betaFeedback

WorksComputer Science

Measuring the Inaccessible: Can AI Evaluate Accessibility in Heritage Tourism?

Abstract

Accessibility assessment underpins the right to participate in heritage tourism, yet evaluation of archaeological sites remains artisanal: fieldwork does not scale, is often carried out without accessibility training, and is reported in forms that reach neither the sector nor disabled visitors. AI appears to offer a shortcut. We argue that assessment is measurement and inference under sparse, expert-mediated evidence: AI transforms how that evidence is captured and communicated, whereas evaluation, which turns evidence into defensible judgements, demands properties current generative systems lack. In a five-park field study in Spain and Italy, an AI desk audit scored only 66% of dimension-site combinations, could not observe embodied and organisational dimensions, and mirrored sites' online self-presentation. Simple Bayesian models instead quantify the confidence rankings deserve and calibrate desk evidence against fieldwork. We propose a separation of roles between AI and interpretable statistical evaluation as the basis for a reliable accessibility atlas.

Introduction

Cultural heritage, from archaeological sites and museums to intangible assets, shapes collective memory, identity, and knowledge, and access to it is internationally recognised as a right: the United Nations [31] affirms participation in cultural life on an equal basis, the Council of Europe [7] recognises everyone’s right to engage with cultural heritage, the European Union [10] sets binding accessibility requirements across the European Union, and UNESCO (1972) and ICOMOS [16] frame inclusive interpretation as integral to responsible heritage management. The population concerned is far from marginal: an estimated 1.3 billion people, 16% of the world’s population, live with a significant disability [32], and on a whole-of-life view access requirements touch an estimated one in three people through impairment, ageing, or travelling with others [8, 3]. Yet a persistent gap separates these principles from lived experience. Policy and practice have traditionally concentrated on museums and urban institutions, whereas archaeological parks present physical, sensory, cognitive, digital, cultural, and economic barriers of their own [23], so that for many visitors they remain effectively closed by design rather than intent. Inclusive archaeology requires more than physical access: because engagement with heritage is embodied, sensory, and cognitive [1], accessibility cannot be confined to the visual dimension, and management depends on understanding how the barriers interact to shape the visit.

The deeper obstacle is that the problem cannot yet be observed at scale. Rigorous assessment remains artisanal, relying on expert fieldwork that is time-consuming, costly, and difficult to replicate, and routine practice compounds the difficulty: assessments are frequently carried out by heritage professionals without specific training in accessibility, and their results are communicated in forms that serve neither the wider sector nor the disabled visitors whose decisions depend on them [25]. The field study underpinning this paper [1] evaluated five parks across Spain and Italy over several years, while managers, policymakers, and disabled travellers increasingly require comparable information on a far larger scale. That challenge has emerged alongside the rapid uptake of AI in tourism [13, 4, 28, 14], where web-scale text, imagery, and review data appear to promise accessibility assessment without fieldwork. The critical questions are what kind of evidence AI-produced accessibility information represents, how reliable it is, and whether it can be compared meaningfully with expert field-based assessment.

This paper argues that accessibility assessment is, at root, probabilistic measurement and inference under sparse, heterogeneous, expert-mediated evidence. AI has transformative potential for the capture and communication of accessibility information. The evaluative step, which turns imperfect evidence into comparable, defensible judgements, demands interpretability, calibrated uncertainty, and accountability to the legal frameworks that underpin the right to cultural participation, properties current generative systems cannot reliably provide on their own. We therefore propose an explicit separation of roles between AI-based capture and communication and statistical evaluation, and provide pilot evidence for it. The contribution is threefold. Conceptually, we develop the distinction between communicating accessibility and evaluating it, and set out the epistemic limits of automated evaluation (§ 2). Empirically, we report what we believe to be the first direct comparison of an AI desk audit with an expert fieldwork audit of the same heritage destinations. Methodologically, we show how simple Bayesian hierarchical models turn audit scores into judgements that carry their own uncertainty and calibrate desk evidence against fieldwork (§ 3), before proposing an interpretable framework and discussing its implications (§ 4 and § 5).

Conceptual framework: accessibility assessment as measurement under uncertainty

Accessibility as a multidimensional, latent construct

Accessibility is not a single, observable property of a heritage site but a multidimensional latent construct inferred from indirect evidence, and it is embodied and interpretive as much as infrastructural. Visitors encounter it through movement, orientation, sensory perception, fatigue, and interpretation, so the same pathway is experienced differently according to age, physical or cognitive profile, sensory abilities, weather, or group composition. Assessment therefore concerns the situated experience of navigating and understanding a site, not only the presence of accessible features: a visitor may physically reach a site while remaining excluded from its interpretive content.

The field study on which this paper draws [1] operationalises it through seven dimensions: physical, sensory, and cognitive accessibility; inclusive programming; staff training; technology and digital provision; and evaluation and monitoring. Each is evidenced by heterogeneous indicators: binary infrastructure checklists, Likert-scale visitor responses, field observation, and documentary analysis. These seven dimensions are the instrument used throughout.

Indicator-based assessment of heritage destinations has a respectable tradition in tourism research [19, 22], and the most direct precedent for archaeological parks is the Heritage Site Accessibility Tool of Salvà Cantarellas [27]: twenty-nine indicators in two weighted areas, applied through fieldwork to eight Mediterranean World Heritage sites, including one studied here. Its value lies in the author’s explicit acknowledgement of its limits: subjective weights, a summary percentage that “can be seen as questionable”, a single person’s judgement, and an access chain only as strong as its weakest link. These are the well-documented pathologies of composite indicators [26, 12]: subjective weighting, information loss under aggregation, and false precision. This paper asks what assessment should look like once those limitations are built into the measurement process, at a time when AI makes accessibility scores cheap to generate at scale.

Three features of the construct govern the approach taken here. First, accessibility is latent and assessment is assessor-mediated: a published score is a potentially noisy reading produced by a particular assessor from particular evidence, not a direct measurement of the site. Second, the dimensions are interdependent: terrain constraints affect sensory experience, orientation, cognitive accessibility, and the interpretation feasible on site. Third, the evidence is sparse and heterogeneous: few sites, many indicators, missing observations, and often a single assessor per site. This regime determines which analytical tools are appropriate, and it is where naive automation is most misleading. The challenge is not primarily technological but epistemological: which aspects of accessibility can be represented as data, how heterogeneous evidence should be combined, and which dimensions remain dependent on situated human judgement.

Communication versus evaluation: two roles of AI

The contributions AI can make to accessible heritage tourism divide into two roles that are routinely conflated and make fundamentally different epistemic demands: communicating accessibility and evaluating it.

