Artificial Intelligence and Eco-Museums: Climatic Challenges and Ethical Perspectives within European Cultural Sustainability

1. Introduction

In recent years, the European cultural sector has moved beyond the so-called phase of digital transformation (DT): platforms, databases, and online workflows are now routine, although an equivalent level of digital literacy has not matched the widespread, almost compulsive use of such tools [1, 2]. Against this backdrop, artificial intelligence (AI) has become the new disruptive factor, capable of profoundly affecting both managerial processes and cultural mediation [3, 4]. For example, one of the most widely discussed recent cases concerns the Albanian state’s decision to “employ” an AI avatar, Diella, as a minister to combat corruption. This move rides the latest trends while cleverly framing them as a celebration of the country’s cultural heritage, as signaled by the traditional attire in which the Albanian Prime Minister Rama has introduced the avatar [5].

The author argues that, as communities approach the DT era, the spread of AI across the GLAM field (galleries, libraries, archives, museums) and the cultural industries raises a twin set of issues: on the one hand, opportunities in terms of efficiency and broader access – particularly around community inclusion; on the other, environmental impacts and ethical dilemmas that require a rethinking of practices and new customs [6]. In particular, growing compute demand for training and using generative systems entails energy and water consumption, as well as hardware-related externalities, which cannot be ignored by institutions that profess a commitment to sustainability [7].

Eco-museums are among the institutions potentially most exposed to this tension. By history and mission, they are not containers of objects but museums-as-process, rooted in place, where communities and landscapes are integral to heritage [8, 9]. In less than 50 years, the eco-museum structure has become an operational pact between inhabitants, local institutions, and places [10], applying both to tangible assets and, above all, to intangible cultural heritage (ICH) [11]. The strong centrality of ICH requires that any digital innovation, AI included, be functional to the living transmission of knowledge, practices, and memories, without reducing them to simulacra or decontextualized repertories.

This nexus between eco-museums and ICH has clear foundations in international frameworks: the 2003 UNESCO Convention defines the intangible as a set of practices, expressions, and knowledge transmitted from generation to generation – living and dynamic [11]; the Faro Convention underscores the link between heritage, cultural rights, and democratic participation [12]. Operationally, this entails community ownership of content and safeguarding mechanisms that foster continuity and transformation, not crystallization. Bringing AI into this ecosystem, therefore, means asking not only what can be automated, but how and within what limits to do so, preserving the voice, context, and responsibility of the communities that generate living heritage.

At the European policy-institutional level, significant steps toward the legitimization and regulation of AI in cultural institutions include the Green Deal (climate neutrality by 2050) [13], the Ethics Guidelines for Trustworthy AI (transparency, human oversight, justice, social and environmental well-being) [14], and the regulatory evolution culminating in the AI Act (risk-based classification with corresponding obligations) [15]. These are not abstract frameworks but, potentially, concrete criteria for eco-museums: for selecting suppliers (disclosure on energy and water), for designing workflows (minimizing generation, prioritizing retrieval and re-use), for labeling generated content, and for local review (human-in-the-loop) before publication.

Within this framework, the author highlights the problematic over-reliance on ChatGPT as the representative case, given that it is one of the best-known and “user-friendly” AIs – a tool fully capable of underpinning user-facing services [16]. The point is not to imply the “automation of the museum,” but to understand whether and how a large language model can assist with low-criticality tasks (short mediation texts, micro-educational trails, and linguistic simplification) while maintaining human control and minimizing computational demand: concise prompts, length limits, retrieval-over-generation, and systematic re-use of approved outputs. The focus is not solely on text quality but also on computational parsimony as an ethical dimension, in line with the community nature of the eco-museum. The balance between benefits (accessibility, clarity, inclusion) and costs/risks (consumption, bias, opacity) becomes the testing ground for responsible adoption.

The resulting position for eco-museums and the cultural sector more broadly is as follows: AI is acceptable when it serves the mission and when it measures its own effects. More specifically, it serves the mission if it broadens access without flattening modes of expression, if it supports the transmission of ICH without supplanting it, and if it renders content more readable for diverse audiences; it measures its effects when it makes computational demand visible (how many requests, how many tokens, how many re-generations), when it makes editorial pipelines explicit (what was generated and from which local sources one started), and when it institutionalizes community review. This human- and planet-centered stance realigns AI with responsible innovation in a European key and with the civic-educational profile of eco-museums, avoiding both uncritical enthusiasm and prejudicial rejection.

