Museums Are Not Data Mines: When AI Enters the Display Cases, Who Has the Right to Speak for Artifacts?

Museums fear not silence but being talked about too loudly by new technologies. AI guides, voice Q&A, collection databases, immersive exhibition spaces—these appear to open display cases so artifacts can finally speak to audiences. But the sharper question is: who is speaking? The artifact itself, the curatorial team, the donors, the communities belonging to the collected objects, or a model that confidently answers once it has swallowed data? When museums place AI into exhibition and education systems, what is truly being tested is not whether technology can dazzle, but whether museums have the capacity to admit: data are not free mines, culture is not a content pack, and interpretive rights do not belong to whoever scans first.

Objects in Display Cases Are Not Just Objects

The International Council of Museums adopted its new definition in 2022, positioning museums as serving society. It emphasizes research, collection, preservation, interpretation, and display of tangible and intangible heritage, while demanding openness, accessibility, inclusivity, diversity, sustainability, and community participation. The key to this definition is not its elegance but its reminder: curation does not end when objects are moved into storage; exhibition does not finish when objects are placed under lights; interpretation is not simply writing official-sounding labels. When AI intervenes, these old questions are amplified. Past mislabeling may have hidden in drawer after drawer of the archive; today it will be repeated by chatbots, turned into intimate voice guides, and cut into short-video knowledge summaries. Errors thus become not just mistakes but expandable authority.

The first step of museum data ethics is not to ban AI but to stop pretending that “data equals facts.” Collection fields typically contain object names, dates, materials, provenance, ethnic groups, collectors, locations, and stories. These fields appear objective but are often traces of knowledge power from a particular historical moment: how did colonial investigators name things? Did donors have the right to transfer ownership? Did Indigenous communities consent to public disclosure? Are certain objects tied to ritual, taboo, or family inheritance? If AI reads only the fields without hearing the silence behind them, it translates that silence into usable data. This is convenient and dangerous—convenient like an automatic coffee machine, dangerous like turning ancestral spirit tablets into product reviews.

From “Owning Objects” to “Negotiated Interpretation”

Recent museum reforms go beyond repatriation; they address co-governance, provenance research, community participation, and digital access. AI makes these issues more urgent because digitization means artifacts exist not only in display cases but also as images, 3D models, metadata, voice scripts, educational materials, training data, and API responses. The object’s body may remain in the museum, but its digital avatars travel everywhere. Once these avatars enter a model, they can be remixed, regenerated, and resold. Traditional collection authorization forms often cannot keep pace with this fluidity.

Thus museums must split “can it be made public” into finer layers: research use, exhibition, commercial use, model training, generative system summarization, audience download, cross-border transfer, children’s education, co-review with specific communities. This is not bureaucratic nitpicking but the minimum cultural responsibility. If an item relates to Indigenous knowledge, religious practice, funeral rites, or family memory, museums cannot simply use “already publicly exhibited” as a pass. Public exhibition does not equal unlimited copying; seeing does not equal owning; learning does not equal extracting.

FAIR Is Good, but CARE Reminds Us Where People Are

Scientific data management often cites FAIR: Findable, Accessible, Interoperable, Reusable. This matters for research infrastructure—otherwise data are like screws hidden in a warehouse corner, useful yet unfindable. For Indigenous Peoples and other historically collected communities, FAIR alone is insufficient. CARE principles remind us that data governance must include collective benefit, authority to control, responsibility, and ethics. In short, data should not only be findable by machines; the people it concerns must decide how it is used, who benefits from its value creation, and who bears responsibility when harm occurs.

This is especially critical for AI museums. Many institutions claim: “We are doing education, not commerce.” Yet education can produce a single version of history; philanthropy can package unequal data access; tech demonstrations can bring communities in to take photos without letting them participate in decision-making. True co-curation does not happen only at the opening ceremony when clan members give speeches; it must be embedded in collection fields, image authorization, voice scripts, model answer boundaries, error-correction workflows, and revenue arrangements. Without these systems, even beautiful AI guides are just old colonial habits displayed in a digital showcase.