The communication role is where the potential of AI is already substantial, and where current practice is visibly deficient: the accessibility information fieldwork produces is typically reported in technical or administrative formats that never reach the visitors whose decisions depend on it. Information has long been the critical resource of the accessible tourism market [20], and its technological provision has evolved steadily [24]. Large language and multimodal models now generate easy-read text, multilingual audio description, and adaptive wayfinding support at very low marginal cost, and domain-tuned models can adapt these outputs to specific audiences [5, 29]; AI here operates as an accessibility intervention in its own right. Closely related is the capture of evidence: multimodal models can structure the scattered, multilingual documentation on which assessment depends, computer vision can extract preliminary cues about terrain and barriers from imagery, and visitor reviews can reveal signals of lived experience that fieldwork rarely captures at scale [4, 28]. By lowering the marginal cost of capture and communication, these technologies make a broad, multi-site atlas conceivable at all, although digital channels can exclude older or less connected visitors, so they should complement rather than replace non-digital provision [13].

Evaluation is a different act: turning evidence into a judgement that scores, compares, prioritises, or certifies sites in ways that affect decisions and should remain open to scrutiny. A communicated description can be questioned by the reader; an evaluative score travels into rankings, funding decisions, and compliance assessments, often while the provenance and uncertainty of the evidence remain invisible. The distinction matters especially in heritage accessibility, where evaluation concerns a latent construct inferred from sparse, expert-mediated evidence: an AI system can efficiently identify and organise relevant information without establishing the validity of the judgement derived from it. The evaluative role demands interpretability, calibrated uncertainty, and accountability, which current generative systems cannot guarantee as autonomous evaluators.

The position advanced here is therefore in favour of appropriately placed AI: transformative in communication and capture, out of place at the evaluative core, where it should be overseen by transparent statistical methods and accountable human judgement. This is consistent with work on hybrid and augmented intelligence in tourism [29], and with the argument of Leonelli [18] that where opacity is institutionally and legally untenable, generative systems should be mediated by transparent probabilistic structures rather than explained after the fact. AI is accordingly one component of a hybrid evidence architecture, expanding capture and communication, while interpretable inference provides the evaluative layer.

Data sparsity, embodiment, and measurement validity

The data regime of heritage accessibility (tens of sites, dozens of indicators, missing entries, and frequently one assessor per site) is where data-hungry, black-box approaches are weakest and explicit statistical models, which state their assumptions and can incorporate expert knowledge, are strongest. Heritage accessibility differs further from that of contemporary built environments because interventions are constrained by conservation and authenticity: archaeological parks occupy uneven terrain, contain fragile remains, and sit under protection regimes that limit physical modification, so assessment must weigh conservation, authenticity, and inclusion together, recognising that some barriers are structural features of the resource rather than failures of management [1].

The question of who, or what, does the scoring then runs deeper than data quality. An automated system scoring a site from its website measures the digital representation of the site, not the site itself; desk-based and fieldwork-based scores are different forms of evidence that should not be treated as interchangeable without explicit calibration. Field observation across the five parks repeatedly revealed discrepancies between documented and experienced accessibility: official materials did not consistently capture terrain, gradients, wayfinding difficulties, sensory demands, or the cumulative effects of fatigue [1]. Accessibility is in part embodied and experiential, realised in the encounter between a particular visitor and a particular place [3]: documentation records that a provision exists, not whether it works for a visitor negotiating uneven terrain or managing fatigue. Multisensory engagement is part of what makes heritage visits meaningful [21], and much of it is invisible from a desk. AI-assisted documentary assessment should therefore not be assumed to reproduce fieldwork more cheaply, since it may be a different measurement process with its own sources of error. Section 3 makes the gap measurable.

False accessibility claims and the cost of unreliable evidence

The failure mode with the most direct human cost is the confident, wrong accessibility claim. Generative systems produce fluent assertions not necessarily anchored to verified evidence; accuracy is among the most consistently flagged risks of their deployment in tourism [13, 4], and inaccurate generative output measurably distorts travel decisions [6, 17]. The costs of error are sharply asymmetric. A site wrongly described as inaccessible loses a visitor; a visitor who travels on the strength of an inaccurate claim about a step-free route, an adapted toilet, or current opening arrangements pays in money, fatigue, and dignity. These consequences fall hardest on disabled travellers, who plan in advance and depend on accurate information to judge whether a visit is feasible at all [8, 20, 27].

Salvà Cantarellas [27] observes that even the blanket label “accessible” misleads when detached from the specific conditions under which accessibility is experienced; a system that synthesises such labels at scale may amplify existing weaknesses in accessibility communication rather than resolve them. What matters is not only whether an individual statement is correct, but whether users can rely on it and understand its evidential basis.

Emerging legal practice reflects the same risk. In British Columbia Civil Resolution Tribunal (2024) a tribunal held an airline liable for negligent misrepresentation after its chatbot gave a customer inaccurate fare information, finding that the company could not disclaim responsibility for what its automated system asserted. The case concerned fares rather than accessibility, but the principle carries: using an automated system does not remove responsibility for the claims users are encouraged to rely upon. Reliability is therefore a prerequisite for any assessment meant to inform decisions, comparisons, or policy, and where AI outputs become evaluative claims their evidential basis, uncertainty, and provenance must remain visible.

Interpretability, accountability, and the limits of aggregation

When a score informs funding allocation, management priorities, or a compliance assessment under the European Union (2019) or United Nations (2006), “the model returned 2 out of 5” is not an adequate basis for a decision. Such decisions require a traceable account of how the score was derived, the evidence, the assumptions, and the uncertainty, which opaque scoring cannot supply. Tourism research has flagged the legal and compliance implications of integrating generative systems with heritage information as an open question [28]; this paper treats the evaluative layer as central rather than as a technical afterthought.

A single opaque number fails accountability in a second way: by aggregating away the structure a decision needs. On the visitor side, needs are heterogeneous: measures that improve access for wheelchair users may offer little to visitors with sensory impairments, autism spectrum conditions, cognitive disabilities, or age-related limitations, so a single score says little about for whom a site is accessible and where barriers remain. On the site side, a mean across dimensions obscures the mechanisms managers must act on: terrain constraints propagate into orientation, sensory experience, fatigue, and cognitive accessibility, whereas digital or interpretive gaps are comparatively easy to address, so two parks can share an average while facing different management challenges. Tourism’s own instruments show it. Reviewing the market-appeal and robusticity matrix of du Cros (2001), McKercher and Ho (2006) note that composite scores cluster near the midpoint as sub-dimensions wash out, and that equal weighting denies a decisive flaw the weight it deserves. Salvà Cantarellas (2023) makes the same point in the language of the access chain: one broken link can void the visit for particular audiences, however good the average.

The objective is not to eliminate summary scores but to place them within an interpretable structure that shows which dimensions drive an assessment, what evidence supports each judgement, where information is missing, and how confident the comparison is, preserving the value of a summary while avoiding the false precision of a single unexplained number.

