Why AI Models Rarely Recognise Scots as a Distinct Language

Paste Scots into a model and you usually get one of two responses: English with the spelling tidied, or a translation into English performed without being asked. Neither is a judgement the model made about Scots. Both follow from what is and is not in its training data, and that part is documented.

What is formally recognised

Scots is not a folk category. Three separate registries and instruments treat it as a language variety in its own right, and it is worth being precise about which claim each one supports.

• ISO 639-3 assigns Scots the code sco, separate from English (eng) and from Scottish Gaelic (gla), which is a different language entirely — Celtic rather than Germanic. The registry is maintained by SIL and published at iso639-3.sil.org. The same code is usable in BCP 47, so sco is a valid language tag anywhere a locale is accepted.
• The United Kingdom recognises Scots and Ulster Scots under the European Charter for Regional or Minority Languages, which it ratified in 2001. The charter and the UK’s ratification record are published by the Council of Europe.
• The Scottish census asks about Scots directly. National Records of Scotland has included Scots language questions since 2011 and publishes the results at nrscotland.gov.uk. Speaker counts run into the millions on the census’s own self-report basis; the exact figure depends on which question and which census year, so check the published table rather than a secondary summary.

There is also a scholarly reference corpus of the language: the Dictionaries of the Scots Language, which combines the Dictionary of the Older Scottish Tongue and the Scottish National Dictionary. Scots is one of the better-documented minority languages in Europe by lexicography. It is one of the worse ones by machine-readable running text, and those are different resources.

The corpus problem, including a specific one

Pretraining corpora are assembled from web crawls and from Wikipedia dumps. For a language with a small web presence, the Wikipedia edition is not one source among many — it is frequently the largest single block of clean, freely licensed, topically varied prose in the language, and its influence on the model’s idea of the language is correspondingly large.

In August 2020 the Scots Wikipedia community discovered that a very large share of its articles — tens of thousands of them — had been written or edited over several years by a single editor who did not speak Scots, and who had been producing English text with individual words respelled. The finding was made and documented on the project itself and reported widely at the time. Whatever share of that content had already been swept into crawl-derived datasets, and into models trained on them, is not something that can be unwound after the fact.

This is the concrete reason to be careful with any claim that a model “knows Scots”. A model can have learned a systematic respelling of English and be reproducing it faithfully. The output will look like Scots to somebody who does not read Scots, which is precisely the failure mode that made the Wikipedia situation last as long as it did.

What language identifiers do with it

Detection and generation fail in the same direction, for the same reason. Meta’s widely used fastText language identification model, distributed at fasttext.cc, does carry a sco label among its 176 languages — and it was trained in part on Wikipedia, which is the corpus described above. So asco prediction from it tells you the text resembles Scots Wikipedia, which is a weaker statement than it appears.

The structural problem is worse than any one tool. Scots and English share most of their high-frequency vocabulary and their entire script. Character n-gram identifiers separate languages by exactly the signal the two varieties hold in common, so short Scots texts land in the English bucket with high confidence — which is the general problem covered in distinguishing similar languages in detection, with an unusually severe instance here. If your pipeline routes by detected language, Scots input will be routed as English before any model sees it.

The features that get erased

What separates Scots from Scottish-accented English is not a spelling convention. It is lexis and morphology, and the items below are the ones that a “correct this” pass reliably destroys.

Scots              English gloss        note
ken                know                 core verb, not slang
bairn              child
kirk               church
greet              cry, weep            false friend with English "greet"
oxter              armpit
messages           shopping, groceries
dinnae / cannae    don't / can't        -nae is a bound negative clitic
Ah'm no wantin     I don't want         negation placement differs
the day / the morn today / tomorrow     definite article, not an adverb
gaun / gaed        going / went         distinct verb paradigm
een / shuin        eyes / shoes         irregular plurals, not -s
yous               you (plural)
outwith            outside of           also standard Scottish English

Two of those are the diagnostic ones. The negative clitic -nae is morphology, not spelling: dinnae is not a respelling of don’t, it is a different formation. And the day for “today” is syntax. A model that renders those correctly has learned something about Scots. A model that produces whit, ye and doon over otherwise untouched English syntax has learned a substitution cipher, which is the thing the 2020 episode consisted of.

Orthography compounds it. Scots has no single official spelling standard; the Scots Language Society and others have published recommendations, and usage in literature varies by author and region. A model has therefore seen the same word spelled several ways with no consistent majority, which is the situation described in languages with no standard AI benchmark: without a standard there is nothing to be evaluated against, so nobody evaluates.

A probe you can run

Rather than take anyone’s word for how a particular model handles Scots, this is a small, honest test you can run in a few minutes. It checks the two failure modes separately.

1. Identification without a hint. Send a paragraph of published Scots — from a book, not generated — with the instruction “name the language of this text and its ISO 639-3 code”. A model that answers eng, or “English with a Scottish accent”, has told you it does not hold Scots as a category.
2. Preservation. Send the same paragraph with “fix the typos, change nothing else”. Count how many of ken, dinnae, bairn, the day and yous survive. Every one that is rewritten is a false positive of the kind covered for another variety in Singlish and grammar checkers.
3. Generation. Ask for four sentences of Scots on a plain topic and check the morphology, not the spelling: is there a -nae clitic, is the morn used adverbially, are the irregular plurals right. Respelled English will pass a spelling check and fail all three.
4. Record the model and the date. This is behaviour of a particular checkpoint, not a property of the technology, and it is the kind of result worth re-running when you change models — see testing prompt consistency across languages.