Why creating a "Tunisian ChatGPT" is harder than you think

Imagine opening ChatGPT and asking it for the recipe for a spicy traditional dish, and it replies with a typical local expression, our humor, and our accent.

We all want an artificial intelligence that truly understands us, an AI that grasps our jokes, our culture, and our unique way of speaking. But despite the huge global enthusiasm for AI, creating a 100% Tunisian model remains a colossal challenge.

Why can’t we just build our own “Tunisian ChatGPT”? It all comes down to a mix of chaotic online writing, exorbitant costs, and the fact that robots simply don’t understand our sarcasm. Let’s break it all down.

1. The data dilemma: We don’t have enough words online

AI is like a super-intelligent child. To learn a language, it needs to read billions of words. Big companies have trained their models by having them “read” half of the internet – books, articles, and websites mainly in English.

Here’s the problem: The Tunisian dialect is primarily spoken, not written. When we write it online, it’s a chaotic mix of Latin letters, numbers, French, and English. Think about it: how do you spell a simple question word in a text message? Everyone uses their own combination of letters and numbers.

If we can’t even agree on the spelling of our own words, how is a robot supposed to learn the rules?

What AI actually has to read online

AI models are hungry for Tunisian content. Here’s a glimpse of what they actually consume on the web.

You can’t build a massive artificial brain using only a few thousand misspelled social media comments and captions.

2. The exorbitant bill

Building an AI from scratch isn’t like creating a standard mobile app. It requires massive, warehouse-sized supercomputers filled with ultra-powerful and extremely expensive chips.

Tech giants spend hundreds of millions of dollars just to pay for the electricity and hardware needed to train their AI. For a local project aimed solely at our population, finding that kind of funding is incredibly difficult.

The reality of AI in numbers

Let’s compare what it takes to build these models:

3. Robots don’t understand Tunisian sarcasm

Tunisians are world champions at code-switching: we start a sentence in our mother tongue, add a verb in French, and finish with a slang word in English.

Current AI models are too rigid for this. If an AI is trained on ultra-classical and formal language, it will speak like a news anchor. It will completely miss the sarcasm, hidden meanings, and cultural flavor of a real Tunisian conversation.

(As the graph shows, AI is super smart in English, but completely lost when it comes to our slang or sarcasm).

So, are we doomed to talk to robots formally?

Not exactly! We don’t need to build a multi-million dollar AI from scratch.

The smartest strategy for Tunisia is to take an existing, free, and open-source AI and send it to a Tunisian school. By gathering clean and organized data, like transcriptions of Tunisian podcasts, plays, and radio shows, local developers can “fine-tune” these models to understand our dialect.

The technological talent already exists here and is ready to take on the challenge. We just need to start transforming our beautiful, chaotic dialect into data that machines can actually read.

Our mission at IyedeX: Taking action

This is exactly where our role at IyedeX comes in. We understood that we are unique, with a rich culture and a complex dialect, and that we need technology that truly resembles us. Our mission is to adapt AI to Tunisia so that it stops being a foreign concept and starts resonating with our daily lives.

That’s why we didn’t wait. We have already taken this decisive step by launching our very first product aimed at the general public (B2C): 3amAli. Through this launch, we have taken the first step to bring an AI tailored directly to Tunisians. We firmly believe it’s time to have an assistant that understands not only our words, but also our identity, our humor, and our culture.