Case study: fieldwork versus AI desk audits of five archaeological parks

The claims above are testable. We use the five archaeological parks assessed in situ between 2022 and 2026 by Armenio (2026): Cueva Pintada (Gran Canaria, Spain), the Valle dei Templi (Agrigento, Italy), Cerro de la Cabeza (Valdepeñas, Spain), Tiermes (Soria, Spain), and Solunto (Sicily, Italy). Together they span the realistic range of the sector, from a purpose-built museum-park to minimally staffed open-air sites, and so allow a direct test of whether online accessibility information corresponds to accessibility as experienced in the field.

The fieldwork benchmark

Table 1 (upper row of each pair) reports the field assessment of each park on the seven dimensions, on a 1 to 5 scale [1]. The benchmark combines binary indicator grids and structured observation under a common protocol at every park, complemented by visitor questionnaires and documentary analysis where available, with each score anchored to explicit descriptors rather than to an overall impression. The same assessor, trained specifically in accessibility assessment, applied the same instrument across all five parks, so between-park comparisons are internally consistent, and the benchmark is deliberately stronger than the routine practice described in § 1. Its own uncertainty is unquantified, however: with one assessor, assessor effects cannot be separated from the site and dimension components, and the noise term of the model in § 3.4 captures within-assessor inconsistency only.

ParkPhys.Sens.Cogn.Progr.StaffTech.Eval.
Cueva PintadaF4434343
Cueva PintadaD544434–
Valle dei TempliF4423341
Valle dei TempliD4423–43
Cerro de la CabezaF2323222
Cerro de la CabezaD2–33–4–
TiermesF3223221
TiermesD2–32–12
SoluntoF3222221
SoluntoD2––––3–

Table 1. Field assessment versus AI desk audit of the five parks across the seven dimensions (1 to 5 scale). Upper rows (F): fieldwork scores of Armenio (2026); lower rows (D): first run of the pilot AI desk audit (6 July 2026), scoring only where retrieved online evidence supported a judgement (– = insufficient evidence).

Two observations matter for what follows. First, no park dominates: Cueva Pintada has the highest overall profile but is not strongest in every dimension, and the weakest dimension differs across parks, which is the structure a single average obscures. Second, evaluation and monitoring is a systematic weakness, at or near the bottom for four of the five parks: provision is more often implemented than evaluated.

The pilot desk audit: design and findings

We therefore ran a pilot desk audit in which an AI system assessed each park on the same seven dimensions, from publicly available online information only, under a fixed protocol: score a dimension only when retrieved evidence supports a judgement, otherwise record it as unassessable, and log the evidence behind every score. The first run, in Table 1, was performed on 6 July 2026, in a single pass, by Claude Fable 5 (Anthropic) with extended reasoning and integrated web search; S1 gives the verbatim prompt and S2 the evidence log, and two replication runs follow in § 3.3. The pilot is illustrative evidence of the measurement problem, not a benchmark of any particular system.

Three findings stand out. First, coverage rather than accuracy was the binding constraint. The model could justify a score for 23 of 35 dimension-park combinations (66%), and the missingness is clearly non-random: physical and technological provision could be scored for every park, whereas staff training was unassessable for four parks and sensory provision and evaluation for three each. The desk saw the dimensions most visible through websites and far less of the embodied and organisational dimensions that § 2.1 identified as constitutive. Coverage also tracked digital prominence rather than accessibility: the two flagship destinations were almost fully assessable, whereas Solunto, the least documented park, yielded evidence for two of seven dimensions. An automated atlas built this way would inherit, and could amplify, that visibility gap.

Second, where the desk and the field disagreed, the disagreement was systematic. Across the 23 scoreable combinations the mean absolute deviation from the fieldwork was 0.61 points, and the direction of several discrepancies is informative. The desk overscored where official self-description or announced provision dominated the record: it gave Cueva Pintada the maximum for physical access on the strength of the site’s own “fully accessible” materials where the field recorded 4, credited the Valle dei Templi with active evaluation (3 against a field score of 1) on the basis of publicised participatory initiatives, and rated the technology of Cerro de la Cabeza 4 against a field score of 2, reflecting announced virtual-reality provision that the fieldwork found less effective in practice. Conversely, it could not score provision that was present but poorly documented, as with the same park’s sensory resources, observed in the field yet absent from the web. On this evidence, the desk audit is a different instrument from the field audit, measuring a different construct, principally a site’s digital self-presentation, so its limitations arise from a difference in construct as well as from prediction error.

Third, the instruments may disagree simply because they observe at different times: the desk captures what was online on the audit date, fieldwork the site as visited, so any atlas combining them must carry time stamps as well as provenance. None of this makes desk assessment useless: where provision is adequately documented, agreement with the field was close, which is why the capture layer of § 2.2 is valuable. The question is whether documentary evidence can be calibrated against fieldwork rather than whether it can replace it, and § 3.5 shows that calibration is a modelling task the data can support.

Qualitative observations from the evidence log

The evidence log (S2) shows why documentary coverage and field-based accessibility diverge. Some information never appeared online: no park documented staff disability-awareness training, the closest evidence being the generic presence of guides, and systematic evaluation left traces only as announcements, publicity for an event or a dated mention of a visitor survey, insufficient to distinguish a monitoring practice from a one-off communication. Sensory provision at the smaller parks was undocumented even where the fieldwork had found it. Other information appeared but could not be adjudicated: for the Valle dei Templi the system retrieved both the official designation of the site as fully accessible and independent reports of steep gradients and scarce accessible parking, a contradiction a human auditor would investigate and a fluent generative system could resolve without making the uncertainty visible. A further pattern was temporal: for two parks the strongest online evidence concerned works announced after the fieldwork period, so the desk partly assessed a site’s anticipated condition. One observation cuts the other way. No run produced fabricated provision, but this is not evidence that generative AI is inherently reliable: the protocol required supporting evidence and permitted leaving a dimension unassessed, so the restraint is attributable to the procedure rather than the model, which is the design lesson § 4 encodes.

Generative systems are stochastic and the web changes, so reproducibility was tested directly: the audit was repeated on 18 September 2026 in ten fresh sessions, one per park for each of two models (Claude Fable 5.1 and Claude Opus 5), under the same instruction and without access to the field scores or the first-run log (S4). Coverage was 23, 23, and 26 of 35; whether a cell was assessable agreed in 74 to 86 percent of cells between pairs of runs, and six cells (staff training at three parks, evaluation at two, sensory at one) were unassessable in every run. Where two runs both scored a cell they agreed within one point in 90 to 100 percent of cases, and the mean absolute deviation from the field was 0.61, 0.65, and 0.69. The three discrepancies above kept their direction in every run, with smaller magnitude in September, while Cueva Pintada’s sensory provision fell to 2 against a field score of 4 because the park’s own accessibility page could not be retrieved that day: run-to-run variation arose as much from what the retrieval layer could reach as from the model.

