A regional bank in the Gulf recently deployed a chatbot they described as 'bilingual.' In testing, the Arabic version answered questions about account balances by switching mid-sentence into English, misunderstood Gulf dialect entirely, and displayed text left-to-right on a right-to-left screen. Customers complained. The chatbot was quietly switched off after six weeks.
This isn't unusual. It's the norm. Here's what goes wrong — and how to avoid it.
The four failure modes
1. Training data imbalance. Most large language models are trained on internet text that is 85–90% English. Arabic training data is thinner, Gulf dialect thinner still. The model 'knows' Arabic in the way a tourist knows a foreign country — they can read the signs, but they miss the nuance.
2. RTL rendering as an afterthought. Arabic reads right-to-left. If your chat interface is built in a framework designed for English-first, RTL support is often bolted on — which means mixed-direction text, alignment errors, and emojis appearing on the wrong side of messages. Users notice immediately.
3. Code-switching is ignored. Gulf Arabic naturally mixes Modern Standard Arabic, local dialect, and English within the same sentence. 'أبي تبوك appointment' (I want to book an appointment) is a real input your chatbot will receive. Models trained on standard corpora often fail to parse this.
4. Brand voice doesn't transfer. A formal, high-trust brand voice in English often sounds stilted when directly translated to Arabic. Gulf Arabic has its own register — formal but warm, direct but respectful. A literal translation sounds like a manual.
What actually works
The chatbots that perform well in bilingual Gulf deployments share three characteristics:
First, they are trained or fine-tuned on Gulf Arabic specifically — not just Modern Standard Arabic. This means using real customer service transcripts, support ticket data, and Gulf-dialect examples in the training data.
Second, the interface is built RTL-native. Not RTL-added — RTL-native. Every UI element, every animation direction, every input field is designed for Arabic first and flipped for English, not the other way around. The chat agent on this site is built this way.
Third, the system prompt explicitly instructs the model to handle code-switching — to accept and respond in whatever language mix the user initiates, not to force them into one register.
The practical test
Before launching any bilingual chatbot, run this test: give it ten real messages from your Arabic-speaking customers, including some in Gulf dialect and some mixing Arabic and English. If it fails on more than two, you have a problem that won't fix itself in production.
The fix isn't always expensive. Sometimes it's a better system prompt. Sometimes it's a fine-tuned model. Sometimes it's the interface. But you need to know which one before you deploy.
If you're planning a bilingual deployment, I offer a pre-launch review as part of the Chat Agent service — I'll test your configuration against real Gulf Arabic inputs and tell you what needs fixing before your customers find it.