The politics of cultural sovereignty has usually been framed in the language of rescue. A language is endangered. A ritual tradition is fading. A museum collection is fragile. A sound archive needs cataloguing before magnetic tape degrades. Digitisation, in that account, is a defensive act: convert memory into durable files, place them in trusted repositories, and widen access before loss becomes irreversible. That logic remains necessary. But by mid-2026 it is no longer sufficient.
Once digitised cultural materials enter the data supply chains of machine learning, they cease to be passive records. They become raw material for systems that classify, imitate, summarise, translate, recommend and generate. Digitisation preserves the past; model training manufactures possible futures from it. The sovereign question therefore changes. It is no longer only who holds the artefact, the scan or the recording. It is who decides how computational systems infer from that material, under what constraints, for whose benefit, and with what rights of refusal.
From preservation to projection
Cultural institutions spent two decades building digital collections on the premise that more access was broadly democratic. Europeana, national libraries, research repositories and museum digitisation programmes were shaped by that public-interest ethos. It was a reasonable response to scarcity. Yet generative AI has altered the meaning of access. A photograph of regalia, a corpus of transcribed oral literature, or a catalogue entry describing ceremonial use can now be incorporated into systems that generate stylistic pastiche, automated educational content, synthetic voices or predictive cultural taxonomies.
The archive is no longer merely consulted; it is mined for patterns and turned into a behavioural substrate. That matters because culture is not just information. It is authority, context, relationship and protocol. A ceremonial song does not become politically neutral because it has been converted into metadata. Nor does a language corpus become a free-standing technical resource simply because it is machine-readable.
The rise of inferential appropriation
Classic cultural appropriation involved visible borrowing: motifs, stories, garments, symbols. The AI era introduces a quieter form, which might be called inferential appropriation. Here, the extraction is not only of content but of latent structure. Machine systems learn association patterns, stylistic regularities, kinship terms, ritual sequences, place-name correlations or culturally specific modes of narrative compression. They can then reproduce approximations detached from the norms that governed their original use.
This is why the familiar legal distinction between protected expression and unprotected facts is proving inadequate. Communities may retain nominal ownership over digitised records while losing practical control over what can be inferred from them. A community can lose control without losing possession. That is the central asymmetry of machine learning applied to heritage data.
Digitisation preserves the past; model training manufactures possible futures from it.
Open by default meets sacred by context
Digitisation preserves the past; model training manufactures possible futures from it.
Many public-sector digitisation projects were built around open-access assumptions that made sense in scholarly and civic terms. But cultural materials are often governed by layered permissions: seasonal access, gender-specific knowledge, clan stewardship, restrictions on the deceased, obligations tied to place, or rights that are collective rather than individual. These conditions are difficult to encode in systems designed for binary states such as public or private, copyrighted or not, licensed or unlicensed.
The result is a governance mismatch. Institutions may publish low-resolution images, catalogue descriptions or transcriptions in good faith while omitting the contextual protocols that give those materials their legitimacy. Once copied into training datasets, those nuances disappear entirely. UNESCO’s work on intangible cultural heritage and AI ethics, together with the CARE Principles for Indigenous Data Governance, points toward an alternative logic: data about communities should be governed not merely by openness but by collective benefit, authority to control, responsibility and ethics.
Why cultural data rights are not solved by copyright
Copyright remains relevant, but it is a poor container for many cultural claims. It is time-limited, individualised and expression-focused. Cultural sovereignty claims are often perpetual, collective and relational. They may concern secret or sacred knowledge, genealogical information, customary law, or heritage whose misuse causes harm even when no formal intellectual-property right is infringed. WIPO’s work on traditional knowledge has long recognised this gap, and the adoption of the 2024 treaty on intellectual property, genetic resources and associated traditional knowledge was an important signal that conventional IP categories do not capture the full problem.
Still, the challenge in AI goes beyond unauthorised copying. The issue is derivation by statistical abstraction. A model may not reproduce a protected image verbatim yet may still emulate its ceremonial structure or generate misleading variants that circulate as plausible heritage. Existing doctrine struggles to address the politics of plausibility: who gets to decide whether a machine-generated retelling is an educational aid, a distortion, or a form of epistemic trespass.
The archive as infrastructure of machine judgement
It is tempting to regard heritage collections as special-interest material at the margins of the AI economy. In fact they are part of the infrastructure through which systems learn to sort the world. Language archives shape translation tools. Ethnographic records shape knowledge graphs. Digitised newspapers shape historical summarisation. Museum taxonomies shape object-recognition labels. Place-name databases shape mapping systems. Once cultural archives are piped into models, they begin to influence automated judgement far beyond the cultural sector.
The crucial political question is no longer whether culture is online, but how online culture becomes machine judgement. If the underlying record is skewed toward imperial collection practices, salvage anthropology, missionary transcription or state-centric cataloguing, those biases do not remain in the reading room. They are propagated at scale into systems that seem technically neutral.
What homogenising AI actually does
The standard fear is that AI will flatten culture by flooding the public sphere with generic output. That is true, but incomplete. Homogenisation occurs through at least three mechanisms. First, high-resource languages and heavily digitised traditions become disproportionately legible to models, which in turn makes them more searchable, more teachable and more likely to be recommended. Secondly, minority traditions are often rendered in majority-language metadata, forcing them into external descriptive categories. Thirdly, model optimisation rewards recurrence and statistical confidence, which can marginalise ambiguity, locality and exception.
