The Paradox at the Heart of Digital Preservation
In September 2026, UNESCO and the International Council of Museums released a landmark report on the use of artificial intelligence in the global museum sector. The findings were striking in their contradictions. Approximately 57 per cent of responding museums are currently utilising AI — a figure that would have seemed extraordinary a decade ago. Yet 55 per cent of those same institutions have no internal AI strategy or guidelines, and 62 per cent reported that their countries lack specific legislation or policies governing AI in the museum sector. The technology is racing ahead of the governance, and the stakes could not be higher: the institutions responsible for preserving humanity's cultural memory are deploying powerful tools they do not yet fully understand, without the frameworks needed to ensure those tools serve the communities whose heritage they hold.
This paradox sits at the centre of one of the most consequential debates in contemporary cultural policy. Artificial intelligence offers genuinely transformative capabilities for cultural heritage preservation: the ability to digitise and analyse vast archives at speeds no human team could match, to reconstruct damaged or destroyed artefacts through computational modelling, to transcribe and translate endangered languages, to detect the illicit trafficking of cultural property through pattern recognition. These are not trivial benefits. Approximately 3,000 indigenous languages are at risk of disappearing by the end of the century. Millions of cultural artefacts remain undocumented in understaffed museum storerooms. Climate change and armed conflict are destroying irreplaceable heritage at an accelerating rate.
But the same technologies that offer these capabilities also carry risks that are only beginning to be understood: the risk of "digital colonialism," in which the cultural materials of marginalised communities are extracted, processed, and interpreted by systems built on Western assumptions and controlled by external institutions; the risk of "epistemic violence," in which AI systems trained on dominant-culture data render minority knowledge systems inexpressible or distorted; the risk of "museumification," in which living, performative, relational heritage is flattened into static data and stripped of the social, emotional, and spiritual contexts that give it meaning.
When living, performative, and relational heritage is flattened into static data, it risks being museumified — stripped of the social, emotional, and spiritual contexts that define it.
What AI Can Do: The Technical Landscape
To assess the risks clearly, it is necessary first to understand the capabilities. UNESCO's Culture Sector has developed a strategic portfolio of AI projects that illustrates the breadth of potential applications. In heritage conservation and risk management, AI is being used for maritime and underwater heritage monitoring through the AI STREAM project, for AI-assisted documentation of heritage sites in crisis zones, and for AI-powered debris profiling to ensure safer recovery of cultural materials following conflict or natural disaster. In the fight against illicit trafficking, UNESCO is developing a "Due Diligence Room" within its Virtual Museum of Stolen Cultural Objects, an AI-enhanced version of the CultNatLaws database, and pilot monitoring programmes to detect the online trade of cultural property.
For museums specifically, the most common current applications are pragmatic rather than transformative: text drafting, translation, and administrative assistance. But specialised applications are gaining traction. Handwritten-text recognition is enabling the digitisation of archival documents that would otherwise require years of manual transcription. Computer vision is being used to catalogue and analyse large collections, identifying patterns and connections that human curators might miss. Accessibility features powered by AI — audio descriptions, real-time translation, personalised navigation — are making museum collections available to audiences previously excluded by language, disability, or geography.
For intangible heritage — the living practices, oral traditions, performing arts, and knowledge systems that UNESCO's 2003 Convention on Intangible Cultural Heritage was designed to protect — natural language processing and automatic speech recognition are the critical technologies. The challenge is acute: intangible heritage is, by definition, not written down. It exists in performance, in conversation, in the transmission of knowledge from practitioner to apprentice. Capturing it requires recording, transcribing, and making searchable vast quantities of audio and video material in languages for which training data is scarce and for which standard AI tools frequently fail.
When living, performative, and relational heritage is flattened into static data, it risks being museumified — stripped of the social, emotional, and spiritual contexts that define it.
The Data Sovereignty Imperative
The most instructive case study in community-controlled AI for cultural preservation is Te Hiku Media, a non-profit broadcaster in New Zealand dedicated to the preservation of te reo Māori. Facing the threat of language loss and a 30-year archive of recordings that would take centuries to transcribe manually, the organisation developed its own automatic speech recognition models using the open-source NVIDIA NeMo toolkit. The models achieve 92 per cent accuracy for te reo and 82 per cent accuracy for bilingual speech — performance levels that rival commercial systems for major languages.
