Artificial Intelligence and Cultural Heritage Preservation in GLAM Institutions: A Review of Computer Vision, Natural Language Processing, and 3D Reconstruction for Tangible and Intangible Heritage

Abstract

Artificial intelligence (AI) is increasingly implicated in the preservation, interpretation, and accessibility of cultural heritage across galleries, libraries, archives, and museums (GLAM). This review article synthesizes interdisciplinary scholarship and authoritative standards on the use of machine learning, computer vision, natural language processing (NLP), and three-dimensional (3D) reconstruction for cultural heritage preservation, with attention to both tangible collections (objects, images, monuments, and documentary heritage) and intangible heritage (performance, oral tradition, craft knowledge, and living practices). Drawing on work in heritage studies, digital humanities, information science, computer vision, and preservation science, the review maps major application domains—digitization and capture; automated documentation and metadata enrichment; multilingual retrieval and knowledge organization; 3D reconstruction, monitoring, and simulation; immersive interpretation; and public engagement—alongside persistent risks such as algorithmic bias, colonial epistemologies embedded in data, legal constraints, and long-term sustainability challenges. The article contributes a conceptual framework—the Responsible Heritage AI Lifecycle —to align technical pipelines with heritage values: authenticity, provenance, stewardship, community governance, and durable access. The review argues that the most consequential shift introduced by AI is not simply operational efficiency, but a reconfiguration of what counts as evidence, description, and authority in cultural memory institutions. Responsible deployment therefore requires integrated methods for model transparency, data sovereignty, uncertainty communication, and preservation of both digital surrogates and AI-derived outputs.

Keywords: cultural heritage; digital preservation; GLAM institutions; machine learning; intangible heritage

Introduction

Over the past three decades, digitization has moved from a primarily documentary and access-oriented practice toward a more expansive reengineering of cultural heritage work: collections description, cataloging, conservation assessment, exhibition design, educational outreach, and community knowledge-sharing. In that trajectory, artificial intelligence—especially machine learning methods in computer vision and NLP—has become a major force shaping how heritage is rendered legible and discoverable in digital form. This transformation is not limited to speed or scale. By automating description, classification, transcription, translation, and recommendation, AI can subtly shift interpretive authority and reshape institutional memory practices. At the same time, new 3D reconstruction techniques and immersive interfaces are altering the ways visitors encounter heritage as image, environment, and simulation, reconfiguring relationships among originals, surrogates, and experiences.

In GLAM institutions, AI is typically framed as an instrument: improved digitization workflows, automated documentation, multilingual access, and broader public engagement. Yet in cultural studies terms, AI is also a mediation regime—an assemblage of data, models, infrastructures, labor, and institutional priorities that changes how heritage is categorized and valued. Digitization once raised urgent questions about aura, authenticity, and reproduction; AI intensifies these questions by generating new descriptive layers (e.g., inferred tags, summaries, entities, attributions), new cross-collection linkages (knowledge graphs and embeddings), and new synthetic representations (reconstructions and generative outputs). ¹

This review addresses AI and cultural heritage preservation as both technical practice and cultural formation. It focuses on three core technical families—computer vision, NLP, and 3D reconstruction—because they currently dominate applied deployments in GLAM institutions and because they speak directly to long-standing heritage concerns: what can be documented, what can be retrieved, what can be authenticated, what can be re-experienced, and who gets to define meaning. The article also foregrounds intangible heritage, where AI promises expanded documentation and access while raising distinctive questions about consent, cultural protocols, and the dangers of extracting “data” from living traditions.

Two foundational claims orient the review. First, AI applications in cultural heritage preservation are most effective when treated as sociotechnical systems rather than as software add-ons: they require curatorial judgment, domain-specific data governance, and preservation planning for both inputs and outputs. Second, because cultural heritage data are historically structured by colonial collecting, cataloging bias, and institutional asymmetries, AI systems trained on such data can reproduce and intensify inequities unless mitigated through participatory governance and explicit ethical design.