Advertisement

2. Origins, definitions, and principles

The eco-museum was conceived as a process-based museum and a community pact. This participatory, place-based institution sustains silent knowledge over time, preventing the loss of traditions and cultural heritage. As an alternative to the object-centered model, the earliest French formulations of the 1970s [17] shifted attention to an inseparable triad – territory, community, and heritage – linking memory, environment, and local development. Early cases are emblematic: at Le Creusot (1971), in a mining and metallurgical district, the eco-museum became a tool for culturally repurposing industrial heritage [18]; at Fresnes, marked by the detention of Jews during the Second World War, the eco-museum approach helped restore civic meaning to a site of memory through collective storytelling and regeneration [19]. From the outset, these trajectories showed that the “collection” extends beyond material objects to the practices, traces, and testimonies that bind places to the lives passing through them. The concept finds a clear synthesis in the Altkirch Declaration (2004): the eco-museum is a territory-based museum grounded in participation, integrating cultural, natural, and intangible heritage to foster social cohesion and sustainable development [20]. This is not a slogan but a codification of principles already in use, reaffirming that territory is not a backdrop but a constitutive element of the museum.

Accordingly, the eco-museum acts as a community instrument for caring for, interpreting, and enhancing cultural and environmental heritage for local well-being – effectively embodying the community. Emphasis falls on purpose (care and transmission) and on cultural ownership (the community decides what counts and how it is told). The result is a focal point that goes beyond exhibition as an end in itself: the eco-museum activates processes and enables collective choices and practices.

Operationally, eco-museums employ tools such as landscape charters, participatory mapping, oral-history archives, intergenerational workshops, thematic itineraries, and place-based storytelling. The goal is not accumulation but to keep alive the relationships among places, knowledge, and people, maintaining continuity while allowing transformation. This aligns with the UNESCO 2003 Convention on Intangible Cultural Heritage: intangible heritage is living, context-dependent, and requires safeguarding that preserves use and meaning for the communities that generate it.

The territory itself is the primary “object-document,” living evidence of collective identity. It must be understood beyond administrative boundaries as a relational space in which safeguarding means transmission and reactivation rather than crystallization. Here, the eco-museum takes shape as a pact among community, institutions, and landscape: the community recognizes, selects, and narrates; institutions facilitate and steward; the landscape provides the weave within which practices and memories make sense [21].

This approach entails an ethics of participation: the community does not hand over its heritage to an institution – it governs it [20]. Hence, the shift of focus from digital products to living contexts, ensuring that digitization supports rather than disembodies practices [22].

Advertisement

3. Organizational forms and functions

In the shift from an “alternative territorial institution” to a dynamic, shared space, the eco-museum has consolidated a flexible organizational morphology: not a single building–container, but a form that interweaves places, practices, and people within a pact among community, institutions, and landscape. In this configuration, the eco-museum serves as a community instrument for caring for, interpreting, and enhancing its cultural and environmental heritage, in the service of local well-being and sustainable development [23].

Operationally, recurring tools give concrete shape to the “distributed museum”: participatory mapping to identify meaningful places and practices; oral-history archives and intergenerational workshops to activate transmission and learning; place-based narratives (trails, heritage walks, events) that circulate stories and know-how anew. These devices are not designed to “fix” traditions but to accompany their vitality within contexts of practice, avoiding drifts toward the decontextualized musealization of ICH [24].

The educational function is structural: eco-museums operate as community laboratories and partners to schools, supporting forms of active, place-rooted learning (local history, ecological knowledge, craftsmanship). At the same time, they experiment with accessible mediation for diverse audiences, fostering inclusion and cultural citizenship. Where appropriate, they promote slow, identity-rooted tourism attentive to local ecologies and proximity economies, without slipping into a spectacle.

From an organizational standpoint, an eco-museum works when decisions are shared, roles are clear, and community participation is not reduced to symbolic consultation. This approach enables the integration of scarce resources (volunteerism, distributed skills) with institutional support and project funding, while safeguarding the cultural ownership of those who inhabit the places. Absent this architecture, the risks grow: standardized language and the outsourcing of narrative choices.

Typical functions can be summarized along four interlinked axes:

1. Documentation and care of memory – collecting testimonies, recordings, and materials that, once returned to the community, support processes of recognition and awareness; not a closed archive but a common good in evolution.

2. Mediation and narration of place – producing situated narratives (briefs, micro-itineraries, events, performative formats) that connect places, practices, and lived landscapes, while maintaining the link between the tangible and the intangible.

3. Education and inclusion – co-designing with schools, associations, and local groups pathways that foster experiential learning, intergenerational exchange, and participation by new or marginalized publics.

4. Sustainable local development – supporting economic and social practices consistent with cultural diversity and proximity economies (craft know-how, agro-pastoral value chains, festivals), while avoiding commodifications that distort community meanings.

These functions call for hybrid competencies: curatorial, educational, project design, and documentary.

At the system level, building networks of eco-museums – national and translocal – proves crucial for sharing criteria, tools, and minimum standards (participatory processes, consent models, formats for returning materials to the community) and for strengthening organizational capacity in fragile contexts. Networks also help navigate heterogeneous regulatory frameworks, fostering operational convergence and the exchange of good practices.

