How AI can support catalogue raisonné research
Artificial intelligence can assist with narrow, inspectable parts of catalogue raisonné work, including text recognition, source discovery, exploratory translation, and data preparation. Its usefulness depends on whether scholars can check its output and retain responsibility for the historical conclusions that follow.
A catalogue raisonné brings together many kinds of evidence. Artwork records may draw on archival documents, exhibition catalogues, correspondence, photographs, conservation files, sale records, and earlier scholarship. Researchers must locate relevant material, identify the people and works it concerns, evaluate its reliability, and record what it supports. Some of this work can be assisted by artificial intelligence. But “AI” does not name a single method or level of authority. Optical character recognition, machine translation, semantic search, automated data extraction, and image comparison perform different tasks and produce different kinds of results.
A useful evaluation therefore begins with the research task rather than with a general promise that AI will make scholarship faster. The team needs to know which material the system will examine, what it should return, how the result will be checked, and what could happen if it is wrong.
Paul Gauguin, Nature morte à la coupe et aux fruits, c. 1892, oil on canvas, 20.3 x 31.4 cm, Virginia Museum of Fine Arts. Image courtesy of the Virginia Museum of Fine Arts.
Turn page images into searchable text
Catalogue raisonné projects often work with publications and archival sources that have been digitized as images. A scan allows someone to view the page, but the words within it may remain unavailable to a search tool.
Optical character recognition, usually abbreviated to OCR, identifies printed characters in page images and produces text that a computer can search. This can make it easier to locate an artist’s name in an exhibition catalogue, search a dealer’s stock book, or find repeated references to an artwork across a lengthy publication. The U.S. National Archives distinguishes OCR from human transcription and cautions that its automated text is not always accurate, even though it can improve discovery.
OCR does not create a definitive transcription. Decorative typefaces, damaged paper, narrow margins, multiple columns, handwritten notes, and older spelling can all affect the result. Proper names and numbers are especially consequential in catalogue research because a single incorrect character may change an artist, collector, date, or inventory number.
The page image should therefore remain available beside the recognized text. Researchers can use OCR to find possible evidence more quickly, then verify the relevant wording against the original scan. Corrections to the recognized text should also remain distinguishable from the historical document itself.
Find relevant passages without relying on exact keywords
Traditional keyword search works best when a researcher already knows which term appears in the source. Historical materials rarely cooperate so neatly. An artwork may have several titles, a place may be identified in different languages, and an individual may appear under a married name, initials, or an inconsistent spelling.
Semantic search compares a query with passages that have related meaning or context. It can therefore surface material that does not repeat the researcher’s exact words. This may be useful for exploring variant descriptions, locating passages about a recurring subject, or identifying sources that deserve closer reading.
Some systems combine this retrieval process with a language model that composes an answer. The method is commonly called retrieval-augmented generation. The generated answer and the retrieved evidence should be evaluated separately. A clear sentence may exceed what the selected passage supports, while a relevant passage may still require considerable interpretation.
For scholarly research, the most useful interface may be one that returns the passage, page number, publication details, and enough surrounding text to judge its relevance. A summary can help with orientation, but it should not replace the evidence or conceal disagreement among sources.
Use machine translation for discovery, not final interpretation
Catalogue raisonné research frequently crosses languages. Correspondence, exhibition reviews, sale catalogues, inscriptions, and earlier scholarship may have been produced in several linguistic and historical settings.
Machine translation can help a researcher obtain a provisional sense of a document, test queries in another language, or decide which material merits specialist attention. Cross-language retrieval may also reveal a source that an English-language search would miss.
These uses remain exploratory. Historical terminology, idiom, abbreviations, obsolete words, and culturally specific references can be difficult to translate. Names and titles may also be translated inconsistently. The Europeana Translate project evaluated its language tools with linguists and cultural heritage specialists, demonstrating why the field and language of the source matter when assessing a translation.
A quotation used in a catalogue entry or scholarly essay should still be checked by someone with the appropriate linguistic and subject knowledge. The researcher should retain the original passage and identify when a translation was produced or assisted computationally.
Propose information for a researcher to review
Automated extraction can propose names, dates, places, organizations, artwork titles, and other entities within a text. It may also suggest that two differently written names refer to the same person or that a date belongs to an exhibition rather than to an artwork.
These proposals can accelerate review, particularly when a team is processing a large body of repetitive material. They can also introduce errors that are difficult to notice once they have been incorporated into a database.
Catalogue information is not simply found and copied. Researchers decide whether a person acted as an owner, dealer, lender, photographer, or author. They determine whether two titles refer to the same artwork, whether a date is documented or inferred, and whether an attribution is accepted, questioned, or historical. A system can propose a value, but it cannot supply the missing scholarly justification.
