement, as reported by TechCrunch]. techcrunch.com
[Trend context] OpenAI. (2024, August 29). ChatGPT now has more than 200 million weekly active users [Announcement, as reported by Axios]. axios.com
[Hallucination benchmark] OpenAI. (2025, August 13). GPT-5 system card. openai.com
[Hallucination benchmark] Vectara. (n.d.). Hughes Hallucination Evaluation Model (HHEM) Leaderboard. GitHub. github.com
[Hallucination benchmark] Vectara. (n.d.). Hallucination Evaluation Leaderboard [Hugging Face Space]. huggingface.co
[Hallucination benchmark] The Mariners' Museum and Park. (2025, September). Art, history and AI: Should museum curators and researchers use AI? marinersmuseum.org
[Audio guide] Mannion, S., Sabiescu, A., & Robinson, W. (2015). An audio state of mind: Understanding behaviour around audio guides and visitor media. In MW2015: Museums and the Web 2015. mw2015.museumsandtheweb.com
[Audio guide] Soichot, O. (2010). Suivez l'(audio)guide ! [Interview with Alain Eisenstein, CEO of OPHRYS Systèmes]. La Lettre de l'OCIM, 132, 48-53. journals.openedition.org
[Hallucination benchmark] Schonfeld, R. C., Dayan, M., & Ruediger, D. (2024, October 15). Tracking the licensing of scholarly content to LLMs. The Scholarly Kitchen. scholarlykitchen.sspnet.org
[Trend context] Edison Research. (2026, March). The Infinite Dial 2026. edisonresearch.com
[Trend context] Pew Research Center. (2025, September 17). AI in Americans' Lives: Awareness, Experiences, and Attitudes (source for the 62% "at least several times a week" figure cited in How fast AI is spreading). pewresearch.org
These sources informed the background research but are not directly cited in the body of the report.
Newman, N., Fletcher, R., Robertson, C. T., Ross Arguedas, A., & Nielsen, R. K. (2025). Digital News Report 2025. Reuters Institute for the Study of Journalism, University of Oxford. reutersinstitute.politics.ox.ac.uk
Stanford Institute for Human-Centered AI. (2025). The 2025 AI Index Report. Stanford University. hai.stanford.edu
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2025). Hallucination-free? Assessing the reliability of leading AI legal research tools. Journal of Empirical Legal Studies. onlinelibrary.wiley.com
Microsoft & LinkedIn. (2024, May 8). 2024 Work Trend Index Annual Report: AI at work is here. Now comes the hard part. microsoft.com
Stack Overflow. (2025). 2025 Stack Overflow Developer Survey: AI. survey.stackoverflow.co
CREDOC. (2026, February). Baromètre du numérique 2026: Rapport. Centre de recherche pour l'étude et l'observation des conditions de vie. credoc.fr
Eurostat. (2025). Digital economy and society statistics: Households and individuals. European Commission. ec.europa.eu
Mann, L., & Tung, G. (2015). A new look at an old friend: Re-evaluating the Met's audio-guide service. In MW2015: Museums and the Web 2015. mw2015.museumsandtheweb.com
Reweighting uses country-level shares of weekly+ AI use as the anchor for each country (Reuters Institute Generative AI and News Report 2025 for US, UK, FR; Bitkom KI-Studie 2026 for Germany), applied to our sample's Q8 = Yes and Q9 = AI rates within frequent-AI-user and occasional-AI-user classes. See the reweighting section. ↩
Simon, F. M., Nielsen, R. K., & Fletcher, R. (2025). Generative AI and News Report 2025. Reuters Institute for the Study of Journalism, University of Oxford (see References, ref 1). ↩
OpenAI announcement, February 2026. Global figure including non-target-country and non-adult users; reported here for trend triangulation, not as a population-share number. ↩
Ofcom, Online Nation 2025, December 2025 (see References, ref 7). ↩
Singla, A., Sukharevsky, A., Yee, L., Chui, M., & Hall, B. (2025, March). The state of AI: How organizations are rewiring to capture value. McKinsey & Company. mckinsey.com ↩
Adult-population estimates compiled from Eurostat (DE, FR), the UK Office for National Statistics, and the US Census Bureau; rounded to the nearest million for an order-of-magnitude estimate. ec.europa.eu ons.gov.uk census.gov ↩
Averaged across publicly available cultural-participation surveys with different definitions of "museum visit": DCMS Taking Part / Participation Survey for the UK, NEA Survey of Public Participation in the Arts (SPPA) for the US, the German Federal Statistical Office's cultural-participation series, and the French Ministry of Culture's Pratiques culturelles survey. Definitions vary by survey; the 35% figure is best read as an order-of-magnitude estimate. ↩
Mannion, Sabiescu, & Robinson, "An Audio State of Mind: Understanding Behaviour around Audio Guides and Visitor Media", Museums and the Web 2015 (see References, ref 18). ↩
Soichot (2010), interview with Alain Eisenstein of Ophrys Systèmes (audio guide operator), La Lettre de l'OCIM 132 (see References, ref 19). ↩
Scholarly Kitchen, "Licensing Scholarly Content to LLMs", October 2024. ↩
Mariners' Museum, "Should Curators Use AI?", 2025. ↩
Vectara HHEM hallucination leaderboard. ↩
OpenAI GPT-5 system card, SimpleQA evaluation. ↩
Stanford HAI, AI Index 2026, Responsible AI chapter. ↩