The limits of the pilot remain: five parks in two countries, one vendor, three single-pass runs on two dates. The findings are evidence for the conceptual and methodological argument, not a general benchmark of AI performance; a fuller evaluation would add other vendors’ models and more runs under a frozen prompt, with coverage, inter-run agreement, deviation from the field, and the rate of unsupported claims as pre-registered metrics.

A Bayesian hierarchical model of the fieldwork scores

The desk audit shows what happens when evidence collection is mistaken for evaluation; the constructive question is what evaluation should look like once assessment is recognised as probabilistic inference from imperfect evidence. Bayesian models suit this problem because they make uncertainty, structure, and prior information explicit objects of inference that others can inspect and contest, rather than by-products of aggregation [11]. We illustrate with the simplest useful instance, applied to the fieldwork data.

Each field score is decomposed additively into an overall level, a site component, and a dimension component, all estimated with full posterior uncertainty; the specification, priors, and computation are given in S3. This is the least structured model sufficient to demonstrate uncertainty propagation; ordinal likelihoods, latent-variable formulations, and explicit dependence between dimensions belong to the validated instrument of Section 5. Two caveats follow. With thirty-five observations the exercise illustrates the reasoning rather than establishing findings about the parks. A site-level summary may also seem to reproduce the aggregation problem of Section 2.5; the choice is deliberate, because league tables publish such one-number summaries, and the point is to show how one can carry explicit uncertainty and dimension-level information.

Figure 1 reports the site-level result. The raw averages produce an apparently precise ordering: 3.57, 3.00, 2.29, 2.14, 2.00. The posterior supports a more cautious reading. Cueva Pintada is credibly ahead of the Valle dei Templi (posterior probability 0.96), and both are clearly separated from the remaining three, but the bottom three parks are statistically indistinguishable: the probability that Cerro de la Cabeza exceeds Tiermes is 0.68, that Tiermes exceeds Solunto 0.68, and the probability that the raw ordering of all three is correct is 0.41, barely better than chance. Treating those decimals as real differences would claim a certainty the evidence does not support. The dimension components carry information of their own: evaluation and monitoring is credibly the weakest dimension, with posterior effect −0.87-0.87 (90% credible interval −1.45-1.45 to −0.32-0.32) and a 0.92 posterior probability of being the lowest of the seven, confirming the exploratory observation above.

Figure 1. Latent site accessibility for the five fieldworked parks: posterior means with 90% credible intervals (filled circles with horizontal bars), against the raw mean of the seven dimensions (open squares). The intervals of the three lowest-scoring parks overlap almost entirely: the ordering implied by their raw averages holds with posterior probability 0.41.

Calibrating desk evidence against fieldwork

§ 2.2 and § 2.3 argued that desk and field audits measure different constructs; the statistical question is whether the relationship between them can be learned. If it can, documentary evidence becomes a calibratable complement rather than a rival, supplying the scalability an atlas requires while fieldwork remains the reference. We therefore extend the model so that both audits are generated by the same underlying site and dimension quantities, while desk scores are allowed a dimension-specific systematic shift and their own noise level (S3), drawn as a graph in Figure 3. Estimating the shift is what calibration means.

Calibration model diagram showing latent site level $a_i$, dimension profile $b_j$, and desk distortion $d_j$ generating observed field score $F_{ij}$ and desk score $D_{ij}$, with field noise $\sigma_F$ and larger desk noise $\sigma_D$

Figure 3. The calibration model of Section 3.5. Circles are latent quantities, squares observed scores, and arrows read “is generated from”: field and desk audits are noisy readings of the same underlying site and dimension quantities, with the desk additionally subject to a dimension-specific distortion and larger noise. The fuller programme adds arrows among dimensions and nodes for assessors and visitor profiles.

The results, again illustrative, are consistent with the qualitative findings. The estimated desk shift is centred above zero where announcement can substitute for evidence of implementation: evaluation and monitoring (posterior mean +0.5, posterior probability 0.89 of a positive shift), cognitive provision (+0.5, probability 0.89), and technology (+0.4), and close to zero for physical provision, programmes, and staff. The desk also remains the noisier instrument after correction, with residual standard deviation 0.77 against 0.56 for the field. Combining the two sources narrows the site-level intervals of Figure 1 by roughly ten percent, a gain that comes from estimating the relationship between the instruments rather than from taking desk scores at face value. One limitation motivates future work: the desk’s blanks are informative, falling on the dimensions and parks least visible online, so a full atlas must model the missingness rather than condition on it. Conditioning assumes a cell is blank at random given site and dimension, which the coverage pattern contradicts, so the calibration estimates should be read as conditional on online visibility.

An interpretable framework for accessibility assessment

The case study motivates the separation of roles summarised in Figure 2. Heterogeneous evidence, from documents and imagery to visitor reviews, is captured with AI assistance and handed to the evaluative core as provenance-tagged, time-stamped material, never as a final judgement. The core is an interpretable probabilistic model, anchored by expert fieldwork and the lived expertise of disabled visitors, that turns evidence into judgements carrying uncertainty and an auditable derivation, which then serve decision support for managers, funders, and regulators, and accessible communication for travellers. AI sits at the edges; accountable inference sits at the core.

Proposed separation of roles for AI in heritage accessibility assessment

Figure 2. The proposed separation of roles for AI in heritage accessibility assessment: AI-assisted capture structures heterogeneous evidence, an interpretable probabilistic core turns it into judgements with uncertainty, AI-assisted layers support decisions and communicate accessibly, and governance carries provenance, calibration, and compliance throughout.

Figure 3 draws that model as a graph that can be read aloud. Circles are quantities nobody observes directly: each park’s underlying accessibility level, each dimension’s profile across the sector, and the desk’s dimension-specific distortion. Squares are the recorded scores. The arrows say that both audits are generated by the same underlying quantities, with the desk score additionally shifted by the distortion and blurred by its own, larger noise. Every assumption is on the page: to contest an assessment is to point at a node or an arrow. The same grammar extends directly, in the spirit of glass-box architectures in which transparent probabilistic structure mediates between model output and institutional decisions [18]: arrows among dimensions would encode how terrain constrains sensory and cognitive provision, and further nodes would represent assessors or visitor profiles, in a tradition already familiar in tourism [15].

Within this core, the answer to the limitations that Salvà Cantarellas [27] acknowledges and the composite-indicator literature formalises [26, 12] is direct: where subjective weights are unavoidable, make them explicit objects of inference; where a scalar conceals information, report the structure; where a point score overstates, report the uncertainty. Faced with the same problem, McKercher and Ho [19] retreat to a deliberately ordinal scheme; ours is the complementary move, keeping a graded assessment but modelling what a flat score discards.

Discussion: from pilot evidence to a living accessibility atlas

What stands between this pilot and the atlas the field needs is a set of connected developments rather than a single missing technology or dataset. Table 2 summarises what the case study showed and what would close each gap. The rows are not independent: a validated instrument makes calibration meaningful, calibration makes AI-assisted capture safe to use, and compliance-ready interpretability lets any of it enter the decisions the European Union [10] and United Nations [31] regulate.