Method and Scope of the Review

Review approach

This is a narrative review grounded in cross-disciplinary synthesis rather than a strict meta-analysis, reflecting the heterogeneity of evidence in arts and cultural studies: peer-reviewed computer science papers, heritage standards and charters, museum informatics research, preservation frameworks, and critical scholarship on digitization and cultural authority. The review emphasizes sources that are either peer-reviewed or widely recognized as authoritative in the heritage and preservation communities (e.g., UNESCO recommendations, ISO standards, ICOMOS/charter documents). ²

Definitions and boundaries

Cultural heritage is treated as encompassing both tangible and intangible forms, consistent with UNESCO’s distinction between safeguarding living heritage and preserving documentary and material culture. ³ Digital preservation refers to the long-term stewardship of digital assets and their associated metadata, including the durability and interpretability of files, databases, and—crucially for this review—machine learning outputs and models. The GLAM acronym is used to emphasize institutional commonalities in stewardship, access, and public mission, while acknowledging differences in professional norms across archives, libraries, museums, and galleries.

AI in this article primarily denotes machine learning approaches (including deep learning) used for perception and language tasks, rather than symbolic AI. The review also addresses multimodal foundation models (models trained across text and image) because of their growing importance for cross-collection discovery and multilingual retrieval. Where relevant, the article distinguishes between (a) AI applied to digitized representations of heritage and (b) AI used for in situ monitoring and analysis of heritage sites and materials.

Literature Review: From Digital Heritage to Heritage AI

Digitization, authenticity, and the politics of representation

Digital cultural heritage scholarship has long stressed that digitization is not neutral reproduction but interpretive transformation—“recoding” institutional priorities into databases, interfaces, and metadata practices. ⁴ Early debates about mechanical reproduction and authenticity are frequently revisited in digital heritage contexts, especially when surrogates become primary access points and when computational systems mediate interpretation at scale. ⁵ In museums, archives, and libraries, digitization introduced new tensions between preservation and access, and between universalist discovery and culturally specific restrictions. These tensions intensify with AI, which often requires large-scale aggregation, standardization, and reuse of data that may be culturally sensitive or legally constrained.

Heritage charters and standards provide a complementary lineage. The Nara Document on Authenticity emphasizes that authenticity is culturally contingent and not reducible to material fact alone. ⁶ The London Charter and Seville Principles—developed to guide computer-based visualization—stress transparency, documentation, and intellectual rigor in digital reconstructions and interpretive simulations. ⁷ These principles anticipate current debates about AI-generated reconstructions, synthetic media, and the evidentiary status of algorithmic outputs.

Digital preservation frameworks and the problem of “AI outputs”

Digital preservation is often organized around reference models such as OAIS (Open Archival Information System), which conceptualizes ingest, archival storage, data management, preservation planning, and access as interlocking functions. ⁸ Trustworthy repository standards (e.g., ISO 16363) and community frameworks (e.g., the NDSA Levels of Digital Preservation) operationalize these functions into audit criteria and maturity levels. ⁹

However, the literature has only begun to grapple with how AI complicates preservation planning. AI produces new classes of artifacts: training datasets, data splits, labels and annotation guidelines, model weights, embeddings, vector indexes, prompts and system instructions, evaluation reports, and model cards. Some of these artifacts are necessary for reproducibility, while others are difficult to preserve because they are platform-bound, rapidly evolving, or legally encumbered. The resulting preservation problem is not simply file format obsolescence; it is the preservation of computational interpretability —the ability to explain how outputs were derived, with what uncertainty, using what data and assumptions.

Computer vision and collections: from images to evidence

Computer vision in GLAM settings typically begins with digitized images of artworks, photographs, manuscripts, and objects. Convolutional neural networks (CNNs) catalyzed performance gains in classification and feature learning, enabling more robust tagging and similarity search than earlier handcrafted-feature pipelines. ¹⁰ More recently, transformer-based vision architectures and large-scale pretraining have shifted the focus from task-specific models to reusable representations. ¹¹

In heritage contexts, vision tasks include: (a) object detection and segmentation for item isolation and condition assessment; (b) style, iconography, or material classification (with strong caveats about interpretive reduction); (c) visual similarity and “more like this” discovery; (d) reading text in images through OCR; and (e) detecting change over time in site monitoring. Foundation models and segmentation models trained at scale can reduce the need for bespoke labeling, but their generality also introduces opacity and domain mismatch risks for heritage materials (e.g., non-Western iconographies, degraded manuscripts, low-resource scripts, or specialized material textures). ¹²

NLP, documentary heritage, and multilingual access

Libraries and archives have long used language technologies for information retrieval and metadata normalization. Contemporary NLP—especially transformer architectures—supports entity recognition, relation extraction, topic modeling, summarization, and cross-lingual retrieval at unprecedented scale. ¹³ These capabilities matter for cultural heritage because documentary collections often have sparse metadata, inconsistent descriptive practices, and multiple languages across cataloging histories, colonial contexts, and diasporic collections.

