Adapting the Interpretive Layer

A different beginning, not a different object

Imagine a worn bowl whose surface carries a small repair. An adult asks about its material. A child notices the crack. Someone else asks whether its presence in the collection is disputed. Each question can support a serious encounter.

An adaptive system changes the interpretive route in response to a visitor’s needs or choices. It might select an approved introduction, offer a definition, retrieve a relevant source or generate a response from a bounded body of material. These are different mechanisms, with different risks. Calling all of them “AI” conceals decisions that a museum needs to make explicitly.

Begin by distinguishing stable evidence from variable presentation. An accession number, a documented acquisition date and an attributed maker belong to the collection record. Tone, sequence, reading depth and the form of a prompt belong to an interpretive layer. Even that distinction needs care: a record may contain outdated terminology or contested claims. Adaptation must not turn an inherited assertion into unquestionable fact.

Offer choices before making inferences

Ask visitors what they want to explore. “How was it made?”, “Who used it?” and “Why is it here?” provide understandable entry points. A request for a short explanation is useful evidence about the current task. It is not evidence that a person has limited knowledge, a particular disability or a fixed learning style.

Keep choices reversible. A visitor should be able to move from a brief introduction to a detailed account without starting again. The interface can remember a selection within the encounter, but it should not silently carry assumptions from one person to the next on a shared screen. An obvious reset is part of the interpretation, not merely a technical housekeeping feature.

Build an answer from accountable material

For many exhibitions, reviewed branching content will be enough. Educators can prepare several introductions, follow-up questions and explanations of specialist terms. This approach is less flexible than open generation, but its limits are legible. Every route can be checked before opening and maintained as collection knowledge changes.

Where free-text questions add genuine value, retrieval can bring relevant passages into the response process. Retrieval-augmented generation does not guarantee accuracy: a system can select the wrong passage, omit a qualification or produce a claim unsupported by what it retrieved. Sources should therefore be attached at the level of the answer, not merely listed on a general “about” screen.

Collection models such as CIDOC CRM can help teams distinguish people, objects, events and relationships. They do not decide what an interpretation should say. Similarly, the International Image Interoperability Framework can support access to images and annotations without resolving questions of ownership, cultural permission or interpretive balance. A standard is infrastructure, not a substitute for editorial judgement.

Adapt the invitation; keep the grounds of the claim available.

Editorial principle — Wafflerocketznd reading room

Make uncertainty part of the experience

A museum answer may need to distinguish observation, attribution, inference and disagreement. “The surface shows…” is different from “Researchers suggest…”. A concise response can retain those distinctions. Simplification should remove unnecessary difficulty, not erase the uncertainty on which an honest interpretation depends.

Prepare useful responses to questions outside the reviewed material. “The collection record does not establish that” is better than a plausible biography. Offer a relevant source, a related question or a route to a member of staff. Avoid making refusal feel like visitor error. The system’s boundary belongs to the system.

Test the conversation, not just the prose

Evaluate complete encounters with educators and visitors. Can someone identify which object is being discussed? Can they return to the standard label? Does a follow-up preserve the qualification in the first answer? Do questions about acquisition history receive the same care as questions about technique?

Keep a small set of difficult test questions and repeat them after content or model changes. Record the approved answer, the supporting source and the reason an alternative would be misleading. An answer that sounds warm but changes a date is not an improvement. A shorter exchange that helps someone look again may be.

Success is not the longest session or the largest number of generated words. It is a visitor leaving with a clearer relationship to the object, a sense of what remains uncertain and the freedom to continue the conversation elsewhere.