In sum, the eco-museum is a social and cultural infrastructure that succeeds when it braids light, distributed organizational forms with functions of documentation, mediation, education, and local development – all rooted in participation and community ownership. This framework – process museum, shared governance, participatory tools, and networks – forms the basis on which, in the sections that follow, we will assess whether and how AI tools can be aligned with the mission, without slipping into the substitution of practice or the homogenization of languages.

Advertisement

4. Essential international experiences

International experiences confirm that an eco-museum works when cultural ownership remains with communities, practices are enacted (not merely documented), and mediation keeps its bond with territory and landscape. The four cases cited here are not a part of an exhaustive inventory but directional indicators showing how the model adapts to diverse contexts without losing its founding traits.

• Lai Chi Wo (Sinosphere) [25, 26]. In a historic village marked by terraced agriculture and tight interrelations among settlement, forest, and water cycles, the eco-museum approach has operated as a device for socio-ecological regeneration. Care for the agricultural landscape, the revival of traditional techniques, and the intergenerational transmission of knowledge have been brought back to the center not as folklore, but as living practices with contemporary utility (environmental education, local resilience). The community’s role has remained decisive in selecting themes, narrating places, and defining educational activities, with governance that unites local actors and technical expertise without dispossessing the former of their voice.

• Écomusée du Fier Monde (Canada) [27]. Born within an urban fabric shaped by layered histories of labor, migration, and social movements, this eco-museum has made co-curation with associations and neighborhoods its distinctive hallmark. The memory of work and struggles has been presented not as a repertoire of objects but as oral history, neighborhood routes, co-produced temporary exhibitions, and educational programs developed with schools and civic networks. Its public function is explicit: to give space to plural narratives, make social transformations legible, and foster participation and inclusion through accessible language without sacrificing rigor and accountability to the communities involved.

• Eco-museum of Nordfjord (Norway) [28, 29]. In a coastal and rural context organized as dispersed nodes and thematic centers, the eco-museum connects fishing, seamanship, craftsmanship, rural life, and oral storytelling. Its polycentric structure avoids the “headquarters-museum” risk and allows ties to be rewoven among trades, landscapes, and everyday practices. Mediation favors walks, demonstrations, workshops, and small situated displays, with communities playing a strong role in selecting knowledge and testimonies. The integration of the material (tools, boats, architectures) and the intangible (technical know-how, local vocabularies, life stories) is treated as a continuum: what matters is the relationship to places and to the people who inhabit them.

• Ecomuseu de Itaipu (Brazil) [30]. In a territory transformed by major infrastructure, the eco-museum approach has focused on reweaving memory and environmental education, holding together the effects of territorial transformations with the persistence of local knowledge and identities. Its programming puts ecosystems and communities into dialog, with workshops, visits, and narrative devices that show how innovation can be oriented toward environmental justice and active citizenship, avoiding the spectacle of places and keeping the focus on rights, responsibilities, and long-term horizons.

From these four cases emerge operational constants that are also useful for the chapter’s discussion of AI: (a) the centrality of the community in deciding what is told and how; (b) distributed, polycentric forms of organization capable of holding together places, practices, and memories; (c) accessible mediation that privileges clear language, situated narratives, and experiential learning; (d) an integrated material/intangible vision in which objects make sense because they are connected to knowledge, relationships, and landscapes; and (e) an explicit relationship to sustainability and rights that prevents drifts toward appropriation or the digital musealization of practices. It is on this warp and weft that the scope and conditions of any potential use of AI will later be assessed: if it does not reinforce these elements, the innovation is off-target; if it supports them with proportionality and transparency, it can become a tool that serves the eco-museum’s mission.

Advertisement

5. Proportion over volume: AI that serves cultural and environmental ends

As discussed above, eco-museums live in symbiosis with nature. That bond sets a conditio sine qua non: digital tools, however helpful, should never erode the eco-museum’s mission. Hence, the following are theoretical reflections on a conscious, nonintrusive use of AI. Against this background, with digitization now embedded in heritage workflows, the critical question is not whether to use technologies but how to align them with cultural and environmental ends [1, 3]. The lexicon of Industry 4.0 – automation, interoperability, and data integration – has streamlined cataloging, management, and access [31]. At the same time, the use of a shared digital language, from digitization to digitalization, has increased the computational loads that accompany next-generation AI. The next step, for a deliberate use of this tool, is the 5.0 inflection: technology no longer as a neutral multiplier of output, but as a means proportionate to human and planetary objectives, evaluated by impacts rather than by performance alone [32].

This inflection entails a shift in the decision-making scale. The unit of account is not the algorithm in the abstract but the editorial deliverable: every use of AI must be situated in places and practices, traceable in its sources and local review, reversible without loss of meaning, and measurable in the marginal computational load it introduces. The European framework on climate, trustworthy AI, and risk-based regulation provides an external constraint. Still, the real threshold lies inside editorial workflows: here, one decides whether AI adds public value or merely volume.