At the same time, archival description and museum cataloging are not merely technical tasks; they encode institutional perspectives and power relations. Automated NLP can reproduce those perspectives if trained on historical descriptions that include outdated terminology, racialized language, or incomplete provenance. Moreover, language models can produce fluent but unfounded assertions (“hallucinations”), making them risky as direct descriptive authorities without verification and provenance controls. ¹⁴

3D reconstruction and the turn to spatial heritage data

3D capture and reconstruction in heritage settings include laser scanning, structured light, and photogrammetry. Landmark projects such as the Digital Michelangelo Project demonstrated how high-resolution scanning can support conservation, study, and public engagement, while also raising challenges of storage, documentation, and interpretive framing. ¹⁵ Photogrammetric pipelines—especially structure-from-motion and multi-view stereo—have made 3D reconstruction more accessible by using overlapping photographs to compute camera poses and dense geometry. ¹⁶ Heritage applications include monument documentation, reconstruction hypotheses, deformation monitoring, and immersive interpretation in AR/VR environments.

In cultural heritage preservation, 3D data are both measurement and interpretation. Even when reconstruction is grounded in physical measurement, choices about resolution, cleaning, meshing, texture mapping, and missing-data inference are consequential. Visualization principles therefore matter as much as technical accuracy, especially when reconstructions are presented as public-facing “restorations” rather than as conditional hypotheses.

Ethics, governance, and documentation for machine learning in heritage

ML ethics scholarship provides tools for documenting datasets and models in ways aligned with accountability. Datasheets for datasets and model cards are now widely cited mechanisms for clarifying data provenance, collection contexts, intended use, limitations, and performance characteristics. ¹⁷ In heritage, such documentation intersects with community governance principles, including Indigenous data sovereignty and culturally grounded access protocols. ¹⁸ The CARE Principles (Collective Benefit, Authority to Control, Responsibility, Ethics) are especially salient where heritage materials are tied to living communities and cultural rights, complementing FAIR principles focused on technical discoverability and reuse. ¹⁹

International policy documents increasingly foreground these concerns. UNESCO’s Recommendation on the Ethics of AI emphasizes human rights, cultural diversity, and impact assessment, offering a policy backdrop for GLAM institutions adopting AI at scale. ²⁰ Yet translation into operational practice remains uneven, especially in resource-constrained institutions and in cross-border collaborations where legal regimes and cultural protocols differ.

Synthesis of Studies: Where AI Is Transforming Cultural Heritage Preservation

1) Digitization and capture: scaling surrogates while preserving context

Digitization in GLAM historically focused on imaging standards, metadata, and access systems. AI changes digitization by making capture computationally actionable : images and scans become inputs for segmentation, transcription, reconstruction, and feature extraction. Three trends stand out.

(a) Intelligent quality control and workflow triage. Computer vision can assist in identifying out-of-focus images, skew, color calibration issues, page curvature, and missing pages. While these uses are often underreported in scholarly venues compared to “headline” applications, they are consequential for scale digitization because they reduce rework and improve downstream OCR and retrieval.

(b) OCR and HTR as preservation-adjacent technologies. Optical character recognition has long been part of library digitization, and open-source engines like Tesseract represent a durable technical lineage. ²¹ Modern deep-learning OCR and handwritten text recognition (HTR) have improved performance on complex layouts and historical scripts, but results remain uneven for low-resource languages, degraded documents, and non-Latin scripts. From a preservation standpoint, OCR/HTR outputs must be treated as derivative texts with uncertainty, not as neutral transcriptions. The preservation object becomes a bundle: image + transcription + confidence + model/version metadata.

