If your records are not in English, the model will not notice.

If your collection is cataloged in a language other than English, and you are thinking about generating image descriptions, there is a default waiting to catch you out. It caught me last week, across 3,468 records, and I did not notice until the run had finished.

The short version: your prompt is probably written in English, so the model will answer in English, whatever language your records and your visitors are in. Nothing warns you. The descriptions come back fluent and sensible and completely unusable as alt text on your own site.

Here is how I found it.

The collection was Joconde, the French Ministry of Culture's national database, which aggregates catalog records from museums across France through the Plateforme Ouverte du Patrimoine. Worth saying plainly that this is a slice of a national aggregation rather than a museum's collection: LinkedCulture holds 10,000 Joconde records out of a database running to several hundred thousand, and only 3,468 of those carry an image URL at all. The other 6,532 are catalog records with no picture attached.

So what finished is narrower than it sounds. Every Joconde record in the corpus that has an image now has a description, which is 3,468 of them.

Every description came back in English

The Joconde records are in French. Their titles are in French, their catalog fields are in French, and the collection is published by a French institution for a largely French-speaking public. The model was given the photograph alone, with no text at all, and it wrote every description in English:

A religious figure sits on a throne-like chair, surrounded by a tree with fruit, holding a book.

A cartoon depicting three men in period clothing, one holding a violin, another a sword, and the third a staff.

A clock face with a rooster and a tree branch above it.

Those are reasonable descriptions. They are also useless as alt text on a French-language museum page, and I generated 3,468 of them without stopping to think about it. If you run this over a French, Dutch or Italian collection, you will get the same result and it will look like it worked.

Nothing in the pipeline chose English. The instructions are written in English, so the model answered in English, and I never told it otherwise. The corpus happens to be mostly English-language institutions, so this default went unnoticed until it reached a collection that is not.

If you are running something like this over a multilingual corpus, catch this first. The language of the description needs to be a decision, tied to the record or the institution, not a side effect of the language your prompt happens to be written in.

What I plan to do about it

The descriptions are stored with their model, prompt version and generation date, so I can find every one produced by an English-language prompt and regenerate it. Since the model never saw the record's text, regenerating in French costs the same as the first pass did, which for this collection is a few cents.

The harder question is which language a description should be in, and I do not think there is one answer, so you will have to make the same call. A French museum record on a French museum's site wants French. The same record surfaced in a cross-institutional search alongside English and Dutch records may want both.

If it helps, here is where I landed on the storage. Language is its own column on the description, not part of the version string, and it is part of the uniqueness key, so a record can hold the same description in several languages rather than one overwriting the other. The language is chosen per institution rather than per record, since it is a property of who the collection is published for. Storing several is easy. Deciding which one a given page should announce is the part I have not solved.

One thing I would not do is throw the English versions away. They are the wrong language for alt text on a French page and exactly the right thing for search: an English-language query describing what an object looks like has nothing in a French catalog record to match against, and now it has something. If your collection is non-English and you want it findable by English speakers, those descriptions are the bridge.

The catalog descriptions here are a third kind again

I have been tracking what museums put in their description fields, because it varies more than the field name suggests, and Joconde adds a category I had not seen.

Getty's description field carries genuine visual description. Cleveland's carries art-historical context about the maker. MIA's carries dimensions and place. Joconde's carries technical manufacture and conservation language, written for specialists:

Incision fine, incision en v dissymétrique profonde, champlevé, incision en v profonde, incision en u, coloration.

amulette en forme de truie Faïence Egyptienne (?)

That first one is a precise account of how the surface was worked. It tells a conservator a great deal and tells a blind visitor nothing about what the object looks like. The second is closer to useful, and carries the museum's own uncertainty in that "(?)", which I would want preserved rather than resolved away.

I keep relearning the same lesson here: you cannot infer what a description field contains from its name. If you are planning work like this across several institutions, sample each source separately and read what actually comes back. I have now been wrong about this for four different museums by generalising from one record.

What is missing from these records

Two other gaps in the Joconde data limit what can be done with the result.

The institution field is empty across the records I hold, so I cannot say which French museum holds any given object. Joconde is an aggregation, and the holding institution is what would let a description be sent back to the people able to check it. Without that there is nobody specific to offer this to, unlike the Museo Egizio.

And two thirds of the records have no image. A description pipeline has nothing to say about a record with no picture, so those 6,532 sit outside this work entirely.

Where this leaves the count

Two collections are now finished in the sense that every image they contain has a description: the Museo Egizio at 5,228, and Joconde at 3,468. Roughly 8,700 objects, for well under two dollars of rented compute between them.

The Egizio one produced a finding about quality, where the museum's own curators clearly outperformed the model. This one produced a finding about language, which I would not have found by testing on English collections. Both took a completed collection rather than a sample to surface, so I would rather finish collections than benchmark across them.

The larger institutions are running now. If you work with a multilingual collection and have solved the description-language question in a way you are happy with, I would genuinely like to hear how: contact@EdwardMonk.com.

---

Descriptions generated 2026-09-05 with Qwen2.5-VL-7B-Instruct from the photograph alone, no catalog text supplied. All are marked machine-generated and unreviewed. Earlier notes: describing one museum's whole collection and what this costs at different collection sizes.