The standard story about cultural preservation in the digital age has been reassuring. Scan the manuscript, record the language, photograph the object, upload the catalogue, and a threatened inheritance becomes safer than it was in a box, a storeroom or a fading human chain of transmission. That account is not wrong. It is, however, incomplete. By mid-2026, the more consequential development is that digitised culture no longer sits still. It is sorted, scraped, vectorised, captioned, segmented, translated, clustered and absorbed into machine-learning pipelines. What was once an archive becomes training material.
This changes the politics of cultural sovereignty. The old anxiety was disappearance through neglect, decay or theft. The newer one is dispossession through computation. A songline, ceremonial image, weaving pattern, place-name register or oral history may be digitised in the name of preservation, yet become available for uses that neither custodians nor originating communities recognise as legitimate. The dispute is not only about ownership of files, but about authority over meaning.
That matters because AI systems do not merely store culture. They re-express it. They generate summaries, imitations, classifications and derivative outputs that circulate far from the institutions that first digitised a collection. A society can lose control of its cultural memory even while appearing to save it.
From conservation to computability
Digitisation preserves, but it also re-specifies culture in a format optimised for extraction. A recording becomes metadata plus waveform; a manuscript becomes text-recognition output; a carved object becomes a 3D model; a corpus of oral testimony becomes machine-translatable language data. Each conversion creates new possibilities for scholarship and access. It also turns context-rich cultural material into machine-readable units that can be recombined outside the norms that gave them meaning.
For years, heritage institutions treated openness as the natural endpoint of digitisation. The moral case was often strong: public collections should be publicly accessible. Yet open access and open science frameworks, including those promoted by UNESCO, were built largely around scientific and educational ideals, not around the asymmetries introduced by large-scale AI training. Once cultural material is broadly available in standardised digital formats, downstream control becomes difficult. The archive ceases to be only a memory institution and becomes part of the substrate of model development.
The legal categories inherited from the twentieth century fit this awkwardly. Copyright can protect some expressions, but not facts, styles, languages or many forms of inherited knowledge. Privacy law can shield personal data, but not necessarily collectively held cultural sensitivities. Heritage law often governs objects and sites more clearly than derivatives and embeddings. The result is an expanding zone of cultural material that is preserved in digital form but weakly governed when machines learn from it.
The problem is not only theft
Public debate often defaults to the language of theft. That is sometimes apt. Images have been copied from museum repositories into image datasets; transcribed texts have been ingested without meaningful consent; creators and communities have found fragments of their traditions echoed in generated output. But a sole focus on theft is too narrow, because many of the most significant harms occur even where the initial acquisition was nominally lawful.
A museum may hold rights in a photograph while lacking moral authority to license unrestricted computational reuse. A state archive may release ethnographic records under a permissive framework while the people described in them dispute the legitimacy of such openness. A language corpus may be assembled for revitalisation and then repurposed for systems that treat the language as just another optimisation problem. In such cases the injury lies not simply in unauthorised copying, but in the collapse of governance boundaries that once distinguished access, interpretation, ceremony, pedagogy and commerce.
Digitisation preserves, but it also re-specifies culture in a format optimised for extraction.
This is why cultural sovereignty is increasingly about protocol rather than possession alone. The key question is not merely who owns the file, but who sets the conditions under which the file can be seen, indexed, linked, trained on or transformed into synthetic output.
Digitisation preserves, but it also re-specifies culture in a format optimised for extraction.
Indigenous data governance points to a broader settlement
Some of the most sophisticated responses have emerged from Indigenous data governance rather than mainstream technology policy. The CARE Principles for Indigenous Data Governance, developed through the Global Indigenous Data Alliance, insist on collective benefit, authority to control, responsibility and ethics. These principles sit alongside, and sometimes in tension with, the more familiar FAIR framework for data stewardship. The point is not to reject technical interoperability. It is to insist that interoperability without legitimate authority can amount to a more efficient route to extraction.
Practical tools have followed. Traditional Knowledge and Biocultural Labels developed by Local Contexts are designed to carry provenance, permissions and community protocols with digital items. Platforms such as Mukurtu have shown how access rules can reflect cultural distinctions, including gendered, seasonal or kin-based restrictions that conventional content-management systems are poorly equipped to express. These initiatives matter not because they romanticise closure, but because they treat context as part of the data itself.