(c) Audio and audiovisual heritage. Although this review emphasizes vision and text, the preservation of oral histories, language recordings, and performance documentation increasingly uses speech recognition, diarization, and alignment. These are particularly relevant for intangible heritage safeguarding, yet high-quality models are still uneven across languages and dialects, and speech technologies may misrecognize culturally specific names and terms. Uncertainty and community review are therefore integral to responsible use.

Digitization also raises sustainability concerns: AI can encourage “digitize everything” logics while shifting costs downstream into storage, compute, and ongoing model maintenance. Conway’s account of digitization in libraries remains pertinent: preservation goals can be displaced by access metrics, and the meaning of “preservation” can be narrowed to scanning rather than long-term stewardship. ²²

2) Automated documentation and metadata enrichment

Metadata is central to cultural heritage preservation because it carries provenance, context, and interpretive frames. AI has been deployed to enrich metadata by suggesting tags, extracting entities from text, clustering items by similarity, and mapping descriptive fields across schemas. These practices can increase discoverability, but they also risk flattening cultural specificity and embedding model assumptions into catalog records.

(a) Vision-based tagging and similarity search. Deep learning representations enable “visual search” across collections, supporting discovery when textual metadata are sparse. Multimodal models that align images and text (e.g., CLIP-style approaches) can retrieve images with natural-language queries and vice versa. ²³ For GLAM, this can help users find visually similar artifacts across institutional boundaries, but interpretive caution is required: similarity in embedding space is not equivalence in iconographic meaning or cultural function.

(b) Entity extraction, linking, and knowledge graphs. NLP can identify people, places, organizations, dates, and works in finding aids, exhibition catalogs, and legacy documentation. Linked data initiatives commonly rely on controlled vocabularies and conceptual models such as CIDOC CRM (ISO 21127), aiming to formalize relationships among events, objects, and actors in ways that support interoperability. ²⁴ AI can accelerate entity reconciliation and suggest links, but human oversight is essential where names are ambiguous, transliterations vary, or colonial naming practices obscure Indigenous or local identifiers.

(c) Description as a contested domain. Automated description intersects with well-documented critiques of archival and museum description as historically shaped by institutional power. AI can assist in identifying harmful language and surfacing alternative terminologies, but it can also perpetuate bias if trained on existing catalogs. Responsible deployments increasingly combine algorithmic suggestions with editorial workflows, audit trails, and community consultation—particularly for collections involving Indigenous peoples, enslaved persons, and other historically marginalized communities.

To clarify what “enrichment” means across institutions, Table 1 maps common AI techniques to heritage tasks and the types of heritage most affected.

3) Multilingual information retrieval and cross-cultural access

Multilingual access is central to cultural heritage because collections circulate across linguistic boundaries via collecting histories, migration, and colonial extraction. AI contributes in three main ways: translation/normalization of metadata, cross-lingual retrieval, and multilingual interfaces for public engagement.

(a) Cross-lingual embeddings and retrieval. Transformer-based multilingual language models (e.g., XLM-R) enable retrieval across languages by mapping semantically similar phrases to nearby representations. ²⁵ In GLAM contexts, this can support users searching across multilingual catalogs even when metadata are unevenly translated. Yet “semantic equivalence” is not guaranteed; culturally specific terms, kinship categories, and place names may resist direct translation. A responsible retrieval system should therefore expose language mappings and allow user feedback rather than silently substituting terms.

(b) Standards, interoperability, and the limits of “universal” description. Interoperability initiatives (e.g., linked data models, metadata crosswalks, and image interoperability frameworks) aim to reduce fragmentation across institutions. But cultural studies scholarship warns that universalizing metadata can erase local epistemologies and impose externally defined categories. This concern is sharpened when AI systems “learn” categories from standardized schemas that were not designed for all cultural contexts.

(c) Multilingual public engagement. AI-assisted translation can expand access for audiences underserved by dominant-language catalogs. However, automated translation errors can be reputationally damaging and culturally harmful, especially when dealing with sacred items, culturally restricted knowledge, or politically sensitive histories. Institutions increasingly need translation governance: what may be machine-translated, what requires human review, and how corrections feed back into both catalog and model improvement.