In the eco-museum context, this means combining tools inherited from 4.0 – helpful in describing and connecting knowledge and lived landscapes – with 5.0 safeguards: community ownership, protection of ICH, transparency in choices, and local control over steps. Responsible innovation is not a checklist of principles but an operating method: anticipate effects (environmental and cultural) by setting ex-ante limits on use; include territorial actors in the editorial chain; monitor language and representations to avoid bias; and correct course in response to documented community feedback.

Within this profile, the question “Why AI in eco-museums?” must be posed rigorously. The issue is not whether to adopt AI, but when and under what conditions. From this follows the proportionality criterion that governs the whole chapter: rely on AI only where it yields a recognizable cultural gain, and in the minimum necessary amount. In practical terms, this points toward brief, contextual texts; retrieval-over-generation where validated local materials exist; systematic reuse of approved outputs; local human review before publication; clear labeling of AI use; targeted prompts and lengths within budget; at most one re-generation; and traceability to local snippets and a community glossary. If even one of these prerequisites is missing, it is preferable not to generate AI content; that said, there are concrete reasons why some uses of AI can align with the eco-museum mission:

• AI can serve as a linguistic and editorial aid to improve clarity and accessibility, useful in resource-constrained settings and with heterogeneous audiences (local visitors, schools, families, older adults). Generating short briefs, plain-language materials, and local translations can support educational work without supplanting community voices, provided the algorithm’s ancillary role is clear and local human review occurs before publication [33].

• The centrality of ICH – practices, know-how, rituals, performing arts, and ecological knowledge – requires tools that help contextualize and interpret without “fixing” traditions. The 2003 UNESCO Convention defines ICH as living heritage, transmitted across generations and continually renewed. AI can help draft comprehensible explanations or thematic companion itineraries to accompany practices – not replace them – safeguarding communities’ cultural ownership of content and representations, with editorial control and transparency about tool use [34].

• In territories with infrastructural gaps and limited professional staffing, AI – if used sparingly – can facilitate distributed curation: drafting panels, micro-educational itineraries, and summaries for intergenerational pathways. Yet, there are risks of “disembodying” the digital experience: poorly mediated technology can turn ICH into a simulacrum detached from context, encouraging passive consumption and spectacle. Adoption must, therefore, remain instrumental and supervised, with review practices, citation of local sources, and accountability of knowledge-bearers within the editorial process.

• Finally, the environmental question arises from AI’s cumulative effects on energy, data, and computation. Responsible use in an eco-museum adopts a demand-before-supply principle: generate only what is needed, in short texts, limiting re-generations, and favoring retrieval and the reuse of validated community materials. This keeps the local “computational cost” auditable and consistent with Europe’s sustainability horizon [35].

Advertisement

6. European framework: Sustainability, cultural rights, and trustworthy AI

In the European context, the digital transformation spread gives way to a phase in which innovation is assessed in terms of sustainability and responsibility [13]. In its most recent outlooks, the European Union ties the use of AI to climate and social objectives, on the assumption that rising computational loads affect the cultural sector’s environmental footprint and call for corrective action [13, 32]. The reference framework is the European Green Deal, with climate neutrality on the horizon for 2050 and the idea that the transition is both a challenge and an opportunity for “a better future for all” [13]. In parallel, the Commission has advanced the Ethics Guidelines for Trustworthy AI, which set requirements for human oversight, robustness, privacy, fairness, and social and environmental well-being, steering adoption toward a balance between innovation and responsibility [15, 36].

Alongside this strategic dimension is an evolving regulatory framework for AI. Europe is moving toward a regulation that classifies systems by their risk to health, safety, fundamental rights, the environment, and democracy, with graduated obligations for transparency, risk management, and technical documentation [15]. For cultural institutions, this entails assessing use cases, tracing data flows, and making the use of generative systems in mediation and education transparent, with specific attention to the overall environmental impact of the tools adopted [36].

The climate–culture link is explicit: used without restraint, AI amplifies energy and water consumption and produces externalities along the hardware supply chain; governed responsibly, it can serve educational purposes, access, and inclusion in line with European values. An indiscriminate increase in computational demand (training, storage, inference) would have negative climate repercussions; for this reason, the GLAM sector should define metrics and mitigation practices. In this sense, the EU framework is not ornamental: it invites measurement and reduction of the footprint, tying AI adoption to choices of parsimony (reuse, short texts, retrieval) and to policies of accountability toward visitors and communities.

For eco-museums, the European framework translates into operational consequences:

• Editorial transparency: When an LLM is used in mediation, it is appropriate to inform the public about AI use and to cite the local sources employed in retrieval, while maintaining human-in-the-loop review and terminological coherence with community glossaries [36].

• Minimization of computational demand, consistent with the EU’s environmental approach – produce short, situated texts; limit re-generations; prefer retrieval-over-generation; and reuse approved snippets.

• Governance of cultural data, with informed consent and clear licenses, in line with Faro and the centrality of ICH as living heritage (not “digitally musealized”).

