Keeping Curatorial Authority

The label has authors, even when the answer varies

An object’s acquisition history includes a gap. A confident digital voice fills it with a plausible story. The answer is easy to understand, but the museum has lost something essential: the visible boundary of its knowledge.

Curatorial authority in an adaptive system does not mean that one voice must settle every question. It means that responsibility for claims, permissions and corrections remains identifiable. Curators, educators, conservators, registrars, access specialists and source communities may each hold knowledge that changes what can responsibly be said.

Establish who can approve a source, who can revise an interpretation and who can pause the service. A vendor may operate the software, but it should not silently become the institution’s editorial decision-maker. Put these responsibilities into a working process before an exhibition opens, not into a policy document that nobody can use during a live incident.

Make the source boundary explicit

A collection catalogue, a conservation report and an old exhibition essay may disagree. Material can also contain personal information, restricted cultural knowledge or language that the museum no longer endorses. Inclusion in an internal database does not automatically make a document appropriate for a public-facing answer.

Build a reviewed source set with records of ownership, permissions, review status and intended audience. Preserve the distinction between a historical quotation and the museum’s current account. Where a source community has conditions on access or reuse, those conditions need technical and editorial enforcement, not just a note in a spreadsheet.

Give disagreement an honest form

Some questions have several supported interpretations. The system should not collapse those into a falsely neutral summary or choose whichever account sounds most decisive. Explain whose interpretation is being presented, what supports it and where disagreement remains. Do not ask an algorithm to resolve an ethical dispute that the institution has not addressed.

The ICOM Code of Ethics for Museums offers a professional reference for museum responsibilities. The Museums Association’s ethics resources provide another useful point of reflection. Neither turns a contested acquisition into a simple software problem. The relevant people still need to be involved, and the visitor should be able to see the limits of the account.

A source is not a decoration at the end of an answer. It is a limit on what the answer may claim.

Editorial principle — Wafflerocketznd reading room

Test for pressure, ambiguity and unsupported certainty

Ordinary visitor questions are only part of testing. Include prompts that demand an invented quotation, ask the system to ignore its sources or introduce a false premise. Test ambiguous object names and follow-up questions that could refer to more than one display. A technically correct statement about the wrong object is still an interpretive failure.

Prepare a documented set of expected behaviours: answer from an approved source, explain uncertainty, ask for clarification or decline. Keep examples of harmful failures and test them again after model, retrieval or content updates. The NIST AI Risk Management Framework can help structure risk conversations, but it does not certify an installation’s accuracy or replace local testing.

Retrieved documents also need scrutiny. An embedded instruction in a document should not be allowed to override the system’s rules or expose unrelated material. Restrict access by design, separate internal from public sources and treat model instructions as one layer of defence rather than a security boundary on their own.

Make correction practical

Provide a straightforward way to report a problematic answer without demanding unnecessary personal information. A useful report includes the object, the question, the answer and enough context to reproduce the problem. Explain whether a report is sent immediately or prepared locally, and do not quietly retain a full conversation when a smaller record will do.

Assign an owner and a route for urgent issues. Teams need to be able to withdraw one answer, remove a source or switch to approved static content while investigating. Keep version history so a correction can be traced to the material visitors actually encountered. If the problem is significant, explain the change publicly where appropriate.

Plan for life after launch

Budgets often favour a striking opening and underfund maintenance. Yet collection knowledge changes, links break and model behaviour can shift. Schedule review, preserve an exportable source set and define how the experience will end or migrate if a supplier relationship changes. A stable label is a better outcome than an abandoned interactive service.

Curatorial control is not resistance to experimentation. It is the condition that makes experimentation responsible. Visitors can explore new questions while the museum remains answerable for the materials, judgements and limits behind each response.