If the goal was to create an Italian chatbot capable of making people laugh, then EMMA is probably the country’s most successful artificial intelligence project.
Too bad that wasn’t the goal.
In recent days, the AI assistant developed by LIA has gone viral due to a series of surreal responses, glaring errors, and delusions that have quickly turned the project into an inexhaustible source of memes. On social media, amid screenshots and jokes, some have already dubbed it Italy’s first “stand-up comedy AI.” And judging by the results, that’s not even such an unfair label.
Behind the laughter, however, lies an interesting question: why does EMMA make so many mistakes?
The Database Problem: When Italy Isn’t Enough
The most obvious explanation has to do with the quantity and quality of the available data.
Modern language models don’t become intelligent because someone teaches them the rules of the world.
They learn by analyzing enormous amounts of text, often collected from the entire global web. GPT, Claude, and Gemini were trained on gigantic datasets comprising billions of pages, books, articles, forums, and documents—although in this specific case, the issue remains quite controversial.
A national project aimed at building an assistant based primarily on Italian publishing content inevitably starts from a much more limited foundation.
It’s a bit like trying to train an Olympic champion by having them run only in their backyard. No matter how beautiful the backyard may be, it eventually comes to an end.
When the information base is limited, the model tends to fill in the gaps by inventing details, connecting nonexistent information, or producing answers that seem plausible but are completely wrong. This is the phenomenon known in the industry as “hallucination”, and in EMMA’s case, it seems to have evolved into a true art form.
Emma-5, the Italian AI: I don't know if it has any flaws, but so far it's doing very well pic.twitter.com/qwOVkz20on
— Matte Galt 𐀏 (@mrk4m1) June 25, 2026
I haven't laughed this much in a long time—not even as much as when I was testing Emma, the amazing Italian AI. pic.twitter.com/Yz0nF2yPEf
— Lorenzo Pregliasco (@lorepregliasco) June 25, 2026
It’s not just about the model—training matters most
The most common mistake when discussing artificial intelligence is thinking that simply choosing a good model is enough.
In reality, the model is just the engine. What makes the difference is the data, its quality, the training phase, testing, and the continuous refinement of responses.
The recent history of AI shows that training is often more valuable than the algorithm itself. Even relatively small models can achieve surprising results if they are fine-tuned with high-quality data and rigorous procedures.
Conversely, a powerful model fed with incomplete or poorly controlled data can produce unpredictable results.
In this sense, EMMA serves as a reminder of a lesson the industry continues to learn: artificial intelligence isn’t a matter of marketing, but of data. And good data requires time, expertise, and investment.
The EMMA incident also raises a broader question.
In recent years, many European countries have sought to build national alternatives to American and Chinese tech giants. The idea is understandable: technological sovereignty, data control, and reduced dependence on Big Tech.
But the reality is that developing competitive artificial intelligence requires enormous resources.
France is perhaps the most interesting case. Mistral AI, founded in 2023 by former researchers from Google DeepMind and Meta, has managed to become Europe’s leading AI player thanks to a combination of billion-dollar funding, international talent, and a strategy based on open-weight models. Today, it is consideredthe only European company truly capable of challenging OpenAI, Google, and Anthropic in certain market segments.
The difference is not merely financial. Mistral began by building globally competitive foundational models, investing in research, and attracting some of the industry’s top specialists. In just a few years, it has gone from an emerging startup to a symbol of European ambition in artificial intelligence.
EMMA, on the other hand, seems to represent the opposite approach: starting with a limited database while still attempting to offer an experience comparable to that of the major international chatbots.
The result is that comparisons with ChatGPT, Claude, or Gemini become inevitable—and often merciless.
The Real Lesson from EMMA
Artificial intelligence isn’t built by simply adding a chatbot on top of a document repository. It requires a critical mass of data, expertise, training capabilities, and investment—resources that only a few players in the world are currently able to bring together.
With Mistral, France has demonstrated that a European alternative is possible. But it has also shown just how difficult it is.
EMMA, meanwhile, has demonstrated something else: that the line between artificial intelligence and unintentional comedy is much thinner than we thought.