• Measurement: Pragmatism would align with the spirit of European policy: not offloading sustainability to generic pledges, but supporting the institutionalization of monitoring routines and continuous improvement proportionate to eco-museum resources [31].

In this way, the EU framework – Green Deal, trustworthy AI, and risk-based regulation – ceases to be mere “context” and becomes a working method capable of combining innovation, the protection of cultural rights, and footprint reduction.

The hypothesis is that, in this near-post-digital transformation European landscape, AI, and in particular generative language models, will enter heritage workflows, promising accessibility and editorial simplification, but at the cost of increasing processing demands (energy, water, hardware) and ethical–cultural risks (prejudice, rights, opacity).

In eco-museums, where heritage, place, and community coincide, the issue becomes more pressing. Every innovation must support the living transmission of ICH and the community’s ownership of content, avoiding the “digital musealization” of practices and knowledge, and governing the uses of cultural data.

Advertisement

7. Implications for AI in the eco-museum context

In light of the foregoing, the adoption of AI in eco-museums is not a question of if but of conditions: to align with the mission and with European frameworks, especially ecological ones, it must be tied to a principle of proportional use and to accountability mechanisms that hold together living heritage, community, and sustainability [13, 36].

• Centrality of the community and cultural ownership. In eco-museums, the community is not a target audience but the subject of the museum process; consequently, AI may be used only within editorial pipelines that preserve community voice, control, and the right to review or refuse content that concerns them. This entails clear procedures for informed consent, explicit licensing rules, and traceability of local sources used in mediation. This is consistent with the Faro perspective and with the participatory approach the draft identifies as a condition for the future: “reclaiming responsible cultural ownership” is the precondition for any innovation, digital or algorithmic.

• Avoiding “disembodiment” and the digital musealization of ICH [12, 34]. Algorithmic mediation must accompany practices, not replace them. Digital tools risk turning the intangible into a decontextualized simulacrum or spectacle; the same holds for AI: generated texts, translations, or summaries must remain anchored to use contexts, places, and relationships that give heritage its meaning. Misuse – for example, standardized productions or tech “showcases” – encourages passive consumption and weakens the experiential learning eco-museums promote.

• Sustainability: reduce computational demand. The escalation of AI workloads (training, storage, inference) has direct and indirect climate impacts, calling for corrective measures and a culture of responsible use. In eco-museums, this translates into parsimonious choices: generate only what is essential for mediation (plain language, context notes, micro-itineraries), limit re-generations and length, and prioritize reuse of already approved community materials and contextualization over mass production of new texts. The goal is clear: secure gains in access and intelligibility without pushing computational demand beyond what is proportionate to the mission.

• Transparency and human control along the chain. Whenever AI shapes texts and mediation apparatuses, this should be disclosed to the public in plain terms, with local human-in-the-loop review before publication. Transparency is not a box to tick; it makes choices and responsibilities auditable, aligns the eco-museum’s actions with European guidelines on trustworthy AI, and protects trust with visitors.

• Avoiding technological shortcuts and managing divides. It is also essential to avoid showcase-style drift (e.g., low-interactivity 360° tours touted as innovation), which can exacerbate infrastructural divides affecting many inland areas. In such contexts, AI must be designed to fit the territory: restrained applications oriented toward mediation and education, interoperable, with sustainable organizational costs, and without shifting decision-making and control away from communities. Eco-museum networks can reduce inequalities by sharing models, tools, and baseline skills [37, 38].

Advertisement

8. Implications for AI in the eco-museum context

In order to operationalize a responsible and sustainable approach to AI adoption in eco-museums, this section outlines a set of proportionality criteria that should guide any decision to implement generative technologies. These criteria are derived from the ethical, environmental, and cultural imperatives (Table 1) discussed in the previous sections and are aligned with the European framework for trustworthy and sustainable AI (Green Deal, Ethics Guidelines, AI Act).

The application of AI is deemed acceptable only when all of the following conditions are satisfied:

1. Mission alignment: The use of AI must demonstrably support the eco-museum’s mission of safeguarding ICH, enhancing accessibility, and promoting sustainable local development. AI should function as an assistive instrument, not a substitute for embodied practices or community voice.

2. Retrieval over generation:Whenever possible, AI should draw upon validated, community-approved local content through retrieval mechanisms rather than generating new content from scratch. This reduces both cultural risks (such as decontextualization and misrepresentation) and environmental costs.

3. Human-in-the-loop editorial review: All outputs must undergo local human review before public dissemination. Editorial control ensures cultural accuracy, guards against bias, and maintains the legitimacy of the eco-museum’s interpretive authority.

4. Transparent labeling and source attribution: Content generated with AI must be clearly labeled, and any local sources used must be explicitly cited. This transparency is crucial for both editorial accountability and public trust.