The lesson extends beyond Indigenous collections. Minority languages, diaspora archives, religious manuscripts and community memory projects all face versions of the same problem: a digital object can be technically accessible yet normatively misgoverned. Once AI systems flatten these distinctions, culture becomes available chiefly in the forms machines can process.
Metadata is where sovereignty quietly lives
Much argument about AI and culture focuses on datasets and outputs. Less attention goes to metadata, where many sovereignty decisions are actually made. Metadata determines discoverability, categorisation, attribution, rights statements and semantic relationships. It is the layer through which institutions encode what an item is, who may use it, and how it should be understood.
When metadata is sparse, generic or built around external taxonomies, communities lose representational control. Sacred material may be described as folklore, a politically contested territory as a neutral place-name, or an oral tradition as anonymous content. Such descriptors then travel. Search engines privilege them, aggregation platforms inherit them, and AI systems treat them as ground truth. Misdescription becomes machine-amplified authority.
Europeana’s rights statements and data-space practices have helped standardise access across institutions, a major achievement for public heritage. Yet standardisation has limits. The subtler challenge is whether shared infrastructures can represent differentiated cultural permissions rather than reducing everything to a small menu of open or closed statuses. In the AI era, simplistic metadata is not an administrative inconvenience. It is a sovereignty risk.
Why language preservation is entering a harder phase
The politics of endangered languages illustrates the shift. For decades the principal scarcity was recording material before speakers were lost. That remains urgent. But when language archives become training resources, a different scarcity appears: institutional capacity to govern reuse. A community may welcome speech technology that aids learning, transcription or translation. It may not welcome models that generate ceremonial phrases out of context, absorb rare linguistic forms into commercial systems, or produce inaccurate synthetic speech that then circulates as authentic.
Language is especially vulnerable because it often slips between legal regimes. Grammar and vocabulary are not owned in any simple way. Yet language carries relations, status hierarchies, humour, taboo and place-based knowledge. To treat it solely as neutral text or audio data is to detach utterance from social life. UNESCO has repeatedly stressed linguistic diversity as a condition of cultural pluralism. AI intensifies the issue because scale rewards dominant languages while making smaller ones newly visible mainly as low-resource inputs to be mined.
The paradox is sharp: computational tools may help revitalise a language while also standardising it around the forms most legible to models. Dialectal variation, ritual speech and non-written forms can be pushed to the margins. Preservation then occurs at the price of simplification.
The dispute is not only about ownership of files, but about authority over meaning.
The rise of inferential appropriation
Another underappreciated problem is inferential appropriation. AI systems do not need to reproduce a source verbatim in order to extract value from it. By training on many examples, models can infer style, structure, themes and relationships. This matters for cultural sovereignty because some knowledge systems are governed less by secrecy in an absolute sense than by rules about who may connect which meanings under which circumstances.
A set of publicly visible motifs may, in isolation, seem harmless. Combined computationally with catalogue notes, geospatial records and ethnographic texts, they may reveal patterns that communities regard as sensitive. A language model trained on oral histories may not quote them directly, but may still generate culturally loaded associations that originated in restricted contexts. The harm lies in the bypassing of social protocols through statistical synthesis.
The dispute is not only about ownership of files, but about authority over meaning.
This is one reason the conventional transparency demand of AI governance, while valuable, is insufficient on its own. Knowing that a model trained on heritage material does not settle whether it should have done so, whether certain inferential capacities ought to be constrained, or whether communities should have standing to contest downstream uses that were never envisaged at the point of digitisation.
European regulation helps, but only at the margins
Europe has moved further than most jurisdictions in regulating AI. The EU AI Act creates obligations around risk, transparency and certain prohibited practices. It is an important piece of public law, especially as a framework for accountability. Yet it is not, and was never designed to be, a complete settlement for cultural data rights. Its categories are oriented towards safety, fundamental rights and market governance. Cultural authority appears only indirectly.