What distinguishes Te Hiku Media's approach is not the technology but the governance. The organisation built its own distribution platform, Whare Kōrero ("house of speech"), which houses over 1,000 hours of archival audio and content from 20 Māori radio stations. Data collection was conducted through the Kōrero Māori crowdsourcing campaign, governed by the Kaitiakitanga licence — a community-developed framework that ensures collected data is used exclusively for the benefit of the Māori people. The campaign gathered over 300 hours of labelled speech data in just ten days, not by surrendering community control but by building infrastructure the community owns.
Te Hiku Media's Kōrero Māori campaign gathered over 300 hours of labelled speech data in just ten days — not by surrendering community control, but by building infrastructure the community owns.
The contrast with extractive approaches is stark. Many AI projects in the cultural heritage space have been conducted by external researchers or institutions that digitise community materials, train models on them, and retain control of both the data and the resulting systems. The community whose heritage was used as training data may receive no benefit, may have no ability to correct errors or remove sensitive materials, and may find that their cultural knowledge has been incorporated into commercial products without their consent. This is the pattern that scholars have characterised as "digital colonialism" — the reproduction, in digital form, of the extractive relationships that characterised physical colonialism.
The Governance Gap: UNESCO's 2026 Survey in Detail
The UNESCO-ICOM 2026 report provides the most comprehensive picture yet of where the museum sector stands on AI governance. The headline figures — 57 per cent adoption, 55 per cent without an AI strategy — tell only part of the story. The report reveals a sector in which AI use is often exploratory or staff-led rather than institution-wide, driven by individual enthusiasts rather than institutional policy. This pattern creates significant risks: without institutional frameworks, there is no mechanism for ensuring that AI use is consistent with the museum's ethical commitments, no process for identifying and mitigating harms, and no accountability when things go wrong.
Fifty-five per cent of museums have no internal AI strategy, and 62 per cent report that their countries lack specific legislation governing AI in the museum sector — a governance vacuum at the precise moment the technology is most consequential.
The ethical concerns identified by survey respondents are revealing. Data privacy tops the list — a concern that is particularly acute for museums that hold sensitive materials about living communities or individuals. Copyright is a close second, reflecting the unresolved questions about whether AI systems trained on digitised cultural materials infringe the intellectual property rights of the communities or individuals whose work was used. Institutional credibility is a third major concern: museums are trusted institutions, and the deployment of AI systems that generate inaccurate or culturally inappropriate content could damage that trust in ways that are difficult to repair.
Te Hiku Media's Kōrero Māori campaign gathered over 300 hours of labelled speech data in just ten days — not by surrendering community control, but by building infrastructure the community owns.
Perhaps most concerning is the concern about AI "flattening" cultural narratives. This is not a technical problem but a philosophical one. AI systems are trained to find patterns and generate outputs that are statistically likely given their training data. Applied to cultural heritage, this tendency produces outputs that reflect the dominant patterns in the training data — which, for most current AI systems, means Western, English-language, post-Enlightenment cultural frameworks. The result can be a subtle but systematic distortion of non-Western, non-English, oral, and indigenous knowledge systems, rendering them legible only through frameworks that were not designed for them.
The Intangible Heritage Challenge
UNESCO's scheduled October 2026 webinar on AI and intangible cultural heritage addresses what may be the most difficult dimension of the problem. Intangible heritage — the living practices, oral traditions, and knowledge systems that communities transmit across generations — is defined by its resistance to fixation. A dance is not a video of a dance. A healing practice is not a description of a healing practice. The knowledge embedded in these practices is relational, contextual, and embodied in ways that resist digital capture.
The risk of AI-mediated intangible heritage preservation is not merely that it will produce inaccurate transcriptions or translations — though it will, particularly for low-resource languages. The deeper risk is that the act of digitisation itself transforms the heritage. When a living practice is captured as data, it becomes an object rather than a process. It can be reproduced, modified, and distributed without the consent or participation of the community that created it. It can be incorporated into AI training datasets and used to generate synthetic versions of the practice that may be indistinguishable from the original to outsiders but are experienced by the community as a form of appropriation.