One way to formalize AI-based retrieval is to describe multimodal similarity search. If an image (or text) is embedded into a vector space, similarity is frequently computed by cosine similarity:

Equation (1) is widely used in embedding-based retrieval. In heritage settings, the interpretive challenge is not the mathematics but the epistemology: the meaning of “similarity” depends on training data, labeling histories, and the cultural specificity of descriptions. Therefore, retrieval interfaces should present similarity as suggestive rather than authoritative and should support provenance display for why an item was retrieved.

4) 3D reconstruction, monitoring, and evidentiary visualization

3D reconstruction has become a cornerstone of digital heritage, enabling detailed documentation and new interpretive experiences. The technical literature emphasizes pipelines such as structure-from-motion (SfM) and multi-view stereo (MVS), which infer camera parameters and dense geometry from overlapping images. ²⁶ Heritage practice adds additional requirements: conservation-grade documentation, paradata (documentation of interpretive decisions), and long-term preservation of complex 3D formats and dependencies.

(a) From capture to reconstruction. Photogrammetry’s appeal lies in accessibility: many projects can be executed with consumer cameras and appropriate calibration. Yet capture protocols (lighting, overlap, scale bars, lens distortion) strongly shape results, and heritage objects often pose challenges—reflective surfaces, repetitive patterns, occlusions, and fragile contexts. Remondino and El-Hakim stress that heritage recording requires integrating technical accuracy with pragmatic constraints and documentation of methods. ²⁷

(b) Monitoring change over time. Where repeated captures are possible (e.g., heritage buildings, monuments, landscapes), 3D data can support deformation analysis and risk assessment. AI-assisted change detection can help identify cracks, surface erosion, or vandalism in image sequences. Yet such systems require careful calibration to avoid false alarms due to lighting changes, seasonal variation, or sensor differences.

(c) Reconstruction as hypothesis and the need for transparency. Visualization principles from the London Charter and Seville Principles insist that digital reconstructions should document sources, uncertainty, and interpretive choices. ²⁸ In contemporary practice, that means versioning reconstructions, preserving intermediate outputs, and publishing paradata alongside models—especially when AI methods fill gaps or denoise scans in ways that may be visually compelling but evidentially ambiguous.

Figure 1 provides a conceptual workflow, emphasizing where interpretive decisions enter and where preservation metadata should be captured.

5) Intangible heritage: documentation, safeguarding, and AI’s extraction risks

Intangible cultural heritage (ICH) includes practices, expressions, knowledge, and skills that communities recognize as part of their cultural heritage. UNESCO’s 2003 Convention frames safeguarding as ensuring viability through transmission, education, and community participation, not simply documentation. ²⁹ AI intersects with ICH in at least four ways: documentation and indexing, language revitalization support, creative re-performance and simulation, and risk modeling for heritage endangerment. Each brings distinctive benefits and hazards.

(a) Documentation and indexing of living practices. AI can help segment and index long audiovisual recordings of performance, craft demonstrations, and oral histories. It can support searching within archives by speaker, topic, or musical motif. Yet the premise that an ICH practice is reducible to “content” that can be extracted and indexed can conflict with community understandings of knowledge as relational, situated, and sometimes restricted.

(b) Language technologies and revitalization. NLP and speech recognition may support transcription, dictionary building, and educational tools for endangered languages, but many languages are low-resource with limited training data, and data collection can itself be ethically fraught. Governance frameworks from Indigenous data sovereignty scholarship emphasize that communities should control how language data are collected, stored, and reused. ³⁰

(c) Generative systems and the limits of cultural simulation. While generative AI may be used to create educational demonstrations or speculative reconstructions, it raises acute concerns for ICH: generating “new” songs, chants, or designs in the style of a community can blur boundaries between cultural transmission and appropriation, especially when models are trained without consent or compensation. Current policy guidance often lags behind these scenarios, making institutional restraint and participatory governance essential.

(d) Safeguarding as a socio-technical practice. The ICH literature stresses that safeguarding is not only preservation of information but sustaining conditions for practice. ³¹ AI systems can support safeguarding only if aligned with community goals, governance, and benefit-sharing, rather than treating communities as data sources. In other words, AI can assist cultural heritage preservation, but it cannot substitute for cultural continuity.

6) Immersive interpretation and public engagement

AI also shapes how the public encounters heritage through recommendation systems, personalization, conversational interfaces, and immersive installations. In museum contexts, personalization can increase engagement and accessibility, but it can also narrow interpretive horizons if it over-optimizes for “relevance” and under-exposes users to challenging histories. Conversational interfaces can help visitors navigate complex collections, yet they pose epistemic risks: if a system generates plausible but incorrect explanations, it can quickly become an authority in the visitor’s mind.