ComponentWhat the pilot showedWhat is needed
Validated instrumentEven expert scores support limited ranking confidence; assessor effects not yet modelledMulti-assessor studies; ordinal and latent-variable formulations; explicit reliability
AI-assisted captureThe desk cannot observe embodied and organisational dimensions, is biased toward announcements, and varies with retrieval accessMulti-vendor, repeated-run evaluation under frozen prompts; human validation; provenance and time stamps
Desk-field calibrationDimension-specific desk distortion is estimable from paired data (§ 3.5)Joint models combining evidence streams; treatment of informative missingness
Compliance-ready interpretabilityOpaque scores cannot answer regulator questionsGlass-box models mapped to EAA and CRPD reporting duties
ParticipationProvision invisible online; lived experience absent from web dataCo-design with disabled visitors across the whole assessment cycle

Table 2. From pilot to atlas: what the case study showed and what is needed.

The first four rows are statistical and infrastructural, and § 3 and § 4 have shown their simplest instances working. Scaling them is also concrete. Under the calibration model the desk need not be paired with fieldwork everywhere: a minimal atlas desk-audits every site, field-audits a subsample stratified by digital prominence, since prominence drives missingness, and estimates the desk distortions from the paired subset. Computation is linear in the number of sites; what changes between five sites and five hundred is that assessor effects become estimable once assessors overlap, missingness must be modelled rather than conditioned on, and replicated desk runs become the bulk of the evidence.

The fifth row deserves more than a row. Accessible tourism is defined by collaborative processes among stakeholders [3], and an instrument that scores the experience of disabled visitors without their involvement perpetuates, in measurement, the exclusion it purports to diagnose. Participation should extend beyond consultation to the whole assessment cycle: indicator selection, field testing, interpretation of results, and evaluation of interventions. This is established practice: Naniopoulos and Tsalis [22] developed their monument-accessibility methodology through a structured process that included people with disabilities directly, and Salvà Cantarellas [27] reaches the same conclusion for audit design. Co-design discharges an ethical obligation and earns a methodological return at once, because lived expertise identifies which indicators matter, which barriers compound, and where documentation diverges from experience.

The programme has, finally, a policy face. AI-derived accessibility information circulating through destination platforms needs governance: provenance and time-stamp standards distinguishing verified fieldwork from automated inference, alignment with the reporting obligations of the European Union [10], and recognition by heritage institutions that publishing an accessibility claim is an act with consequences rather than a marketing gesture.

Hybrid arrangements in which computation assists experts, managers, and disabled visitors are where tourism research on AI is heading [28, 29].

Taken together, these developments point to a living, probabilistic accessibility atlas in which each site’s accessibility is a posterior distribution rather than a static score, updated by Bayesian evidence synthesis as field audits, AI-captured documentation, and visitor reports accumulate, calibrated across instruments and countries, and displayed with its uncertainty rather than despite it. Sparse expert fieldwork would then do what it is uniquely able to do, anchor and calibrate the cheaper evidence streams, while every new observation sharpens rather than replaces the assessment.

Conclusion

Accessibility is not a property of places but a relationship between people, heritage environments, and systems of interpretation, and archaeological sites expose the tensions between conservation, authenticity, and inclusion more sharply than most tourism settings.

AI can document, organise, and communicate accessibility information at unprecedented scale, and our pilot shows it doing useful work there. The same pilot suggests why evaluation cannot be delegated to systems of this kind: across three runs the desk instrument measured digital self-presentation and overlooked the embodied and organisational dimensions that constitute accessibility as lived experience, and even expert fieldwork supports far less ranking confidence than its averages suggest. Assessment should therefore use AI for capture and communication and rely on interpretable inference for evaluation, because inclusion is the domain in which being able to say why a judgement was reached, and how certain it is, forms part of the accessibility being measured.

Supplementary material

Declaration of generative AI use

Generative AI was used in two capacities. First, as the object of study: the pilot desk audit of Section 3 was performed by large language models with web search access under the protocol of Section S1, and their outputs are reported as data. Second, a large language model was used as a writing and editing assistant under the direction of the authors, who designed the study, verified all sources, analyses, and claims, and take full responsibility for the manuscript.

Disclosure statement

The authors report there are no competing interests to declare.

Funding

No funding was obtained for this work.

Data availability statement

The fieldwork assessment data derive from Armenio [1]. The desk-audit protocol, evidence logs, and replication runs are reported in § 9, § 10, and § 12.

Supplementary material

This supplementary material contains the desk-audit protocol (§ 9), the first-run desk-audit evidence log (§ 10), the specification of the statistical models (Section S3), and the two replication runs of the desk audit (§ 12) referred to in the main text.

Desk-audit protocol

The pilot audit reported in § 3 was performed on 6 July 2026 by Claude Fable 5 (Anthropic), accessed through the Claude interface with extended reasoning enabled and integrated web search as the only information source, in a single pass; the system was instructed to justify every score from retrieved evidence alone. For each park the system received the same instruction:

Using only publicly available online information, assess [park name] on each of the following seven dimensions of accessibility: physical; sensory; cognitive; inclusive programmes; staff training; technology and digital; evaluation and monitoring. Score each dimension on a scale from 1 to 5. Score a dimension only if the evidence you retrieve supports a judgement; otherwise record it as not assessable. Note the evidence behind every score.

Scores were transcribed into Table 1 and the supporting evidence into § 10 below; the full transcripts, with the sources retrieved for each score, are available from the authors on request. Two replication runs under the same instruction, on a later date and with a second model, are reported in § 12.

Desk-audit evidence log

Condensed from the audit transcripts; scores in parentheses, – marks dimensions recorded as not assessable for lack of evidence.

Cueva Pintada. Physical (5): self-described fully accessible; lifts, walkways, step-free routes. Sensory (4): audio-description; tactile visits; sign-language visits; subtitled audiovisuals. Cognitive (4): comprehensible-content commitment; autism visits with pictograms. Programmes (4): inclusive educational and family programmes; inclusive-museum seminar. Staff (3): staff along route; interpreters on request; no training documented. Technology (4): immersive audiovisual; online ticketing; digital section. Evaluation (–): no monitoring trace.

Valle dei Templi. Physical (4): official fully-accessible designation, free electric wheelchairs, walkways; independent reports of gradients near ten percent, scarce accessible parking, inaccessible sub-areas. Sensory (4): braille panels; sign-language video-guide; sensory garden; visits for blind and deaf visitors. Cognitive (2): little beyond standard panels; route maps. Programmes (3): dedicated visits; participatory events with disability organisations. Staff (–): nothing documented. Technology (4): sign-language application; QR-coded monuments; online resources. Evaluation (3): publicised participatory assessment initiative.

