Missing Fields in Museum Catalog Records: What AI Can and Cannot Supply

A blank catalog field is not a historical answer

Museum catalog records often contain gaps left by earlier cataloging practices, incomplete accession paperwork or information that was never documented. A record may identify an object as a wooden chair without recording its maker, date, place of manufacture or original owner. An assistant asked to complete the record may produce plausible-looking information because these fields commonly appear together in museum databases. Plausibility does not make an answer suitable for permanent collection documentation.

The first step is to classify the gap. A field may be missing because information exists elsewhere, because it was never recorded, because the available evidence is inconclusive or because the field does not apply to the object. These situations require different responses. A missing dimensions field may be resolved by a new measurement. An unknown maker may require documentary research and could remain unknown indefinitely.

• Missing: the value has not been entered.
• Unknown: available evidence does not establish the value.
• Not applicable: the field does not describe this object.
• Unverified: a possible value exists but has not been confirmed.
• Conflicting: different sources provide incompatible information.

Identify the task before asking an assistant to help

AI can contribute to several different cataloging tasks, but the instructions must distinguish them. Extracting a date already present in a donor letter is different from proposing a date based on an object's appearance. Standardizing 'wooden' to an approved material term is different from identifying an unknown timber species. Summarizing a description is different from adding historical context. Treating all four activities as record completion makes it difficult to identify where new claims entered the catalog.

A useful instruction asks the assistant to classify each field as present, missing, uncertain, conflicting or not applicable. It can then identify which fields contain information that might be extracted from the supplied documents and which require staff research or physical examination. Any proposed value should include the source passage that supports it. If no supporting passage exists, the value belongs in a research question list, not the approved record.

The boundary between normalization and invention

Normalization changes the form of information without changing its evidential meaning. For example, an existing date written as 'circa 1900' can be entered into a structured date field using the institution's approved approximate-date convention. A description containing 'brass candlestick' can be mapped to an authorized object type if that terminology is part of the local cataloging system. The original wording should remain available when it carries useful historical or documentary information.

Invention occurs when a suggestion introduces information that the source does not support. A wooden chair with no maker recorded cannot automatically be assigned to a local furniture workshop because that workshop was active in the period. A photograph with no location should not be assigned to the town where the donor lived. Such suggestions may be reasonable research leads, but they must be identified as hypotheses and supported by separate evidence before they enter authoritative fields.

• Normalization: changing a documented value into an approved format.
• Extraction: transferring information from a supplied source.
• Interpretation: explaining what existing evidence may indicate.
• Research suggestion: identifying a question or possible lead.
• Invention: presenting unsupported information as a catalog fact.

A worked example of an incomplete object record

Consider a hypothetical record for a wooden chair. The catalog contains an accession number, a broad description, a measurement and the name of the person who donated it. The maker, date of manufacture and place of manufacture are blank. An assistant can organize the supplied information into the appropriate fields and flag the three gaps. It can also suggest searching accession correspondence for a maker's name or comparing the object with documented examples.

The assistant should not supply a named furniture maker, assign a precise manufacturing year or turn the donor's address into the chair's place of manufacture. The corrected record would retain the documented description and measurement, preserve the donor information in its proper field and mark the remaining values as unknown or unrecorded according to local policy. The resulting record may look less complete, but it represents the available evidence more accurately.

Controlled vocabularies require local rules

An assistant can help identify candidate terms for object types, materials, subjects and places. This is useful when records contain inconsistent natural-language descriptions or when a small museum is gradually adopting a controlled vocabulary. However, a suggested term is not automatically an authorized catalog value. The institution should identify the vocabulary it uses, the relevant version and the level of specificity appropriate to its collection.

A description such as 'small wooden box for keeping sewing items' may suggest a more standardized object term. Whether that term should be a sewing box, container or another category depends on the object's documented function, institutional terminology and available evidence. A broad but defensible term is preferable to an exact classification based on appearance alone. Staff should also retain original descriptions when they contain information that a standardized term would remove.

Completeness checks and public interpretation are different tasks

A catalog completeness check identifies empty fields, inconsistent formats and possible missing relationships. A public-facing description or museum label writing task selects information for visitors. The two tasks can share approved facts, but they should not be treated as interchangeable. A record may need an internal note explaining that the maker is unknown, while a short public label may simply identify the object, its material and its documented use.

An object label template should draw from approved catalog information and clearly distinguish established facts from interpretation. An assistant can make a sentence shorter or adjust the exhibition label reading level, but it should not remove a qualification that changes the meaning. For example, 'attributed to a local workshop' should not become 'made by a local workshop' simply because the shorter sentence reads more smoothly. A public text should never be used to repair an incomplete source record.

Make uncertainty visible in the catalog

Small institutions benefit from a consistent method for recording the status of proposed values. A field can be marked as unknown, unverified or under research using the institution's established conventions. A separate note can identify the documents consulted, the date of the research and the reason a proposed value was not accepted. If the collection system supports source citations or field-level provenance, use those facilities to preserve the relationship between a value and its evidence.

AI assistance should also be documented accurately. A suitable note might read: 'AI writing assistance used to identify incomplete fields and propose standardized wording. Suggested values were evaluated against the cited collection documentation; unsupported proposals were excluded from the approved record.' Adjust the statement if the tool performed extraction, summarization or another distinct task. The note should describe what happened, not imply that an automated suggestion was independently established.

Prioritize the backlog by risk and research value

Not every incomplete record requires the same amount of work. A missing preferred object name may prevent basic discovery. An uncertain donor identity may affect provenance research. A missing secondary description may have less immediate impact. A simple priority system can help a small team decide where to spend limited cataloging time. Consider the significance of the missing field, the availability of supporting documents and the consequences of entering an incorrect value.

Begin with a sample of ten incomplete records and group their gaps into identification, description, attribution, provenance, dates and administrative information. Identify which gaps can be resolved from existing documentation, which require physical examination and which should remain explicitly unknown. This creates a manageable research list and gives the institution a practical basis for deciding where AI-assisted extraction and normalization can contribute.

Questions readers ask