Cultural Data Is Not a Free Stockpile: AI Must Learn to Knock First
Original Chinese title: 文化資料不是免費素材庫：AI請先學會敲門
AI needs data, but cultural data is not ownerless material. Once languages, stories, place names, ecological knowledge and historical memory enter the digital environment, who may collect, train models on them and exploit them commercially has become an important public issue in the age of AI.
山海資料庫筆記
Data Is Not a Neutral Resource
The AI industry likes to refer to data as fuel. This metaphor is convenient but also dangerous. Fuel is characterized by being mineable, convertible, and consumable; but cultural knowledge is not a mineral deposit nor ownerless energy.
A song, an oral history segment, a place name, or a medicinal plant classification may all carry family memory, land relations, ritual order, and care responsibilities. When these contents are stripped of their context, they can still be processed by computers, but they may no longer be called knowledge.
AI is best at slicing the world into data points. Yet much cultural knowledge is not a set of data points; it is a network of relations. Who may speak, when to speak, where to speak, to whom, and where speech must stop—these limits are not backwardness but governance.
From Open Data to Data Sovereignty
Open data has driven public information sharing and made important contributions to public governance. But not all data is suitable for unrestricted openness.
Some knowledge carries cultural permissions; some knowledge carries ritual restrictions; some knowledge is suited only for internal community education; some knowledge should not be digitized at all. These are not issues of data quality but of knowledge sovereignty.
Data sovereignty asserts that the collection, preservation, classification, use, sharing, and commercialization of data must respect community governance and cultural ethics. In the age of AI, this concept is more urgent than ever because once a model is trained, the influence of the data may no longer remain in the original files but will seep into generated content, search results, textbooks, policy documents, and commercial products.
Before Model Training, Ask Who Can Consent
The situations most needing discussion are when interviews, images, language texts, place-name knowledge, and traditional medical data are organized into datasets and become part of external model training.
The issue is not that AI cannot assist in preserving languages or culture; it is who has the right to decide on preservation methods. If researchers, platforms, or enterprises treat communities merely as sources of data and then claim cultural preservation through their published results, they are simply swapping old inequalities for a digital interface.
Truly responsible AI projects must begin with community consent rather than data scraping. They must ask: which data may be used? Which data may not be used? Which data is limited to education only? Which data cannot be commercialized? Which content can be withdrawn? Who is responsible for reviewing outputs? When errors occur, who bears responsibility?
CARE Principles Offer a Lesson for AI
In international discussions on data governance, the CARE principles provide important direction: Collective Benefit, Authority to Control, Responsibility, and Ethics. They remind us that data must not only be findable, accessible, interoperable, and reusable; it must also benefit communities and respect community control over their data.
This is crucial for AI. Mainstream data governance often focuses on efficiency, format, licensing, and reuse, but cultural data governance adds more fundamental questions: Do these data align with community values? Do they harm cultural relations? Do they grant external institutions excessive power? Do communities only provide content without deciding its use?
If AI learns only to grab data without learning to respect the people behind it, it is not intelligence; it is merely fast, rude copying.
Conclusion
Cultural data is neither a free stockpile nor waiting to be saved by AI. It is a living knowledge system with its own classifications, taboos, permissions, and responsibilities.
AI can assist in preservation, translation, education, and dissemination, but only if it learns to knock first. Without community consent, without data sovereignty, and without cultural review, so-called AI cultural preservation may simply be another form of more efficient inequality.
Before cultural knowledge enters a dataset, it must enter community governance. This is not formalism; it is the bottom line.
AI use and content-safety disclosure
This article was generated as a draft by Yuan Media AI's daily article factory, then published after being manually reviewed and approved by Aciang Iku-Silan. When specific community knowledge is involved, it should still return to community authorization and review by the knowledge holders.