Cerro de la Cabeza. Physical (2): steep hillside; universal accessibility only announced in a future plan. Sensory (–): no evidence. Cognitive (3): interpretation centre with models and audiovisuals; reconstruction panels. Programmes (3): school programmes, workshops, guided visits. Staff (–): no evidence. Technology (4): on-site virtual-reality headsets; augmented-reality application; online three-dimensional models. Evaluation (–): no evidence.

Tiermes. Physical (2): works tendered in 2026 to remedy routes in poor condition; short seasonal adapted route. Sensory (–): no evidence. Cognitive (3): well-regarded interpretive panel network with reconstructions. Programmes (2): free seasonal guided visits, one adapted. Staff (–): no evidence. Technology (1): no application, virtual tour, or digital provision found. Evaluation (2): dated mention of a continuing visitor survey.

Solunto. Physical (2): hillside setting, visitor advice implying demanding terrain; no site accessibility information. Sensory, cognitive, programmes, staff (–): no evidence. Technology (3): website accessibility widget; announced QR-code and virtual-tour project. Evaluation (–): no evidence.

Statistical models

Field model. Field scores are modelled as Fij=μ+ai+bj+εijFF_{ij} = \mu+ a_i + b_j + \varepsilon_{ij}^{F}, with site components ai∼N(0,τa2)a_i \sim N(0, \tau_a^2), dimension components bj∼N(0,τb2)b_j \sim N(0, \tau_b^2), noise εijF∼N(0,σF2)\varepsilon_{ij}^{F} \sim N(0, \sigma_F^2), an improper flat prior on μ\mu, and inverse-gamma(2, 1) priors on all variance components. Site-level accessibility is θi=μ+ai\theta_i = \mu+ a_i.

Calibration model. Desk scores, where observed, are modelled as Dij=μ+ai+bj+δ0+δj+εijDD_{ij} = \mu+ a_i + b_j + \delta_0 + \delta_j + \varepsilon_{ij}^{D}, sharing μ\mu, aia_i, bjb_j with the field model, with overall desk shift δ0∼N(0,1)\delta_0 \sim N(0, 1), dimension-specific shifts δj∼N(0,τδ2)\delta_j \sim N(0, \tau_\delta^2), desk noise εijD∼N(0,σD2)\varepsilon_{ij}^{D} \sim N(0, \sigma_D^2), and inverse-gamma(2, 1) priors on τδ2\tau_\delta^2 and σD2\sigma_D^2. The reported desk shift per dimension is Δj=δ0+δj\Delta_j = \delta_0 + \delta_j. Desk missingness is conditioned on, not modelled; the blanks are informative and a full treatment would model them.

Computation and summaries. Both models are fitted by conjugate Gibbs sampling (24,000 iterations, 6,000 burn-in, fixed seed); chains mix rapidly and results are stable across seeds. Posterior means and 90% intervals are reported below; the site levels θi\theta_i are those displayed in Figure 1. Dimension components bjb_j: physical +0.50 (−0.05, 1.05); sensory +0.33 (−0.22, 0.89); cognitive −0.36 (−0.91, 0.19); programmes +0.33 (−0.21, 0.88); staff −0.18 (−0.74, 0.36); technology +0.16 (−0.39, 0.70); evaluation −0.87 (−1.45, −0.32). Desk shifts Δj\Delta_j: physical +0.01 (−0.58, 0.58); sensory +0.32 (−0.41, 1.06); cognitive +0.46 (−0.15, 1.08); programmes +0.04 (−0.58, 0.65); staff +0.01 (−0.86, 0.85); technology +0.37 (−0.21, 0.95); evaluation +0.54 (−0.20, 1.30). Noise: σF\sigma_F 0.56 (0.45, 0.69); σD\sigma_D 0.77 (0.59, 0.99). Under the calibration model the 90% intervals for θi\theta_i narrow from an average width of 1.08 to 0.99. The sensory, staff, and evaluation shifts rest on one or two jointly observed parks and should be read accordingly.

Replication runs of the desk audit

Design. The audit of Section S1 was repeated on 18 September 2026 with two models, Claude Fable 5.1 and Claude Opus 5 (Anthropic), each with integrated web search. Each park was audited in a separate, fresh session with no conversation history and no access to the field scores, to Table 1, or to the first-run log of Section S2; ten sessions in total. Each session received the instruction of Section S1 verbatim, preceded by a procedural wrapper: use only publicly available online information retrieved during the session and cite every score by URL; search in English and in the local language (Spanish or Italian) with between six and fifteen searches; draw on the site’s official pages and social channels, public institutional documents, accessibility portals, news, and visitor reviews; exclude academic theses, articles, or preprints reporting accessibility assessments of the park; do not force a score, since “not assessable” is an expected outcome; and note the date of evidence, distinguishing announced works from existing provision. Sessions returned scores, one-line evidence summaries with URLs, the list of queries run, and any conflicts among sources. On the audit date several official pages could not be retrieved by the search tool (the Cueva Pintada accessibility page, the Valle dei Templi park website including its accessibility document and service charter, the Cerro de la Cabeza park website, and the Tiermes museum site), which the sessions recorded as caveats; scores rest on institutional, press, portal, and review sources.

Scores. Table S1 reports, for every park and dimension, the field score and the three desk readings. Table S2 summarises coverage and agreement. Per dimension, staff training was scored in one park in R1, none in R2, and one in R3; evaluation and monitoring in two parks in every run; sensory provision in two, three, and four parks. Six cells were not assessable in any run: staff training at Cerro de la Cabeza, Tiermes, and Solunto; evaluation at Cerro de la Cabeza and Solunto; sensory provision at Solunto. Twenty-four of the 35 cells had the same status in all three runs. The two September runs, which shared a date, agreed with each other more closely than either did with July, consistent with part of the between-run variation being driven by what was retrievable on the day rather than by the model. The first-run sign pattern (desk above field where announcements dominate) reversed on average in September, when the official pages of the two best-documented parks were unreachable and scores fell back on third-party audits, press, and reviews; the three cells singled out in § 3.2 nevertheless stayed above the field in every run (Cueva Pintada physical: field 4, desk 5/4/5; Valle dei Templi evaluation: field 1, desk 3/2/2; Cerro de la Cabeza technology: field 2, desk 4/3/3).

ParkPhys.Sens.Cogn.Progr.StaffTech.Eval.
Cueva Pintada4/5/4/54/4/2/23/4/3/34/4/4/33/3/–/–4/4/3/33/–/2/2
Valle dei Templi4/4/4/44/4/4/42/2/2/13/3/2/33/–/–/24/4/3/31/3/2/2
Cerro de la Cabeza2/2/2/23/–/2/22/3/2/23/3/2/–2/–/–/–2/4/3/32/–/–/–
Tiermes3/2/2/32/–/–/12/3/–/13/2/–/22/–/–/–2/1/1/21/2/–/–
Solunto3/2/1/12/–/–/–2/–/2/22/–/1/22/–/–/–2/3/2/31/–/–/–

Table S1. Field score and desk readings across the three runs, as F / R1 / R2 / R3. F: fieldwork [1]. R1: Claude Fable 5, 6 July 2026 (Table 1). R2: Claude Fable 5.1, 18 September 2026. R3: Claude Opus 5, 18 September 2026. – = recorded as not assessable.