The African Union's 2025 Continental AI Strategy and similar national policies in Kenya, South Africa, and Nigeria have begun to address these concerns by emphasising the protection of cultural knowledge systems from extraction. The "Decolonial AI Heritage Framework" proposed by researchers in 2026 goes further, requiring developers to conduct an assessment phase that forces them to consider whether the technology serves the community or merely replicates historical exploitation. These frameworks are promising, but they remain largely aspirational in the absence of binding international standards.
UNESCO's Strategic Response
UNESCO's response to these challenges has been to develop a portfolio of projects that attempt to harness AI's capabilities while building in community control and ethical safeguards. The AI STREAM project for maritime heritage monitoring, the Due Diligence Room for combating illicit trafficking, and the UNESCO-ICOM Museums & AI Peer Exchange Platform all represent attempts to use AI in ways that serve the heritage sector's core mission without reproducing extractive patterns.
The partnership with Arius Technology Europa Italia Srl, announced in 2026, focuses on digital preservation using high-resolution 3D scanning and AI-assisted analysis — a technology that can create detailed records of physical artefacts without requiring their physical transfer. The AI for Youth and World Heritage and AI-Powered Youth Storytelling programmes attempt to build AI literacy within communities, ensuring that the next generation of heritage practitioners can engage with these technologies on their own terms.
These initiatives are valuable, but they operate within a broader context in which the dominant AI systems — the large language models, image generators, and speech recognition tools that are increasingly embedded in everyday life — are developed by a small number of technology companies with limited accountability to the cultural heritage sector or the communities it serves. The governance challenge is not merely to ensure that heritage institutions use AI responsibly; it is to ensure that the AI systems themselves are designed in ways that respect cultural diversity, community sovereignty, and the irreducible complexity of human heritage.
Fifty-five per cent of museums have no internal AI strategy, and 62 per cent report that their countries lack specific legislation governing AI in the museum sector — a governance vacuum at the precise moment the technology is most consequential.
Towards a Framework for Ethical AI in Cultural Heritage
The emerging consensus among researchers, practitioners, and policymakers points towards a framework built on four principles. The first is community sovereignty: communities must retain control over their cultural materials, including the right to determine how those materials are digitised, what they are used for, who has access to them, and whether they can be withdrawn from AI training datasets. This principle is not merely ethical — it is increasingly a legal requirement under data protection frameworks including the GDPR and its equivalents.
The second principle is transparency: AI systems used in cultural heritage contexts must be explainable, auditable, and accountable. When an AI system makes a decision about how to classify, describe, or interpret a cultural artefact, the basis for that decision must be accessible to the communities whose heritage is at stake. Black-box systems that produce outputs without explanation are incompatible with the trust relationships that cultural heritage institutions depend on.
The third principle is benefit-sharing: the communities whose cultural materials are used to train AI systems must receive a proportionate share of the benefits those systems generate. This may take the form of improved tools for community use, revenue sharing from commercial applications, or capacity building that enables communities to develop their own AI capabilities. The Te Hiku Media model — in which the community owns the infrastructure and controls the data — represents one approach; others will be needed for communities with different resources and governance structures.
The fourth principle is precaution: where the risks of AI deployment in cultural heritage contexts are uncertain or potentially irreversible, the burden of proof should lie with those proposing to deploy the technology, not with those seeking to protect the heritage. The loss of a living language, the distortion of an oral tradition, the appropriation of sacred knowledge — these are harms that cannot be undone by a subsequent apology or a revised algorithm.
Conclusion: The Memory Keepers Must Also Be the Rule Makers
The institutions responsible for preserving humanity's cultural memory are at a crossroads. The tools available to them are more powerful than anything their predecessors could have imagined. The risks those tools carry are equally unprecedented. The governance frameworks needed to navigate this moment are still being built.
The UNESCO-ICOM 2026 report is a call to action. The 55 per cent of museums without an AI strategy, the 62 per cent operating in countries without relevant legislation, the communities whose heritage is being digitised without their meaningful consent — these are not statistics to be noted and filed. They are evidence of a governance gap that, if left unaddressed, will allow the most powerful tools for cultural preservation in human history to become instruments of cultural extraction.
The memory keepers must also be the rule makers. The communities whose heritage is at stake must be at the centre of the governance frameworks being built, not at the periphery. And the institutions that hold the world's cultural memory in trust must ensure that the tools they deploy serve the communities they represent — not the other way around.