5. Compute minimization: The computational footprint should be minimized by enforcing concise prompts, limiting generations (e.g., to a single regeneration if strictly necessary), and avoiding unnecessary model calls. This includes the reuse of validated outputs and the avoidance of content overproduction.

6. Consent and licensing: Any cultural data used in AI workflows must be governed by informed consent protocols and clear licensing agreements that reflect the community’s rights and expectations.

7. Auditability and measurability: Institutions should adopt proxy indicators – such as the number of AI queries, average tokens per output, reuse ratios, and regeneration rates – to assess and document the environmental and editorial impact of AI use over time.

These criteria are not intended as mere recommendations but as thresholds for ethical and sustainable deployment. Where one or more criteria cannot be fulfilled, the use of AI should be avoided. This proportional approach (Figure 1) ensures that innovation remains subordinate to the eco-museum’s core values: cultural continuity, environmental responsibility, and community ownership.

Advertisement

9. Environmental risks

Among the various issues linked to the unreflective use of AI, one of the primary risks concerns the environment. The escalation of computational demand – training, storage, and inference – cumulatively increases electricity consumption and pressure on water resources, with externalities across the entire hardware supply chain. On current projections, global data center electricity demand could more than double by 2030 (≈945 TWh), with AI as the primary driver [39]. The growth of computing loads, especially since the advent of “large” models, makes a nonnegligible contribution to the ICT sector’s climate-altering emissions, with direct and indirect effects that call for corrective measures and an explicit accountability framework in GLAM.

Training generative models is energy-intensive in itself; however, in the day-to-day practice of eco-museums, inference weighs more because it multiplies into countless requests and re-generations. Peer-reviewed accounting for BLOOM-176B shows approximately 25–50 tCO₂e for the final training run (method-dependent) and highlights that, once deployed, inference can become a major share of lifecycle impact [40]. The technological trajectory of the past decade shows a rapid rise in computing power used for AI; recent metering and estimates suggest a typical general-purpose LLM query today is on the order of approximately 0.3 Wh (model- and load-dependent) [41]; operationally translated: even seemingly harmless but repeated uses – re-generating the same draft multiple times, producing texts that are too long, serializing translations – inflate computational demand beyond what is proportionate to the mission.

The practical rule is simple: fewer calls, higher quality. Work with short texts, prefer adaptation from local materials (retrieval-over-generation), limit re-generation to one motivated attempt, and plan for the reuse of approved outputs. In this way, AI remains a useful tool without turning into a multiplier of consumption.

For the same amount of energy, where and how inference is run matters: data centers with fossil-heavy energy mixes and water-intensive cooling systems worsen the impact. Evidence shows nontrivial water use both on-site (cooling) and upstream (power generation); for example, training GPT-3 at a U.S. site was estimated to directly evaporate approximately 700,000 liters of freshwater, and day-to-day use can amount to roughly 10–30 milliliters per prompt (~0.5 liters over 20–50 prompts), depending on the site/technology [42, 43]. This figure should not be “absolutized,” but it is relevant because it shifts attention from emissions alone to water resources, which are particularly critical in many European regions. Lifecycle studies also indicate that advanced liquid cooling (cold-plates/immersion) can reduce GHG by approximately 15–21%, energy by 15–20%, and blue-water by 31–52% compared to air cooling [43].

A single request to a general-purpose LLM consumes nonzero energy, even when answers look “small” (~0.3 Wh typical), and aggregate query volumes are the main driver of operational impact [39, 41]. Even if these are estimates, the operational message is clear: volumes of use and text length are material drivers of impact. For small, distributed institutions like eco-museums, controlling these two factors is the first lever of sustainability.

Beyond compute, the material chain of devices (servers, networks, endpoints) entails mineral extraction, manufacturing, transport, and electronic waste. Semiconductor fabs use high-GWP fluorinated gases (e.g., PFCs, NF₃, N₂O) and large amounts of energy and water; regulators and industry recognize this as a priority for abatement [44, 45]. Impacts are not exhausted by the kWh consumed: expanding installed capacity and shortening life cycles generate environmental degradation and pressure on value chains, especially in the absence of refurbishment and recovery strategies. Even if eco-museums do not run data centers, they participate as users: the demand they generate still feeds the supply chain.

Another risk is opacity: providers often do not report, in a granular fashion, energy mix, water consumption, and per-service/per-region metrics, making comparable assessment difficult for cultural institutions. Where available, prefer providers offering credible disclosures and, when configurable, regions with lower grid-carbon and lower water stress; note that public initiatives (e.g., IEA’s Energy and AI Observatory) are improving transparency [46].

The ease of generating texts and translations introduces a rebound: the easier it is to produce content, the more content gets produced. In GLAM, this can induce “over-production” of labels, briefs, and micro-narratives, multiplying model calls without a corresponding increase in value for visitors.

Uncritical adoption of high-intensity, general-purpose models can be disproportionate for short editorial tasks. The environmental risk here is organizational: choosing tools more “powerful” than necessary creates structural computational waste. The chapter urges calibrating the model to the use case (reducing complexity, retrieval-over-generation, reuse) and limiting lengths and versions to avoid the model’s authorial drift.