Immersive heritage experiences—VR walk-throughs, AR overlays, interactive 3D viewers—often depend on 3D models and AI-driven interaction (e.g., segmentation, tracking, adaptive narration). The London Charter’s insistence on transparency suggests that immersive systems should disclose when a scene is documented, reconstructed, or speculative. For researchers, immersive systems also raise methodological questions: what is being preserved—the site, the model, the experience design, or the interpretive narrative? Preservation planning must clarify which layers are essential and how they can be re-rendered as hardware and software environments change.

Discussion: Toward a Responsible Heritage AI Lifecycle

Why AI requires a heritage-specific framework

In many applied domains, responsible AI discussions emphasize fairness, accountability, and transparency. In cultural heritage, these concerns remain crucial but must be reframed through heritage-specific values: authenticity (culturally defined), provenance, stewardship, interpretive pluralism, and obligations to communities of origin. Heritage data are also unusually heterogeneous (images, texts, 3D, audiovisual, conservation records) and unusually entangled with rights and restrictions (copyright, moral rights, privacy, cultural protocols). As a result, generic “AI deployment” checklists often fail to address the core heritage problem: how to ensure that AI-derived representations remain intelligible, contestable, and preservable over time.

This section proposes a conceptual framework—the Responsible Heritage AI Lifecycle —intended to help GLAM researchers and practitioners evaluate AI projects not only by performance metrics but by evidentiary integrity, governance, and preservation outcomes.

The Responsible Heritage AI Lifecycle (framework)

The framework comprises six linked phases, each with research and governance questions:

• Phase 1: Heritage problem formulation. What preservation question is being addressed (documentation, risk monitoring, access, interpretation)? Who defines success? What harms are plausible (misdescription, exposure of restricted knowledge, reputational damage)?
• Phase 2: Data governance and stewardship design. What are the rights and restrictions? Are CARE and community protocols applicable? What consent exists for training and reuse? How will sensitive data be protected?
• Phase 3: Model selection and documentation. Is a domain-specific model needed, or is a foundation model acceptable? What are known limitations? Are datasheets/model cards produced and preserved? ³²
• Phase 4: Evaluation and uncertainty communication. What metrics are used, and what do they miss? How will uncertainty be shown to researchers and publics? Is there human review?
• Phase 5: Deployment and interpretive accountability. How are outputs integrated into catalogs or exhibitions? Are AI suggestions clearly labeled? Is there an audit trail?
• Phase 6: Preservation, reproducibility, and exit strategy. What artifacts will be preserved (datasets, code, weights, embeddings, prompts, paradata)? What is the plan if a vendor model changes or is withdrawn?

Figure 2 shows how these phases align with a typical GLAM AI pipeline and highlights where preservation and ethics checkpoints should be embedded.

Evaluation: metrics, meaning, and the problem of interpretive reduction

ML evaluation commonly uses accuracy, precision, recall, and F1. In heritage contexts, these metrics are necessary but insufficient because the “ground truth” is often contested: iconographic categories shift across cultures; names vary across languages; and provenance may be uncertain. Nonetheless, transparent metrics help prevent overclaiming and enable comparison across methods.

The F1 score is commonly defined as:

Equation (2) is useful, but heritage projects should supplement it with error typologies: which misclassifications occur, and do they disproportionately affect particular cultures, languages, or material categories? In addition, confidence scoring should be preserved and surfaced. When outputs enter catalog records, institutions should consider representing AI assertions as separate, attributable statements rather than overwriting curatorial description.

Table 2 proposes reporting elements and evaluation practices tailored to cultural heritage preservation.

Data sovereignty, consent, and decolonial commitments

Digitization and AI can enable “digital return” or “digital repatriation,” offering communities access to dispersed heritage materials. Yet openness is not inherently ethical. Christen’s influential critique of “information wants to be free” highlights that Indigenous knowledge systems often require governance mechanisms incompatible with default open-access models. ³³ Mukurtu and similar platforms exemplify how access controls can be structured around cultural protocols, demonstrating that preservation and access are compatible with community authority when systems are designed accordingly.

