Why is the AI chatbot weaker in Swedish than it seems?

In our CRIX measurement of Swedish consumer lending chats, we asked every assistant the same question in English and in Arabic. Five of twelve answer in Swedish only. Seven manage English. Only three manage Arabic. The strangest finding was not the refusals themselves, it was that several assistants clearly understood the Arabic question, but chose to reply in Swedish anyway. One of them declined English in fluent English.

Those are configuration choices, annoying, but the smaller half of the problem. The larger half is technical and harder to see: even a chatbot that answers in perfect Swedish often misses the Swedish substance. This article is about why.

How much Swedish has a language model actually read?

First, let’s set the basics straight:

• Swedish is a small language, spoken by roughly 10 million people.
• The leading AI companies, such as OpenAI, Google and Anthropic, are all American.
• Most content on the internet is written in English.

This means that the large models are trained overwhelmingly on English text. Published documentation of major training corpora puts English at roughly 90% of the material, with Swedish making up a fraction of one per cent. Swedish consumer finance law is then a fraction of that fraction, against an ocean of English-language content about credit scores, collections agencies and consumer bureaus that work differently.

The result is a model that has learned the pattern but not the details. It knows that consumer disputes go to some independent board, that debts are collected in escalating steps, that minors have limited capacity to contract. The shape is right. But the specifics that decide a Swedish case, for example what the Inkassolag requires when a claim is disputed, which amounts currently apply, what the current wording of a Swedish regulation says, barely exist in the training material. Where the details are missing, the model fills them in from the pattern. Based on probability.

The answer sounds right and is wrong on exactly the points that matter.

Why does this hit Swedish consumers hardest?

Because the pattern the model falls back on is not a neutral pattern, it is (often) an American one. Ask an under-informed model about credit records and you get FICO logic. Ask about handling late payers and you get a process shaped by US collections practice. For an English-speaking user in a large market, the fallback is often approximately right. For a Swedish consumer, it is a confident answer built on someone else’s rules.

Our measurement shows that the questions least likely to be answered correctly are precisely those that turn on Swedish specifics, for example general consumer rights, disputes, defaults, guarantor liability. Questions covered by the company’s own material receive far better answers.

Is the language gap hard to fix?

No, not at all. Modern language models are already multilingual by nature. Our own assistant answers in 50 languages, most of them verified by native speakers. There really is no meaningful technical barrier and, as far as we can see, no regulatory one. A company that has customers writing in other languages has every reason to answer them in that language. Otherwise it risks losing the customer, or gaining a customer who does not know what they signed up for.

Understanding is not a nice-to-have in consumer credit. The person who does not fully understand what they are signing is the person most likely to end up over-indebted. Language is where that understanding starts.

About the author

Tim Bisander, founder and CPTO at Billuminate

The observations come from the CRIX measurement of Swedish consumer lending chats, August 2026. The full report is public on our website.

LinkedIn →