RunCoverageMAD vs fieldMean signed diff.Over / equal / under
R1 (Fable 5, July)23/350.61+0.268 / 11 / 4
R2 (Fable 5.1, Sept.)23/350.65−0.482 / 10 / 11
R3 (Opus 5, Sept.)26/350.69−0.384 / 10 / 12
PairStatus agreementJointly scoredExact / within oneMAD between runs
R1 vs R227/35 (77%)1937% / 95%0.68
R1 vs R326/35 (74%)2035% / 90%0.75
R2 vs R330/35 (86%)2264% / 100%0.36

Table S2. Coverage of each run and pairwise agreement between runs. Status agreement: share of the 35 cells with the same assessable or not-assessable status in both runs. Jointly scored: cells scored in both runs. Exact and within-one: share of jointly scored cells with identical scores, and with scores differing by at most one point. MAD: mean absolute deviation between the two runs’ scores (between runs), or between a run and the field scores (versus field).

Condensed evidence logs, runs 2 and 3. Scores in parentheses; sources are summarised, with full URLs in the session outputs held by the authors. Cueva Pintada. R2: Physical (4): TUR4all audit sheet (undated) and PREDIF recognition (2010) record a step-free chain, lift with braille and audio buttons, adapted toilet, lowered desk; not 5 for lack of a dated independent confirmation. Sensory (2): braille and contrast signage per TUR4all; 2021 press on QR codes and sign-language work “being done”; no audio description, loop, or tactile models found. Cognitive (3): recurring autism-oriented visits with a pictogram document (2024–2025); no easy-read material for all visitors. Programmes (4): three dated autism family workshops, adapted visits listed by TUR4all, 2021 inclusive-museum seminar. Staff (–): seminar participation only; no training documented. Technology (3): virtual visit, 3D collection, museums app, 2022 VR exhibition; no accessibility claims for the digital offer. Evaluation (2): director reports continuous visitor studies with a university partner since 2006; annual visitor statistics; no accessibility-specific monitoring. R3: Physical (5): the same two audits read as a complete step-free chain. Sensory (2): building-level features only. Cognitive (3): pictogram document for autistic visitors, repeated workshops. Programmes (3): adapted visits and autism workshops; no provision evidenced for other groups. Staff (–): island-wide accessibility plan proposes training but does not name the site. Technology (3): 3D and audiovisual displays, multilingual tours, virtual visit, museums app with a Cabildo accessibility statement. Evaluation (2): a 2007–2009 visitor satisfaction study; nothing accessibility-specific or recent. Both runs note that the museum’s own accessibility page was unreachable.

Valle dei Templi. R2: Physical (4): 85% of the site declared barrier-free after 2020 works, free electric wheelchairs and shuttle, accessible toilets; conflicting wheelchair counts (two on the concessionaire’s page, six in the regional release) and a 2025 review calling close access to the temples very difficult. Sensory (4): braille panels and tactile map, tactile Telamon, sign-language videos by QR code, sensory garden (2023), VideoSegnArt app (2024). Cognitive (2): no park-run easy-read or social-story material; the only such provision is in the FAI-run Kolymbethra garden. Programmes (2): episodic events on the International Day of Persons with Disabilities and an autism walk hosted rather than designed by the park. Staff (–): 2019 announcement that sign-language training was initiated, no later confirmation. Technology (3): VideoSegnArt app, QR sign-language panels, audioguide; website conformance unverifiable. Evaluation (2): a 2019 participatory needs assessment with eight disability associations and the director’s own testing; no published audit or survey. R3: Physical (4): as R2, with visitor reports of steps at the Temple of Juno and no buggy cover in the western sector. Sensory (4): as R2. Cognitive (1): no cognitive-specific provision found; a 2019 source names it as still needed. Programmes (3): standing concessionaire offer of multisensory visits plus recurring autism-day events with park patronage. Staff (2): the 2019 sign-language training for custodial staff, reported as under way, with no later evidence.

Technology (3): as R2. Evaluation (2): as R2; the service charter page could not be retrieved.

Cerro de la Cabeza. R2: Physical (2): a 2019 third-party audit of the interpretation centre (17% kerb ramp needing assistance, 4 cm entrance step, no reserved bays, two adapted toilets); the hilltop site has no documented accessible route and is temporarily closed for works whose stated goal is an accessible itinerary. Sensory (2): high-contrast panels and toilet pictograms, low lighting; no braille, tactile, audio, loop, or sign-language provision. Cognitive (2): panels, scale models, audiovisuals, VR and AR reconstructions; nothing easy-read; cognitive accessibility is a future project aim. Programmes (2): generic guided visits, workshops, and school visits; nothing for disabled visitors. Staff (–): no evidence. Technology (3): VR sessions and the regional AR app since 2024, rated 3.2 with signal complaints; no accessibility features documented. Evaluation (–): no surveys, statistics, or monitoring found. R3: Physical (2): as R2. Sensory (2): as R2, plus a regional statement that the XR experiences are adapted for hearing loss. Cognitive (2): as R2, with visitor reports of poor signage. Programmes (–): school and workshop provision only; nothing targeted found. Staff (–): a 2014 tourism-quality award with generic awareness sessions; nothing disability-specific. Technology (3): as R2; municipal tourism site claims an obsolete web-accessibility standard. Evaluation (–): as R2. Both runs note a funded but unexecuted project promising universal accessibility.

Tiermes. R2: Physical (2): regional works completed August 2026 (regravelled paths, restored walkway, extra steps) improve comfort rather than step-free access; guides and reviews state the site is unsuitable for wheelchairs. Sensory (–): no provision mentioned anywhere. Cognitive (–): standard panels and a numbered map only; reviews report broken signs and an unclear route. Programmes (–): generic guided tours; museum education page empty. Staff (–): no evidence. Technology (1): no app, audioguide, QR codes, or virtual tour; only the regional museums portal’s accessibility features. Evaluation (–): regional aggregate statistics only. R3: Physical (3): the same 2026 works, plus a 45-minute reduced-mobility circuit documented in 2010 and 2017 sources while the regional directory lists wheelchair access as limited. Sensory (1): service listings enumerate other facilities and nothing sensory. Cognitive (1): no easy-read or structured-visit material; wayfinding criticised. Programmes (2): free summer guided visits, archaeologist-led tours, astronomy events; free entry for visitors with a disability certificate; nothing designed for disabled audiences. Staff (–): regional training references only, not site-specific. Technology (2): accessible regional portal; an iOS app listing found but not retrievable; no on-site digital provision. Evaluation (–): a 2010 announcement of visitor surveys; nothing since.

