Getting Austrian German Vocabulary Instead of Standard German From AI

#ai #llm #machinelearning #programming

Ask for German and you get Germany’s German. That is not a rendering of a neutral standard with a regional flavour missing — Austrian German is its own codified national variety, and some of its vocabulary is written into European Union law by name.

Austrian German is codified, not slang

German is a pluricentric language: it has several national standard varieties, each with its own codifying reference. Germany’s is Duden. Austria’s is the Österreichisches Wörterbuch, published since 1951 under the Austrian education ministry and used as the orthographic authority in Austrian schools and administration. Switzerland’s differs again, most visibly in abolishing ß — a point worth keeping separate, because Austria keeps ß exactly as Germany does, so ß is not a tell for anything except Swiss text. See Swiss German vocabulary and umlaut and eszett normalisation for that half.

The distinction between a national standard and a dialect matters for prompting because it changes what you are asking for. Asking for Viennese dialect asks the model to write non-standard speech. Asking for Austrian Standard German asks for the formal written register used in Austrian newspapers, contracts and government forms — fully standard, fully correct, and lexically different from Germany’s in a specific and enumerable set of places.

The twenty-three treaty words

When Austria joined the European Union, its accession treaty carried a protocol on the use of specifically Austrian terms of the German language, listing twenty-three terms — nearly all foodstuffs — which have the same legal status in Union texts as the Germany-standard equivalents. The treaty text is published by EUR-Lex. It is an unusually clean thing to build a prompt around: a closed list, in primary law, of words that are not regional colour but officially recognised alternatives.

Austrian           Germany-standard      English
Erdäpfel           Kartoffeln            potatoes
Paradeiser         Tomaten               tomatoes
Marillen           Aprikosen             apricots
Ribisel            Johannisbeeren        currants
Weichseln          Sauerkirschen         sour cherries
Karfiol            Blumenkohl            cauliflower
Kohlsprossen       Rosenkohl             Brussels sprouts
Fisolen            grüne Bohnen          green beans
Melanzani          Auberginen            aubergines
Vogerlsalat        Feldsalat             lamb's lettuce
Eierschwammerl     Pfifferlinge          chanterelles
Kren               Meerrettich           horseradish
Obers              Sahne                 cream
Topfen             Quark                 curd cheese
Faschiertes        Hackfleisch           minced meat
Lungenbraten       Filet                 tenderloin
Powidl             Pflaumenmus           plum butter
Grammeln           Grieben               cracklings

That is a selection; the protocol’s list also covers cuts of meat (Beiried, Rostbraten, Schlögel, Hüferl, Nuss) whose Germany-standard equivalents differ because the butchery itself is cut differently. A model asked for an Austrian recipe will often produce Kartoffeln and Sahne with a couple of Marillen sprinkled in, for the same reason a Nynorsk request produces Bokmål with ikkje in it: the famous items get substituted and the substrate does not.

Calendar and administrative vocabulary

The food list gets quoted because it is in a treaty, but the vocabulary that will actually break a business document is administrative. If you are generating an Austrian invoice, employment letter or form, Januar instead of Jänner is the single most visible tell in the language.

Austrian                Germany-standard        English
Jänner                  Januar                  January
Matura                  Abitur                  school-leaving exam
Spital                  Krankenhaus             hospital
Sackerl                 Tüte                    bag
heuer                   dieses Jahr             this year
Trafik                  Tabakladen              tobacconist
Erlagschein             Überweisungsträger      payment slip
Meldezettel             Meldebescheinigung      residence registration
Bezirkshauptmannschaft  Landratsamt             district authority
Landeshauptmann         Ministerpräsident       head of a state government
der Akt                 die Akte                the file
Pension                 Rente                   retirement pension
Bundesheer              Bundeswehr              armed forces

Jänner is not optional or old-fashioned in Austria; it is the standard form used in legislation and in dated correspondence. Feber for February exists but is markedly more archaic, and a model that produces it is over-shooting. That asymmetry is typical: the safe instruction is a list of the forms you want, not “maximise Austrian-ness”, which reliably yields dialect spelling in a document that should be formal.

Where it stops being vocabulary

Three differences are grammatical, and they are the ones a vocabulary substitution pass will miss entirely.

Perfect auxiliary with position verbs. Austrian (and southern German generally) takes sein where northern standard takes haben: ich bin gesessen, ich bin gestanden, ich bin gelegen. Germany-standard writes ich habe gesessen. This is invisible to a word list and immediately audible to a reader.
Noun gender. das Cola and das E-Mail in Austria against die Cola and die E-Mail in Germany; das Joghurt against der Joghurt. Gender errors propagate into every article and adjective ending in the sentence, so one wrong gender marks a whole paragraph.
Preterite avoidance. Southern varieties strongly prefer the perfect over the simple past in narrative prose. A text full of ging, sagte and kam reads as northern even when every content word is Austrian.

One text for Germany, Austria and Switzerland

Most teams reach this page because they have one German text and three markets. It is worth being clear that a single text native in all three does not exist, and that the obstacle is not vocabulary.

A neutral lexical subset is achievable: avoid Kartoffeln and Erdäpfel both, write around the divergent nouns, and you have something unmarked. But the perfect auxiliary is not avoidable. Any sentence reporting that something sat, stood or lay somewhere forces the choice between ist gesessen and hat gesessen, and one of the two audiences will read the result as foreign. Noun gender forces the same choice every time an article precedes Cola, E-Mail or Joghurt. Switzerland then adds a purely orthographic fork that Austria does not share, since Swiss usage writes ss where both Germany and Austria write ß.

So there are three honest options, and they should be chosen deliberately rather than arrived at. Publish Germany-standard German everywhere, which is understood in all three countries and reads as imported in two of them — entirely defensible for technical documentation. Generate once per market from a shared brief plus a per-locale glossary, which is what the treaty word list and the administrative table above are for. Or hand-write the small number of documents where belonging matters — onboarding, marketing, anything legal — and generate the rest.

Whichever you choose, store de-DE, de-AT and de-CH as separate locales rather than one de. A single German locale means the decision is made implicitly by whichever component looks at it last, and it will be made in favour of the majority variety every time.

Getting it out of a model

Treat it the way you would treat any minority form of a language whose majority form dominates the corpus: the instruction shifts the distribution, the examples pin it, and the check is mechanical.

Ask for österreichisches Standarddeutsch explicitly, and say formal written register, so the model does not reach for Viennese dialect spelling.
Set the locale you carry through your system to de-AT rather than de, and pass it into the prompt rather than assuming the model infers it.
Paste the vocabulary pairs you care about into the prompt as a glossary block. A model given Jänner, Erdäpfel and Matura as required terms uses them; a model told “Austrian vocabulary” guesses which ones you meant.
Include one short paragraph of genuine Austrian administrative prose as a demonstration, so the perfect-tense and gender behaviour comes along with the words.
Run a deny-list over the output for Januar, Kartoffeln, Sahne, Abitur, Tüte and die E-Mail. Every hit is a real defect and the check costs nothing.

Which forms a given model produces for de-AT shifts between model generations, and Austrian usage itself varies by region and register — Sackerl is universal, Feber is not. Treat the glossary as the contract and the model as something that fills it in.