A language model (an LLM) no longer just reads a document: it identifies it and fills the record with what it knows. From a title page it pulls the author, the subjects and the date; from a title, the ISBN and dozens of fields the source never carried. That ability to enrich is common today, and it is precisely the value of using AI to catalog. AI can fill a record, and easily so. For an institution with a collection, the question is whether that record can enter a permanent catalog and stay there. That is a different bar, and it is where generic enrichment falls short.
The difference is between plausible and reliable. A generic model produces a complete, plausible record: the fields are filled and they sound right. A catalog also needs to know where each value comes from. A plausible but wrong ISBN, a subject that doesn’t apply, or a date inferred from a mistaken identification all stay in the record looking correct. An empty field gets completed later; one filled with a false value contaminates without leaving a trace, and it is expensive to find.
Disciplined cataloging enriches just as much, but it knows what backs each value. At several points it parts ways with generic enrichment.
Knowing where each value comes from. What can be verified is verified and attributed. Subjects and names are checked against authorities and controlled vocabularies, and the record notes which source validated each one. What comes from the model’s knowledge with no external source to back it carries no such note, and that absence is itself a signal of how much certainty to give it. A generic model delivers everything at the same level, without distinguishing what it checked from what it assumed.
Computing what has a rule, instead of asking the model for it. Some fields have a correct answer, not an estimate: the call number (Cutter, Dewey, LC), certain MARC subfields, the archival level of description, the date range of a series. Where that rule exists, the value is computed deterministically instead of being asked of the model. A generic model approximates it, and approximating is not enough when there is a public rule.
Evaluating and flagging each record. Each record passes through an evaluation that scores it on its completeness, its correctness and how much of it could be backed by a source. When it falls below the threshold, it is flagged visibly for review. That score is what makes review by exception possible, looking at what the system flags instead of every record. A generic model delivers a batch without saying which ones are worth looking at.
Treating the standard as logic. MARC21, UNIMARC, ISAD-G, CDWA and Dublin Core are not just an output format: they carry rules and semantics about provenance, hierarchy, field repetition and controlled vocabulary. Telling a generic model “put it in MARC” tends to produce records that are MARC in form and yet incorrect according to the standard.
Reading in many languages and describing in the right vocabulary. The material and its description are rarely in a single language. The system reads in several and describes in the vocabulary of the target standard. It translates what should be translated and leaves intact what should not, such as names, titles and dates. It is not limited to a couple of languages.
Beneath all of this there is a reason why a single generic solution, identical for every institution, cannot be correct: correctness is defined locally. The same document, under the same standard, is cataloged differently at each institution, according to its classification scheme, its authorities and the language it describes in. That is why the result depends on how each institution and each collection is configured, and that configuration is the condition for the record to be the one that institution considers correct.
Enrichment from the model’s knowledge can be wrong. A mistaken identification drags along every value that depends on it, and a value with no external source, such as an ISBN or sometimes the date, rests on that identification, on the score and on the review, not on a check against a real record. That is why provenance, the score and the review are not ornaments: they are what separates enriching from contaminating. For disposable or low-risk descriptions, generic enrichment may be enough. What we describe here applies to the permanent catalog, where a false value is invisible and lasting, and where someone answers for it.
If you are evaluating AI for cataloging and you care about where the line falls between “fills fields” and “is reliable in the catalog of record,” write to us at info@janium.com; we are interested in how you would draw it for your collection.