AI Guides Must Have the Ability to Say “I Don’t Know”

A good museum chatbot should not act like an exam-perfect student but like a trained curator: knowing what can be said, what must be verified, what needs controversy labeling, and what cannot speak for others. AI systems must answer queries such as “insufficient data currently,” “source of this object still pending verification,” “this interpretation requires confirmation by the relevant community,” or “this belongs to knowledge not suitable for public generation.” In the world of generative AI, admitting ignorance is not a functional defect but an ethical feature. Conversely, the most terrifying systems are those that never err but always speak uncertainly as truth in fluent tones.

Therefore museums introducing AI must maintain at least three lists: a data inventory (sources, authorization, update dates, restrictions); a risk register (sensitive cultural content, personal data, unpublished research, controversial provenance, repatriation requests, commercial reuse); and an accountability ledger (who reviews, who updates, who receives complaints, who can remove items, who explains to communities). Without these three lists, AI guides are merely talking shredders in the exhibition hall—they may not cause immediate disaster but will slowly cut cultural context into consumable paper flowers.

Conclusion: Let Technology Learn to Lower Its Voice

AI can do many things for museums: enable audiences to understand exhibits in different languages, let children engage with history interactively, help researchers organize vast data, and bring remote regions access to cultural resources. All these are worthwhile. But if museums forget data ethics, AI will turn “education” into “extraction,” “openness” into “loss of control,” and “diversity” into database classification tags. Truly progressive museums do not make every artifact a talking character; they let technology lower its voice so objects, communities, and history can speak slowly.

A Pragmatic Version for Taiwan Museums

For Taiwan museums this issue cannot stop at international slogans. Many collections involve Austronesian peoples, Han immigrants, colonial governance, religious practice, local industries, and contemporary ethnic politics; the same object may carry different meanings in different communities. If AI systems provide only a single narrative path, audiences will think history has just one voice. A more responsible approach is to present “multiple interpretations”: what museum research says, what local communities say, where controversy remains, and which parts are not publicly disclosed due to taboos or insufficient consent. Such interfaces may not be the most dazzling but best approximate public cultural honesty.

More pragmatically, when museums procure AI systems, tender documents should require suppliers to deliver data source tables, authorization records, sensitive-data exclusion mechanisms, human review workflows, error-reporting channels, model update logs, and removal procedures. Otherwise institutions buy not cultural technology but a very talkative yet unaccountable megaphone. Museums’ professional value lies in slowing down technology, making rights visible, and ensuring stories are not compressed into algorithmic tones.

Another often-overlooked issue is language. Multilingual guides that merely translate Chinese labels into English, Japanese, or Indigenous vocabulary without semantic correction by relevant language communities risk spreading errors internationally. AI translation must preserve review records because cultural concepts cannot always be matched word-for-word; some terms should not be defined by external systems.

This also reflects the public governance discussed: technology can accelerate but should never replace judgment, responsibility, and human context.

Further Reading and Sources

• International Council of Museums (ICOM), “Museum Definition,” 2022-08-24, Source. Verification considerations: confirm core concepts in the 2022 museum definition—“serving society,” “tangible and intangible heritage,” “diversity, sustainability, community participation”—to avoid reducing the definition to an exhibition institution.
• UNESCO, “Recommendation on the Ethics of Artificial Intelligence,” 2021; web update 2024-09-26, Source. Verification considerations: check AI ethics principles—human rights, transparency, fairness, human oversight, environmental and social impacts—as governance background for museum AI guides.
• Global Indigenous Data Alliance, “CARE Principles for Indigenous Data Governance,” 2019/2020, Source. Verification considerations: verify the four CARE principles—Collective Benefit, Authority to Control, Responsibility, Ethics—to avoid misrepresenting Indigenous data governance as merely a privacy issue.
• Wilkinson, M. D. et al., “The FAIR Guiding Principles for scientific data management and stewardship,” Scientific Data, 2016, Source. Verification considerations: verify FAIR components—Findable, Accessible, Interoperable, Reusable—and explicitly state the complementary relationship between FAIR and CARE in the text.