Solunto. R2: Physical (1): steep unpaved hillside road, reviews advising against visiting with walking difficulty, no official accessibility information, no shuttle or mobility aid. Sensory (–): the only sensory aid in the park cluster (sign-language smartglasses) is at Himera. Cognitive (2): bilingual boards and a photocopied map; conflicting reviews on signage quality. Programmes (1): general tours, free days, and an invitation-only event; nothing targeted. Staff (–): review praise for individual staff only. Technology (2): the park cluster’s app, virtual tours, and AR are documented for Himera, not Solunto; website accessibility widget only. Evaluation (–): no survey, charter, or audit found. R3: Physical (1): as R2. Sensory (–): as R2. Cognitive (2): thin interpretation plus an illustrated guide for young readers (2025). Programmes (2): active school and public programme including Solunto; free entry for disabled visitors and a companion; nothing co-designed with disabled audiences. Staff (–): as R2. Technology (3): park app, inclusion in a regional digitisation circuit (2022), a necropolis virtual tour reported as forthcoming, website accessibility widget. Evaluation (–): as R2.

References

  1. [1]Armenio, S. (2026). Arqueología Inclusiva: los retos de la accesibilidad en los parques arqueológicos. PhD thesis, Universidad Autónoma de Madrid.
  2. [3]Buhalis, D. and Darcy, S., editors (2011). Accessible Tourism: Concepts and Issues. Channel View Publications, Bristol.
  3. [4]Carvalho, I. and Ivanov, S. (2024). ChatGPT for tourism: Applications, benefits and risks. Tourism Review, 79(2):290–303.DOI
  4. [5]Chemnad, K. and Othman, A. (2024). Digital accessibility in the era of artificial intelligence: Bibliometric analysis and systematic review. Frontiers in Artificial Intelligence, 7:1349668.DOI
  5. [6]Christensen, J., Hansen, J. M., and Wilson, P. (2025). Understanding the role and impact of generative artificial intelligence (AI) hallucination within consumers’ tourism decision-making processes. Current Issues in Tourism, 28(4):545–560.DOI
  6. [7]Council of Europe (2005). Council of Europe framework convention on the value of cultural heritage for society (Faro Convention). CETS No. 199, Faro.
  7. [8]Darcy, S. and Dickson, T. J. (2009). A whole-of-life approach to tourism: The case for accessible tourism experiences. Journal of Hospitality and Tourism Management, 16(1):32–44.DOI
  8. [10]European Union (2019). Directive (eu) 2019/882 on the accessibility requirements for products and services (european accessibility act).
  9. [11]Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. (1995). Bayesian Data Analysis. Chapman and Hall/CRC, London.
  10. [12]Greco, S., Ishizaka, A., Tasiou, M., and Torrisi, G. (2019). On the methodological framework of composite indices: A review of the issues of weighting, aggregation, and robustness. Social Indicators Research, 141(1):61–94.DOI
  11. [13]Gursoy, D., Li, Y., and Song, H. (2023). ChatGPT and the hospitality and tourism industry: An overview of current trends and future research directions. Journal of Hospitality Marketing & Management, 32(5):579–592.DOI
  12. [14]Hall, C. M. and Cooper, C. (2025). Making tourism smart in the age of artificial intelligence. Current Issues in Tourism, 28(1):1–5.DOI
  13. [15]Huang, Y. and Bian, L. (2009). A Bayesian network and analytic hierarchy process based personalized recommendations for tourist attractions over the Internet. Expert Systems with Applications, 36(1):933–943.
  14. [16]ICOMOS (2008). The ICOMOS charter for the interpretation and presentation of cultural heritage sites.
  15. [17]Kim, J. H., Kim, J., Park, J., Kim, C., Jhang, J., and King, B. (2025). When ChatGPT gives incorrect answers: The impact of inaccurate information by generative AI on tourism decision-making. Journal of Travel Research. doi:10.1177/00472875231212996.
  16. [18]Leonelli, M. (2026). Beyond post-hoc explanation: Toward glassbox AI via probabilistic mediation. arXiv:2606.07113.arxiv.org/abs/2606.07113
  17. [19]McKercher, B. and Ho, P. S. Y. (2006). Assessing the tourism potential of smaller cultural and heritage attractions. Journal of Sustainable Tourism, 14(5):473–488.
  18. [20]Michopoulou, E. and Buhalis, D. (2013). Information provision for challenging markets: The case of the accessibility requiring market in the context of tourism. Information & Management, 50(5):229–239.DOI
  19. [21]Moscardo, G. (1996). Mindful visitors: Heritage and tourism. Annals of Tourism Research, 23(2):376–397.
  20. [22]Naniopoulos, A. and Tsalis, P. (2015). A methodology for facing the accessibility of monuments developed and realised in Thessaloniki, Greece. Journal of Tourism Futures, 1(3):240–253.DOI
  21. [23]Nicolini, E. (2021). Prestazioni di accessibilità nei siti archeologici: criteri per la valutazione e miglioramento. In Germanà, M. L. and Prescia, R., editors, L’accessibilità nel patrimonio architettonico. Approcci ed esperienze tra tecnologia e restauro, pages 64–73. Anteferma Edizioni, Conegliano.
  22. [24]Nigg, J. J. and Peters, M. (2022). The evolution of ICTs in accessible tourism: A stakeholder collaboration analysis. Journal of Hospitality and Tourism Management, 52:287–294.DOI
  23. [25]Nigg, J. J., Plaikner, A., Peters, M., and Haid, M. (2025). Leisure constraints towards accessible tourism: Self-characteristics of people with disability along tourism value chains. Current Issues in Tourism, 28(18):3031–3045.DOI
  24. [26]OECD and European Commission Joint Research Centre (2008). Handbook on Constructing Composite Indicators: Methodology and User Guide. OECD Publishing, Paris.DOI
  25. [27]Salvà Cantarellas, M. À. (2023). Archaeological heritage for all: A heritage site accessibility tool (HSAT) for open-air archaeological sites. Archaeologies, 19(3):515–554.
  26. [28]Sigala, M., Ooi, K.-B., Tan, G. W.-H., Aw, E. C.-X., Buhalis, D., Cham, T.-H., et al. (2024). Understanding the impact of ChatGPT on tourism and hospitality: Trends, prospects and research agenda. Journal of Hospitality and Tourism Management, 60:384–390.DOI
  27. [29]Tuomi, A., Tussyadiah, I., and Ascensão, M. P. (2025). Customized language models for tourism management: Implications and future research. Annals of Tourism Research, 110:103863.DOI
  28. [31]United Nations (2006). Convention on the rights of persons with disabilities. United Nations, New York.
  29. [32]World Health Organization (2022). Global report on health equity for persons with disabilities. Technical report, World Health Organization, Geneva.

Paper details

vlnc:W_68vc3vqq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

a large language model

View statement
Looking for related papers…
Contents