Do AI Chatbots Understand Swiss German? We Tested 5 Models
50 customer messages from 7 regions, 5 current AI models, 1,000 answers: every model understood written Swiss German almost as well as standard German. The differences are in how they reply.
- Author
- intoCHAT Team
- Published
- Reading time
- 6 min read

On this page
Can an AI chatbot understand customers who write in Swiss German? Many Swiss small businesses hesitate for exactly this reason. We wanted to find out, so we tested five current language models with 50 typical customer messages in Swiss German dialect, each also written in standard German for comparison. All messages, answers and scores are available to download at the end.
Key findings
- Understanding is no longer the problem. The five models caught 97.8 to 100% of the key facts in Swiss German messages, and 99.3 to 100% in standard German.
- Clock times and regional words trip models up. Two models read «ab em 18i» ("from 6 pm") as a date, and one didn't understand «Gielä» (Bernese for "boys").
- The big difference is in the reply. GPT-6 Luna wrote back in dialect in 47 of 50 cases. The other four models almost always replied in standard German.
- Invented details are not caused by the dialect. Small promises the business never made appeared in 6 to 24% of replies to Swiss German messages, and in 4 to 16% of replies to standard German ones.
- The German "ß" is no longer an issue (Swiss spelling uses "ss"). It appeared in 1 of 250 replies.
How we tested
- 50 customer messages in written Swiss German from five sectors (hotel, property management, online shop, renovation, garage) and seven regions (Zurich, Bern, Basel, St. Gallen, Lucerne, Valais, Graubünden). The intoCHAT team wrote the messages, modelled on typical customer enquiries; they are not real customer messages.
- Every message also exists in standard German with exactly the same content, so you can see what is down to the dialect and what isn't.
- 5 models, as businesses use them through the providers' APIs: Claude Haiku 5.5 and Claude Sonnet 5.5 (Anthropic), GPT-6 Luna (OpenAI), Gemini 3.8 Flash and Gemini 3.1 Pro (Google). Default settings, one run, October 2026.
- Test 1, understanding: the model restates what the customer wants. We check 134 key facts, such as dates, times, number of people and the actual request.
- Test 2, replying: the model answers as the chat assistant of a fictional business, with a short list of facts about it. We check the reply language, "ß", correctness and invented details.
- Scoring: a separate model, Claude Opus 5.5, scored each of the 1,000 answers against fixed criteria. Because it comes from Anthropic, like two of the tested models, we publish every answer and every score.
Result 1: understanding
| Model | Key facts correct, Swiss German | Key facts correct, standard German |
|---|---|---|
| Claude Sonnet 5.5 | 100% | 100% |
| Gemini 3.8 Flash | 100% | 100% |
| Gemini 3.1 Pro | 100% | 100% |
| GPT-6 Luna | 99.3% | 100% |
| Claude Haiku 5.5 | 97.8% | 99.3% |
Written Swiss German, as customers use it in chat and email, is no longer an obstacle for current models. Even the messages from Valais and Graubünden were understood in full.
Where models stumbled:
- «Ich chan erscht ab em 18i, gits au Bsichtigungstermin am Abig oder am Samstig?» ("I'm only free from 6 pm, are there viewings in the evening or on Saturday?") Two models turned "from 6 pm" into "from the 18th", a date.
- «Gits dä Mantu o für Gielä? Mi Sohn isch 10i.» ("Is this coat available for boys? My son is 10.") One model couldn't translate «Mantu» (coat) and «Gielä» (boys).
- «wie hooch sin d nebechöschte und isch e parkplatz drbii?» ("How high are the service charges and does it come with a parking space?") One model read it as "is there parking nearby".
So the typical sources of error are dialect clock times and individual regional words, not the dialect as a whole.
Result 2: replies
| Model | Reply in dialect | Reply correct (CH / DE) | Invented details (CH / DE) |
|---|---|---|---|
| GPT-6 Luna | 94% | 92% / 90% | 6% / 4% |
| Claude Sonnet 5.5 | 8% | 90% / 90% | 12% / 10% |
| Gemini 3.8 Flash | 0% | 86% / 86% | 14% / 14% |
| Claude Haiku 5.5 | 2% | 82% / 86% | 14% / 14% |
| Gemini 3.1 Pro | 2% | 76% / 80% | 24% / 16% |
CH = customer message in Swiss German, DE = the same message in standard German. "Reply in dialect" counts replies written fully or partly in Swiss German.
Reply quality barely depends on whether the customer writes in dialect or standard German. The differences are between the models.
The invented details were mostly small promises the business had never made: "of course we have room for your skis", "you can choose the new size directly when you send the item back", or a diagnosis for a smell from the car's ventilation. These are exactly the sentences that matter to a business, because customers read them as commitments.
Reply in dialect or standard German?
A reply in dialect feels friendly. But Swiss German has no standard spelling, and a reply in Zurich German to a customer from Bern can feel odd. Most Swiss businesses write to their customers in standard German. If you use an AI chatbot, decide this on purpose rather than leaving it to the model. More in our post Swiss German and AI chatbots: what actually works in 2026.
What this means for Swiss businesses
- Dialect is no longer a reason to go without a chatbot. Written Swiss German is understood reliably.
- Set the reply language. State it clearly in the instructions: standard German with Swiss spelling, or dialect on purpose.
- Have times and dates confirmed. For bookings, have the assistant ask back when something is ambiguous, such as «am 18i».
- Guard against promises. Tell the assistant to say only what is in its knowledge and to hand everything else to your team. And read the first conversations yourself.
Limits of this test
- The intoCHAT team wrote the messages. Real messages are often messier, with abbreviations, typos and strong dialect, and might score lower.
- We tested only written messages. Spoken Swiss German, on the phone or in voice messages, is considerably harder.
- One run per model with default settings. Results may change with new model versions.
- We did not rate how good the dialect replies are linguistically.
- 50 messages is a small sample. Differences of a few percentage points between models are not meaningful.
Download the data
You may use the data freely with the credit "intoCHAT, October 2026".
intoCHAT is an AI agent for websites, WhatsApp and email and can use some of the models tested.
Frequently asked questions
Do AI chatbots understand written Swiss German?
In our test, yes. The five models we tested caught 97.8 to 100% of the key facts in Swiss German customer messages, and 99.3 to 100% in standard German. Mistakes came mostly from dialect forms of clock times and from regional words.
Do AI chatbots reply in Swiss German dialect?
It depends on the model. GPT-6 Luna replied fully or partly in dialect to 47 of 50 Swiss German messages. Claude Haiku 5.5, Claude Sonnet 5.5, Gemini 3.8 Flash and Gemini 3.1 Pro almost always replied in standard German.
Do AI chatbots invent more when customers write in Swiss German?
Not noticeably. Small invented promises appeared in 6 to 24% of replies to Swiss German messages and in 4 to 16% of replies to the same messages in standard German. The issue lies with the model and its instructions, not with the dialect.
Did the test include spoken Swiss German?
No. We only tested written messages, as they arrive in website chat, WhatsApp or email. Spoken Swiss German on the phone or in voice messages is considerably harder.