Implications for Eco-museums:

1. AI’s environmental footprint is not abstract but the result of everyday editorial choices (how many requests, how long, how many re-generated). Where teams set thresholds and practice systematic reuse, impact decreases without sacrificing clarity and accessibility.

2. Transparency toward the public and community concerns not only the ethics of mediation but also environmental ethics: disclose when and why AI is used, and shows that practices of computational parsimony are in place.

3. Provider selection and compute-region choices (when configurable) affect the marginal impacts on energy and water; in the absence of complete data, it is reasonable to work with user-side proxies and prudential criteria.

4. Guard against rebound with editorial rules and retrospective cycles on actual tool use (how many re-generations were truly needed? Which texts can be cached and republished?).

In sum, the environmental risks of AI for eco-museums stem less from “grand projects” and more from diffuse, unguided use of generative tools: energy and water for inference, hardware-chain externalities, informational opacity, and editorial rebound. Impact can be reduced only if demand is measured (with simple proxies), production is contained (short texts, reuse, retrieval), the tool is proportionate to the task, and transparency is required of providers, in line with the European Green Deal horizon and the notion of trustworthy AI [47].

Advertisement

10. Ethical implications

The ethical and cultural risks tied to the use of AI in eco-museums can be read along four interwoven axes: (a) representation and bias; (b) cultural rights, consent, and licensing; (c) editorial transparency and human control; and (d) dependencies and divides – with a transversal risk of disembodying experience and digitally musealizing ICH when algorithmic mediation separates practices from the contexts that keep them alive [34, 48].

Language models tend to flatten registers and vocabularies: without editorial constraints and community glossaries, generative output can level out local differences (dialects, technical terms, variants) and induce stereotypes or oversimplifications, especially around minority cultures or nonstandardized practices. In the eco-museum sphere, where ICH is living heritage, the damage is twofold, as linguistic specificity is lost and situated transmission is weakened [48, 49]. AI governance must be consistent with communities’ cultural ownership, with procedures for informed consent, explicit licensing, and clear labeling of generated content, together with a right of review or refusal for knowledge-bearers; where these steps are missing, the risks rise for appropriation and improper reuse of recorded intangible materials, contradicting Faro’s participatory approach and the processual nature of ICH.

Transparency requires a human-in-the-loop pipeline and explicit disclosure whenever AI intervenes in mediation, so that authorship and sources remain auditable and editorial responsibility remains intact. Even for low-criticality tasks, data minimization and the exclusion of personal or sensitive data are necessary, and proportionality (limits on length and versions) should guide use. Closed or costly infrastructures and connectivity gaps can shift control away from territories, while showcase technologies with low interactivity (e.g., certain 360° tours) risk sterility and distance people from practices, creating organizational and linguistic lock-in. The strongest ethical risk is disembodiment: when AI (or the digital) replaces living practice with its representational double, ICH is uprooted from its vital matrices – body, relationship, place – and slides toward esthetic consumption; mediation should therefore accompany practices, not replicate them, with generated content pointing back to contexts and uses, citing local sources, and being corrected with the community. Operational mitigations follow: proportionality of use (only what is essential, short texts, retrieval-over-generation); reuse and caching (approved snippets, limited re-generations, versioning, and batching); community glossaries and co-curation; transparency by design (labels, source citation, a brief transparency note, prepublication checklist); governance of cultural data (consent, licenses consistent with community wishes, rights to withdraw/correct); and attention to divides (restrained platform choices, portable formats, networks that share consent models, glossaries, and quality rubrics).

AI use is ethically acceptable in eco-museums when (1) it strengthens accessibility and understanding without replacing practice; (2) it respects cultural ownership and community voice; (3) it is transparent and locally reviewed; (4) it limits computational demand and avoids overproduction; and (5) it reduces dependencies and divides through proportionate technical choices. Otherwise, risks of simulacrum, standardization, and appropriation prevail – at odds with the eco-museum’s civic mission and the living nature of ICH.

11. The ChatGPT case

This section presents ChatGPT not as a technical protocol to be implemented but as an illustrative case that helps ground the broader arguments of the chapter. Its purpose is to explore, through a widely known and accessible tool, how generative AI may, or may not, align with the specific missions and constraints of eco-museums. The analysis is deliberately framed from a critical operational standpoint, not a computational one. Within this approach, ChatGPT serves as a proposal to reflect on the conditions under which generative AI can support accessibility and mediation in ways that remain consistent with the values of community ownership, contextual fidelity, and environmental responsibility. Its selection reflects both its pervasiveness in the cultural sector and its symbolic role in current AI discourse. On the benefits side, it can translate, contextualize, and interpret cultural narratives in support of digital and hybrid experiences – repositories, online exhibitions, installations – overcoming linguistic and geographic barriers, facilitating intercultural dialog, and sustaining educational outreach. The generation of short texts, light-touch curatorship, and collaboration tools can help renew educational and mediation practices, provided the community’s voice and ownership remain clear and the algorithm stays an editorial aid under human control. Alongside these outcomes, however, there is an environmental flip side and necessary ethical cautions.

