Unlike Portuguese or Chinese, Dutch has a single written standard governed jointly by the Netherlands and Belgium. There is no Flemish spelling to switch to. What a Belgian reader notices is vocabulary and register, and neither of those is carried by a locale tag.
One standard, two usages
The Nederlandse Taalunie, the Dutch Language Union, is a treaty organisation of the Netherlands, Flanders and Suriname, and it maintains a single official spelling for all of them. The Taalunie publishes the standard. So Vlaams — Flemish — does not name a separate written language in the way zh-Hant names a separate script. It names Belgian usage within a shared standard, plus a family of spoken dialects.
This has a direct practical consequence: the locale tag nl-BE is much weaker as a steering signal than pt-BR or es-MX are, because there is far less locale-differentiated Dutch text in training data to attach to it. Where the Portuguese and Spanish pages in this cluster can lean on the tag, this one cannot. The vocabulary has to go in the prompt explicitly.
The default the model produces is Netherlands Dutch, for the usual reason. The Netherlands has roughly 17 million people to Flanders’ 6.5 million, and Dutch-language web publishing skews further north still. A generic Dutch prompt therefore returns northern usage, and the gap is wide enough that Belgian readers notice.
Belgicisms worth knowing
Belgian Dutch retains more French loanwords, where Netherlands Dutch has more often taken English ones or coined native replacements:
concept Belgian (nl-BE) Netherlands (nl-NL)
-------------------------------------------------------------
mobile phone gsm mobiel / mobieltje
fridge frigo koelkast
umbrella regenscherm paraplu
butcher beenhouwer slager
dry cleaner droogkuis stomerij
reception / desk onthaal receptie
cleaner poetsvrouw schoonmaakster
motorway autostrade / snelweg snelweg
toasted sandwich croque-monsieur tosti
onion ajuin ui
student room kot studentenkamer
desire, appetite goesting zin
job, work job baan
to phone bellen / telefoneren bellen
lorry camion / vrachtwagen vrachtwagen
Two of these are worth singling out. Gsm is universal in Belgium and essentially unused in the Netherlands, and it appears in any product that mentions a phone number, so it is high-frequency and highly visible. Kot and goesting are the words Belgians themselves name when asked what marks their Dutch, which makes them good markers to check for but poor markers to force — overusing them produces output that reads as a parody rather than as Belgian.
There is a smaller set of grammatical preferences too: Belgian Dutch is more comfortable with the -de/-te variation in certain verb forms, prefers different verb-cluster orders in subordinate clauses, and uses some prepositions differently. These are tendencies rather than rules, and no prompt will produce them reliably. The vocabulary is where the achievable difference lies.
Register, and the u problem
The single most consequential difference is not a word. Belgian Dutch uses the formal second person u far more broadly than Netherlands Dutch does. Northern Dutch has shifted heavily toward je/jij in commercial and service contexts, to the point where a Dutch bank or airline addressing a customer as u can read as stiff. In Belgium, the same je reads as overfamiliar in exactly the contexts where a Dutch company would use it.
This matters more than the vocabulary because it applies to every sentence in a customer-facing product, and getting it wrong is a tone problem rather than a comprehension problem. It is also the one thing a model will not infer from “write in Belgian Dutch”, because it is a register decision rather than a dialect fact. Say which form of address you want explicitly, as argued in specifying formal and informal register.
A complication: in Belgian speech, u and the older gij have partly merged in function, and u is not always as formal as its Netherlands equivalent. That is a speech phenomenon and it does not carry into standard written text, so for written output the practical rule is simply that u is the safer default for Belgium and je the safer default for the Netherlands.
Tussentaal, and why you probably do not want it
Between standard Belgian Dutch and the local dialects sits tussentaal — literally “in-between language” — an informal spoken register with its own pronouns (ge, gij), diminutives and verb forms. It is what a great deal of Flemish television dialogue is in, and it is what people often have in mind when they say Dutch output sounds “too Dutch”.
It is also almost entirely a spoken register, with no standard orthography and a thin written corpus, so a model asked for it will produce the same class of approximation described in Swiss German dialect: standard grammar with a few informal forms substituted. For product copy, documentation or support responses, standard Belgian Dutch is both what you want and what is achievable. Tussentaal is for dialogue, and it needs a human.
What to put in the prompt
Because the locale tag is weak here, the prompt has to carry the content. Four things, in order of effect:
- State the audience, not the variety. “Write for a Flemish (Belgian) reader, using Belgian Dutch usage” outperforms “write in Flemish”, which the model may interpret as a request for dialect.
- Supply eight to twelve Belgicisms relevant to your domain, with the Netherlands term beside each as the thing to avoid. Two columns, not one; the contrast is what makes it actionable.
- Specify the form of address explicitly. “Address the reader as u throughout” is a single sentence that fixes the most visible difference.
- Say that the spelling is standard. Adding “use standard Dutch spelling per the Taalunie” prevents a model from inventing phonetic dialect spellings when asked for Flemish — a real failure mode precisely because the word “Flemish” is ambiguous between the standard usage and the dialects.
Verification is easier than for most varieties: grep the output for the Netherlands column above. Unlike spelling checks, this produces almost no false positives, because these are ordinary content words rather than function words.
Top comments (0)