A language model now understands written Swiss German almost as well as Standard German. That is what a test published on 30 October 2026 by Intochat, the developer of a chatbot platform for businesses, shows. Five artificial intelligence (AI) models from Anthropic, Google and OpenAI read customer messages written in dialect. The main risk lies elsewhere: invented promises made in the company's name.

According to the report by the Swiss German trade site Inside IT, the five models picked up 97.8% to 100% of the key information in the Swiss German messages. On the same messages in Standard German, the score ranges from 99.3% to 100%. The gap between dialect and written German has therefore become small for these models.

How Intochat tested the language models

The protocol is simple. According to Intochat's blog, the test brings together 50 customer messages from 7 regions, submitted to five recent models, for 1,000 responses in total. Inside IT specifies that each message existed in dialect (Mundart) and in Standard German (Hochdeutsch), and that the results were then compared.

The models tested are Claude Sonnet 5.5 and Claude Haiku 5.5 from the American company Anthropic, Gemini 3.8 Flash and Gemini 3.1 Pro from Google, as well as GPT-6 Luna from OpenAI, the developer of ChatGPT.

ModelKey information picked up in Swiss GermanLanguage of the responseAccuracy of responses, according to Inside IT
Claude Sonnet 5.5 (Anthropic)100%Almost always in Standard German90% correct; the rest wrong, with invented information
Claude Haiku 5.5 (Anthropic)97.8%Almost always in Standard German14% wrong or invented responses
Gemini 3.8 Flash (Google)100%Almost always in Standard German14% wrong or invented responses
Gemini 3.1 Pro (Google)100%Almost always in Standard German20% wrong, 16% partly invented
GPT-6 Luna (OpenAI)99.3%In dialect in 94% of casesThe best result; 4% with invented information

Comprehension failures remain rare and specific. Intochat cites times expressed in dialect, such as "ab em 18i" (from 18 o'clock), read as a date, and a few isolated regional words.

The language of the response clearly sets the models apart. GPT-6 Luna replied in dialect to 47 of the 50 messages in Swiss German, Intochat writes. The Claude and Gemini models almost always replied in Standard German.

Promises the company never made

The most useful finding of the test concerns wrong responses. According to Inside IT, they occur often, and the language of the question makes almost no difference. Based on this test, dialect therefore does not explain these errors.

The best score goes to GPT-6 Luna, with 4% of responses containing invented information on the questions in Standard German. Claude Sonnet 5.5 answers correctly in 90% of cases, but its other responses are wrong and embellished with inventions. Gemini 3.1 Pro brings up the rear, with 20% wrong responses and 16% partly invented.

These inventions are mostly small commitments that the business never made, according to Intochat. Examples cited: "of course we also have room for your skis", or the option of stating the new size when making a return. One response even offered a diagnosis of a smell coming from the ventilation. For the developer, these sentences are tricky, because customers understand them as a promise.

The limits of the method

This test provides a measurement, not general proof. Several points limit its scope.

  • A vendor's test. Intochat sells a platform where customers choose their model from OpenAI, Anthropic or Google. This is not a publication from a research laboratory.
  • A small sample. With 50 messages, a single error already weighs heavily on a percentage. The differences between 97.8% and 100% therefore remain fragile.
  • Written text only. The test covers typed messages. For voice, Intochat notes that dialect recognition varies: Zurich German is recognised best, while Valais and Graubünden dialects remain weak.
  • Unstable spelling. Swiss German has no recognised written form. The same word can be spelled in five ways depending on the region, Intochat points out.
  • Dated versions. The results apply to the models tested in October 2026, not to their successors.

Understanding dialect, but judging it differently

Understanding a message well does not guarantee treating its author neutrally. A study by the HAIS laboratory (Human-Centered AI-based Learning Systems) at the Bern University of Applied Sciences (BFH), published on 16 July 2026 on the SocietyByte platform, measured this.

Researchers Jeffrey Avila, Thiemo Wambsganss and Roman Rietsche submitted 200 pairs of identical sentences, in Bernese German and in Standard German, to 13 models. These sentences come from SwissDial, a public corpus of spoken Swiss German. The models had to attribute character traits to the author.

The result is clear-cut for closed commercial assistants such as GPT-4o, Claude and Gemini. They attributed negative traits to the Bernese sentences in 34% to 42% of cases, compared with 22% to 33% in Standard German. The words "rude" (unhöflich), "stubborn" (stur) and "stupid" (dumm) cluster on the dialect side.

The authors distinguish what they observe from what they explain: they see where the bias falls, not why. Older German-only models even leaned the other way, in favour of dialect. These models are not the ones in Intochat's test.

And the Swiss models?

Apertus, the large open language model from ETH Zurich and EPFL (the Swiss Federal Institutes of Technology in Zurich and Lausanne), was trained on data in Swiss German and Romansh. According to swissinfo.ch, on 7 October 2025, its developers claim more than 1,800 languages.

Understanding a language is not enough to speak it correctly, swissinfo.ch noted in its test. At the time, Apertus did not rival GPT-4, Gemini or Claude. For its next training phase, it was receiving 20 million francs from the Confederation. Other Swiss players are betting on sovereign AI, such as Giotto, a Lausanne start-up that launched its Solo and Alto models in October 2026.

What this changes for a business or an administration in German-speaking Switzerland

For a customer service team, dialect is no longer the main technical obstacle. Controlling what the assistant promises matters more. A few points emerge from the sources.

  • Choose the response language. Intochat recommends accepting dialect as input and replying in Standard German, with Swiss spelling (ss instead of ß). A dialect chatbot remains possible, but GPT-6 Luna often switches to dialect on its own.
  • Set limits on commitments. An assistant must not promise a service, a deadline or a discount that the company does not offer. Our guide to AI for SMEs reports the case, cited by Alexander Finger, chief technology officer of SAP Switzerland, of a chatbot that had granted discounts on cars.
  • Test with real messages. Intochat advises gathering around ten real messages in dialect, taken from past emails, and comparing two models before choosing.
  • Check for bias before deployment. The BFH researchers call for a university, a company or a public body to check for this dialect bias before adopting a model. For an administration, a citizen who writes in Bernese German must receive the same attention.
  • Respect data protection. The Federal Data Protection and Information Commissioner (FDPIC) pointed out on 24 September 2025 that the Federal Act on Data Protection (nFADP) applies directly to AI.

What we don't know yet

  • Whether the bias against dialect measured by the BFH persists in the versions tested by Intochat.
  • How many customers actually write to a company in dialect. Inside IT raises the question, and Intochat says most switch to Standard German.

Sources