Suggested information should therefore remain separate from approved catalogue data until a researcher has reviewed it. The record should preserve its source, the person responsible for accepting or correcting it, and any uncertainty attached to the decision. The Linked Art model for qualified assertions demonstrates how an attribution such as “possibly by” can retain its qualification, author, date, and source. Controlled vocabularies and stable identifiers can help align names and concepts, but they do not remove the need to interpret the evidence.
Use image analysis to identify material for comparison
Computer vision can group and rank images according to measured visual similarity. Within a catalogue raisonné project, this may help researchers notice related compositions, identify possible duplicate photographs, detect differently cropped reproductions, or assemble images for comparative examination.
Such results are leads rather than determinations. Two works may resemble one another because they share a subject, source image, workshop practice, or later reproduction. Conversely, photographs of the same artwork can appear substantially different because of lighting, cropping, restoration, color correction, or the quality of the photographic file.
Visual similarity cannot independently establish authorship, authenticity, chronology, or provenance. Those conclusions require historical evidence, material examination, specialist judgment, and a clear account of how the comparison group was formed. Image analysis is most useful when it makes potentially relevant material easier to inspect without deciding in advance what that material proves.
Prepare research data for computational use
The available corpus determines what an AI-assisted process can recover. If records contain inconsistent names, missing sources, unclear rights, or unmarked uncertainty, a computational system may reproduce those weaknesses at greater scale.
Good preparation includes stable identifiers, consistent field structures, documented sources, rights information, and explicit relationships among artworks, people, places, exhibitions, publications, and archival records. Approximate dates and disputed claims should be represented as such rather than converted into apparently precise facts for the convenience of a system.
The FAIR Guiding Principles emphasize making research materials findable, accessible under stated conditions, interoperable, and reusable. They also call for detailed provenance and clear reuse terms. These qualities help both people and machines understand what data describes and how it may be used.
Availability does not automatically make information appropriate for every computational purpose. Copyright, privacy, contractual restrictions, and cultural protocols still apply. The CARE Principles for Indigenous Data Governance additionally emphasize collective benefit, authority to control, responsibility, and ethics. A research team must consider who has authority over culturally sensitive data and who may be affected by its use.
Test the method with the project’s own material
An impressive demonstration may reveal little about how a method will perform on the sources in a particular catalogue raisonné. Evaluation should include the project’s routine material as well as its most difficult examples.
For OCR, that may mean faded print, unfamiliar typefaces, handwritten additions, and complex page layouts. Translation tests should include historical phrasing, ambiguous names, and passages whose meaning changes with context. Search tests should include questions for which the corpus contains incomplete, conflicting, or no evidence. Image comparisons should account for reproductions of different quality and works that resemble one another for reasons unrelated to authorship.
The team should record omissions, false matches, incorrect transcriptions, unsupported inferences, lost qualifications, and failures to indicate insufficient evidence. It should also measure the time required for a knowledgeable person to review the output. A method that creates a result quickly may provide little efficiency if confirming it takes longer than completing the original task.
Keep scholarly responsibility visible
Human review is often invoked as a general safeguard, but its meaning must be defined. Someone needs the time, evidence, expertise, and authority to reject or correct what the system proposes. The project should identify that responsibility before AI-assisted output enters an established workflow.
The NIST AI Risk Management Framework recommends defining the particular task supported by an AI system, assigning responsibilities for oversight, and evaluating performance in its intended context. For catalogue raisonné research, the required level of review should correspond to the consequence of the decision. A provisional search lead does not carry the same weight as a published attribution, provenance claim, or decision about catalogue inclusion.
Documentation should distinguish source material, machine-produced suggestions, and approved scholarly conclusions. This separation allows later researchers to understand how a record developed and to reconsider it when new evidence emerges.
Decide whether AI is appropriate for the task
Before adopting an AI-assisted method, a catalogue team should be able to describe the research difficulty it is meant to address and where that difficulty occurs in the existing process. The team should define the evidence an acceptable output must provide, the kinds of error that would make it unusable, and the person responsible for reviewing it.
It should also compare the proposed method with simpler alternatives. Better metadata, a revised naming convention, conventional full-text search, or additional staff training may sometimes address the problem more effectively. Using AI is not itself a measure of innovation or scholarly value.
Artificial intelligence can support catalogue raisonné research when the task is appropriately limited, the source remains available, uncertainty remains visible, and the output can be challenged. It may shorten the path to a relevant page, suggest material for comparison, or help a team prepare records for closer examination. The interpretation of that material, and responsibility for what enters the scholarly record, remain with the researchers.