In other words, homogenisation is not simply the spread of cultural sameness. It is the conversion of uneven archives into uneven probabilities. What survives in machine-mediated culture is not necessarily what matters most to communities, but what is most available, most standardised and least encumbered by context. That is a profound redistribution of symbolic power.
A community can lose control without losing possession.
Consent is the wrong unit of analysis
Policy discussions still lean heavily on consent: did the rights holder authorise digitisation, publication or reuse. But consent is often impossible to define cleanly in cultural settings. The relevant authority may be collective, disputed across generations, or dependent on ceremony and place. Some material was collected under coercive conditions decades ago. Some communities are now reconstructing governance over archives dispersed across national institutions. A checkbox model of permission does not meet this reality.
More importantly, consent addresses a transaction, whereas cultural sovereignty concerns an ongoing relationship. A recording may have been legitimately made for one purpose, then become inappropriate as training data for another. The governance question is therefore dynamic. It requires the capacity to revisit terms, impose downstream conditions and, in some cases, withdraw certain materials from computational use even if they remain preserved for memory or research.
A community can lose control without losing possession.
Towards computational stewardship
If simple openness is inadequate and copyright is partial, what replaces them. The most promising concept is stewardship rather than ownership alone. Stewardship asks how institutions, states and technical actors handle culturally sensitive data across its full lifecycle: collection, cataloguing, storage, access, interoperability, model training, deployment and redress. It turns archives from static repositories into governed infrastructures.
In practical terms, this means at least four shifts. First, digitised heritage should carry machine-readable provenance and protocol metadata that travel with the file rather than remain buried in human-readable notes. Secondly, access controls should distinguish between viewing, downloading, bulk scraping and training use. Thirdly, communities should have formal roles in data governance boards, not merely advisory positions after technical decisions are made. Fourthly, public institutions need audit trails to understand whether and how their collections have entered model-development pipelines.
The case for a right of contextual integrity
One way to name the missing protection is a right of contextual integrity for cultural data. The principle would be simple: information drawn from a cultural setting should not be repurposed in ways that violate the norms, relationships and expectations that gave it meaning. This is not a plea for stasis. Culture changes, borrows and improvises. But change within a field of accountable relationships is different from extraction into opaque systems optimised for scale.
Such a right would not need to ban all machine use. It would instead create rebuttable presumptions around sensitive classes of material: funerary records, sacred narratives, ceremonial imagery, restricted ecological knowledge, minority-language speech data, and community-generated metadata. It would also require impact assessment before training and deployment, much as AI risk frameworks already encourage organisations to examine downstream harms, representational bias and affected populations.
The crucial political question is no longer whether culture is online, but how online culture becomes machine judgement.
Small languages are a revealing test
Much is made of the promise that AI can revitalise low-resource languages through transcription, translation and educational tools. That promise is real, provided communities govern the terms. Small languages present the issue in its clearest form because every corpus is politically consequential. A few hundred hours of speech, a dictionary compiled by missionaries, or a set of school materials can become the basis for systems that effectively standardise spelling, pronunciation and acceptable usage.
This may be useful for education, but it also risks hardening one dialect, one orthography or one historical register as the machine-preferred norm. The resulting feedback loop is subtle. Users adapt to what tools recognise; institutions adopt what tools support; future corpora then reflect those choices. AI can therefore assist language survival while narrowing linguistic plurality within the language itself.
Public memory institutions need a new settlement
Libraries, archives and museums are not the villains of this story. They are among the few institutions with public mandates, conservation expertise and a tradition of balancing access with care. But they now operate inside a wider technical ecosystem that can detach their collections from their ethics. The settlement of the last era was digitise, catalogue, share. The settlement of the next one must be preserve, contextualise, condition and monitor.
That implies investment not only in servers and scanners but in governance design: protocol labels, differentiated licences, community review pathways, and legal capacities to negotiate the use of collections in AI contexts. It also implies intellectual humility. Some material should be digitised but not openly networked. Some should be preserved but not used for training. Some should be described only in terms approved by the relevant custodians. These are not failures of access. They are signs that public institutions have finally recognised culture as a living jurisdiction rather than a stockpile of content.
Sovereignty over the future tense
The deepest stake here is temporal. Heritage policy has often treated culture as inheritance from the past. AI turns heritage into a resource for constructing the future: what children encounter in educational systems, what search and recommendation engines foreground, what synthetic media imitates, what languages software accommodates, what historical analogies automated systems retrieve. When archives feed models, they do not only remember. They anticipate.
That is why cultural sovereignty in the age of homogenising AI cannot be reduced to preservation or representation. It concerns authority over projection. Who may turn a people’s memory into a template for generated texts, voices, classifications and norms. Who may decide which fragments become canonical through repetition by machines. And who has standing to contest those outcomes when they erode the protocols that made the material meaningful in the first place.
The crucial political question is no longer whether culture is online, but how online culture becomes machine judgement.
The answer will not come from one doctrine or one technology. It will come from recognising that cultural data rights are governance rights, and that governance must extend beyond the moment of digitisation into the inferential life of data. In that sense, the frontier of cultural sovereignty is not the archive alone. It is the model, the pipeline and the feedback loop through which memory is converted into social prediction. The communities that gave the archive its meaning should not arrive last in deciding what that prediction becomes.