The same is true of intellectual-property debates now running through courts and legislatures. They may clarify whether training on copyrighted works requires licences in specific contexts. But even a robust copyright settlement would leave much untouched: public-domain heritage, uncopyrightable traditional knowledge, collective claims over representation, and non-economic harms linked to dignity, ceremony or miscontextualisation. Cultural sovereignty cannot be reduced to a licensing dispute.
What European policy does contribute is a demonstration that public infrastructure and public rules can shape technical development. The unresolved question is whether states and memory institutions will build governance layers for cultural datasets that are as deliberate as those now being built for privacy, cybersecurity and AI risk.
Openness needs to be disaggregated
One conceptual error has been to bundle several different goods under the single banner of openness. There is openness for research, openness for education, openness for public memory, openness for creative reuse and openness for machine training. These are not identical. A manuscript available for close reading in a digital library is not thereby normatively available for ingestion into a model that can generate derivative text at scale.
Once these forms of openness are disaggregated, a more realistic policy landscape appears. Some materials may be open for viewing but not for bulk download. Others may be shareable for teaching but not for commercial training. Some may require community review before computational use. Such distinctions are administratively cumbersome, but that is not an argument against them. It is evidence that culture carries obligations which generic internet-era defaults failed to represent.
A society can lose control of its cultural memory even while appearing to save it.
The deeper point is political. Friction is often treated as a defect in digital systems. For cultural sovereignty, carefully designed friction can be a constitutional feature. It forces negotiation where extraction would otherwise be automatic.
What heritage institutions must now admit
Museums, libraries and archives occupy an uncomfortable position. They are custodians, but also processors; stewards, but also publishers of machine-readable content. Many entered digitisation with public-service motives and now find themselves upstream of AI supply chains they did not anticipate. It is no longer credible for institutions to claim neutrality once they know that digitised collections can be repurposed at scale.
This does not mean retreating into wholesale closure. It means accepting that preservation decisions are now also model-governance decisions. Collection policies, rights statements, APIs, watermarking practices, download settings and procurement rules all influence whether archives become resources for extraction or for accountable cultural exchange. Institutions that continue to behave as if publication is the endpoint of responsibility are operating with an outdated theory of the archive.
Some will worry that stronger controls threaten scholarship. There is a real trade-off. But scholarship has always operated within ethical constraints, especially around human subjects, sacred material and vulnerable communities. The novelty here is only that machine use has made old ethical questions impossible to ignore.
Towards computational non-alignment
A useful way to frame the challenge is as a form of computational non-alignment. Smaller cultures and communities need not reject digital technology or AI as such. Their interest is in avoiding automatic alignment with the categories, incentives and extractive habits of large-scale model development. That requires insisting that not everything of value should become frictionless training data, and that digitisation should not entail surrendering interpretive authority.
In practice, this means building layered governance around collections: provenance that remains attached in downstream systems, permissions that travel with files, differentiated access conditions, audit trails for bulk use, and institutional review processes that include source communities. None of this will be perfect. Leakage is inevitable; norms will be contested; public-domain doctrines will continue to constrain what can be restricted. Yet imperfect governance is not pointless governance. It changes expectations, allocates responsibility and creates grounds for challenge.
The larger significance is civilisational. AI is often described as a technology that learns from humanity. That phrase conceals a political choice about which humanity is rendered legible, under what terms, and for whose benefit. Cultural sovereignty, in this setting, is not nostalgia. It is a claim that memory should not be converted into machine resource without regard to the communities that produced, preserved and still live by it.
Preservation without surrender
The coming settlement will turn less on dramatic court cases than on mundane design choices in catalogues, repositories, contracts and standards bodies. Sovereignty is often won or lost in these low-visibility layers. If archives remain configured chiefly for maximum computability, then digitisation will continue to look like preservation from the front end and feel like dispossession from the back end.
The alternative is not to halt the digital recording of endangered cultures. It is to recognise that preservation now has two stages. The first is survival in digital form. The second, increasingly decisive one, is survival under terms that prevent memory from being stripped of context and fed into systems indifferent to origin, obligation and meaning.
A society can lose control of its cultural memory even while appearing to save it. The task of cultural sovereignty in the AI age is to prevent that contradiction from becoming the default architecture of preservation.