From an AI perspective, this means that training datasets cannot be assumed to be a neutral pool of reusable content. CARE principles demand authority to control and collective benefit, shifting attention from technical accessibility to governance and reciprocity. ³⁴ For researchers, this implies methodological changes: participatory dataset design, culturally appropriate consent, restrictions on model training, and benefit-sharing agreements. It also implies that “performance” is not the only success criterion; legitimacy and community-defined benefit are equally central.

Transparency, provenance, and the preservation of AI as heritage infrastructure

Heritage institutions preserve artifacts and records, but increasingly they also preserve the infrastructures that make artifacts legible: databases, metadata standards, controlled vocabularies, and interfaces. AI extends this infrastructural layer into model weights and embedding indexes. If a GLAM institution uses embeddings to power search across millions of items, the embeddings themselves become a preservation object because they mediate access and interpretive discovery.

Here, OAIS-like thinking is instructive: the “representation information” needed to understand digital objects must now include how AI representations were created, what training data were used, and what versioned model produced them. ³⁵ Without this, future researchers may inherit opaque similarity systems whose epistemic assumptions cannot be reconstructed.

This challenge is amplified by vendor dependence. Many GLAM institutions lack capacity to train large models and instead rely on commercial APIs. An exit strategy is therefore part of preservation planning: how will the institution maintain continuity of access if a model is deprecated, terms of service change, or the vendor withdraws a product?

Generative AI and the evidentiary boundary

Generative models can assist in summarization, transcription cleanup, translation drafts, and even speculative visualization. But they can also erode evidentiary boundaries by producing outputs that look authoritative without being grounded in sources. The “stochastic parrots” critique argues that large language models can reproduce biases, amplify harmful content, and generate plausible text without understanding, raising risks when used as knowledge sources. ³⁶ In heritage contexts, these risks are acute because visitors and even researchers may interpret fluent text as curatorial fact.

Accordingly, a conservative principle is warranted: generative AI outputs should be treated as proposals or finding aids unless they are explicitly verified, sourced, and preserved with provenance metadata. This aligns with visualization charters that emphasize transparency and documentation for interpretive claims. When generative AI is used for public-facing interpretation, institutions should disclose the role of AI, document the prompt and model version, and provide pathways for correction.

Energy, sustainability, and institutional capacity

AI systems can impose significant compute and storage costs, creating sustainability pressures that are unevenly distributed across institutions. Large institutions may run on-premise GPU clusters, while smaller museums and community archives may depend on external services. From an equity perspective, this can deepen asymmetries: well-resourced institutions can produce “AI-enhanced” catalogs and immersive experiences, while others cannot. Collaborative infrastructures, shared open models, and consortial approaches can mitigate this, but they require governance and sustained funding.

Conclusion

AI is transforming cultural heritage preservation in GLAM institutions by expanding what can be digitized, described, retrieved, reconstructed, and experienced. Computer vision supports scalable documentation and discovery across images and objects; NLP enables transcription, entity extraction, and multilingual retrieval; and 3D reconstruction underpins both conservation documentation and immersive interpretation. These capabilities can materially improve access and stewardship, especially for under-described collections and for geographically dispersed heritage.

Yet AI’s most significant impact is epistemic. By producing new descriptive layers and automating interpretive steps, AI systems reshape what counts as knowledge about heritage objects and practices. In doing so, they amplify longstanding concerns in arts and cultural studies: authenticity is not only a technical property; description is not only informational; and preservation is not only storage, but a set of social obligations and interpretive responsibilities.

The article’s proposed Responsible Heritage AI Lifecycle argues for integrating governance, documentation, evaluation, and preservation planning into AI pipelines from the outset. For researchers, key priorities include: (1) robust documentation of datasets and models; (2) uncertainty communication and auditability; (3) community governance and data sovereignty, especially for intangible heritage and culturally sensitive collections; (4) preservation strategies for AI-derived artifacts, including embeddings and paradata; and (5) sustainable infrastructures that do not exacerbate inequities among institutions.

AI can support cultural heritage preservation most effectively when it is treated not as an automated substitute for curatorial work but as an interpretive instrument whose outputs remain attributable, contestable, and preservable over time.

Notes

References