On the former, currently available orders of magnitude suggest that a single query can require 10 to 100 times the energy of an email, and that flows of hundreds of millions of queries could match the daily electricity consumption of tens of thousands of U.S. households; on water, processing 5–50 queries can require up to half a liter for cooling, with seasonal variation and dependence on data-center location, while recent water-use increases attributable to AI applications (on the order of + 34% and + 30% between 2021 and 2022 for major providers) signal a systemic risk if use is not measured and contained [50].

These figures should not be taken as absolute – because they depend on infrastructure and energy mix – but they are sufficient to justify parsimony and to discourage ornamental or redundant uses. Ethically and culturally, risks include language standardization, bias, and supply-chain opacity (models, data, footprint), which call for participatory frameworks and transparency: whenever ChatGPT intervenes in mediation, institutions should disclose this to the public, cite local sources, document editorial choices, and maintain local human-in-the-loop review.

What follows is a disciplined approach grounded in proportionality and computational parsimony: generate only what is needed for legibility and access, set thresholds on length and versions to avoid unnecessary re-generations, prefer retrieval-over-generation where validated community materials exist, and reuse approved snippets via caching. This is combined with an editorial pipeline that recognizes community ownership (outputs as drafts to be reviewed with knowledge bearers, shared glossaries, rights of review or refusal, labeling of AI involvement, and attribution of local sources), minimal metrics and auditability via user-side proxies (number of requests, average tokens per request, reuse/cache rate, re-generations) to enable periodic retrospectives and guide containment. Infrastructure and localization choices are oriented, where configurable, toward renewable mixes and more efficient cooling with provider disclosure on energy and water, or, where that is not possible, toward prudential criteria (shorter texts, batching, caching) made explicit to audiences. Finally, there is an emphasis on avoiding showcase uses that fuel sterile overproduction and decontextualized simulacra of ICH, reaffirming that AI is acceptable only when it strengthens transmission and in-person experience rather than replacing them.

Within this frame, the opportunity set (multilingual and plain-language access; assembling micro-narratives and maps across material and intangible heritage; editorial support for small teams; expanded educational outreach when integrated into experiential pathways) must always be weighed against costs and risks (energy and water for inference scaling with volumes and length; bias and register standardization; supply-chain opacity; editorial rebound as production becomes easier), which can be mitigated through parsimony, retrieval, reuse, local review, labeling, and monitoring proxies. The implication for the chapter is that the ChatGPT case, read through eco-museum constraints, does not deny the usefulness of LLMs; it specifies the conditions under which that usefulness can be converted into public value without offloading environmental and cultural costs. In line with the European framework (Green Deal, trustworthy AI) and with the 5.0 approach, adoption makes sense when it is proportionate, transparent, and auditable; otherwise, it yields simulacra and footprints disproportionate to the mission. The methodological section that follows translates these principles into a replicable test design – tasks, rubrics, proxies, and safeguards – to verify in practice how far ChatGPT can help without exceeding the bounds of parsimony.

12. Conclusion

Eco-museums, as process-based and place-bound institutions, safeguard ICH by sustaining silent knowledge – embodied, practice-based competencies transmitted across generations. In this context, AI is not an end in itself but a contingent means. Its acceptability depends on proportionality and accountability: technology must remain subordinate to the cultural mission and to ecological constraints.

The article argues that the decisive unit is the editorial deliverable, not the algorithm. AI use should be situated in places and practices, traceable to local sources, reversible without loss of meaning, and measured for its marginal computational load. Read through the European framework (Green Deal, Ethics Guidelines for Trustworthy AI, AI Act), this translates into a concise set of conditions: community ownership and the right of review; governance of cultural data (consent, licensing); minimization of compute (short, contextual texts; retrieval-over-generation; reuse/caching; limits on re-generations and length); labeling of AI involvement with human-in-the-loop review; and prudential choices regarding infrastructure and providers.

Under these conditions, AI can support access, mediation, and inclusion – multilingual plain-language materials, micro-itineraries, and succinct narrative links between material and intangible heritage – without supplanting practice or diluting local registers. Where conditions cannot be met, restraint is the responsible choice: unmeasured use risks disembodiment of ICH, standardization of language, rebound effects in production, and disproportionate environmental footprints.

The resulting stance is clear. For eco-museums and, by extension, for heritage institutions committed to participation and sustainability, AI is optional and conditional: adopt only when it serves transmission, protects community voice, and measures its effects. Otherwise, decline. This proportional approach preserves the primacy of place and practice, ensuring that digital means do not erode the silent knowledge on which ICH rests.